<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/">
<channel>
<title>Flavoureak Build log</title>
<link>https://flavoureak.co.uk/blog/</link>
<atom:link href="https://flavoureak.co.uk/feed.xml" rel="self" type="application/rss+xml"/>
<description>What Flavoureak learns building and running SaaS: MVP scope, billing, AI costs, search, and what each sector needs.</description>
<language>en-gb</language>
<lastBuildDate>Fri, 16 Oct 2026 08:00:00 GMT</lastBuildDate>
<item>
<title>What goes into a SaaS MVP, and what waits</title>
<link>https://flavoureak.co.uk/blog/saas-mvp-scope/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/saas-mvp-scope/</guid>
<pubDate>Fri, 16 Oct 2026 08:00:00 GMT</pubDate>
<category>SaaS</category>
<description>An MVP is the smallest product someone will pay for, not the smallest product you can build. Here is what belongs in the first release and what can wait for real users.</description>
<content:encoded><![CDATA[<ul><li>Include what a paying customer needs on day one: sign-up, the core workflow, billing and a way for you to support them.</li><li>Leave enterprise features, integrations and mobile apps until customers ask for them.</li><li>Instrument the MVP from the start, so the second release is decided by data.</li></ul><h2>Minimum, but viable</h2>
<p>The point of an MVP is to learn whether people will pay, as early and cheaply as possible. That means it has to be good enough for a real customer to use for real work. A demo that only works when you drive it teaches you very little.</p>
<h2>What belongs in the first release</h2>
<ul><li><b>Sign-up and login</b>, with password reset and email verification.</li><li><b>Workspaces</b>, so a customer can invite their team. Adding multi-tenancy later is expensive.</li><li><b>The core workflow</b>: the one job your product does better than a spreadsheet.</li><li><b>Billing</b>: plans, a trial and a way to take payment. Charging from day one is the strongest signal you will get.</li><li><b>An admin console</b>, so you can see accounts, fix data and help customers without a developer.</li><li><b>Analytics and logging</b>, so you know what people actually use.</li><li><b>A marketing site</b> that explains the product and can be found.</li></ul>
<h2>What can wait</h2>
<ul><li><b>SSO and enterprise controls</b>, until an enterprise customer asks and will pay for them.</li><li><b>A public API and integrations</b>, until customers tell you which systems matter.</li><li><b>Mobile apps</b>, unless the core job happens away from a desk.</li><li><b>Usage-based pricing</b>, until you know which usage customers value.</li><li><b>White-labelling</b>, until a reseller is waiting for it.</li></ul>
<h2>Build the foundations properly</h2>
<p>Some things are cheap to do right at the start and painful to change later: the data model, how tenants are separated, how permissions work, and how deploys happen. An MVP can be small without cutting these corners.</p>
<h2>Decide the second release with data</h2>
<p>Every MVP should ship with product analytics and a short list of questions it is meant to answer. The second release is then planned from what customers did, not from a backlog written before anyone used it.</p>
<div class="callout"><b>Scope yours.</b> The <a href="https://flavoureak.co.uk/founders/">MVP scope builder</a> lets you sort features into now and later.</div>]]></content:encoded>
</item>
<item>
<title>What drives the cost of a custom app</title>
<link>https://flavoureak.co.uk/blog/what-drives-app-cost/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/what-drives-app-cost/</guid>
<pubDate>Wed, 14 Oct 2026 08:00:00 GMT</pubDate>
<category>Working with us</category>
<description>Two apps that look the same in a screenshot can differ in cost several times over. These are the things that move the price, and how to keep them under control.</description>
<content:encoded><![CDATA[<ul><li>Cost follows the number of user types, rules and integrations more than the number of screens.</li><li>Offline working, payments and compliance each add real work, and are often worth it.</li><li>A fixed-price first release with a tight scope is the best way to control cost.</li></ul><h2>Screens are the wrong unit</h2>
<p>People often estimate an app by counting screens. Screens are the visible part. The cost sits underneath: who can do what, what the rules are, what has to talk to what, and what happens when something goes wrong.</p>
<h2>What moves the price</h2>
<h3>Types of user</h3>
<p>An app used only by your office staff is one thing. Add customers, subcontractors and an admin who can see everything, and each needs its own permissions, screens and notifications.</p>
<h3>Rules</h3>
<p>Tax treatments, approval steps, pricing rules and deadlines are where the real work is. They need to be written down, built and tested against real examples from your business.</p>
<h3>Integrations</h3>
<p>Every connection to another system, such as accounting software, a payment provider or a supplier's API, has to handle failures, retries and changes on the other side.</p>
<h3>Platforms</h3>
<p>A web app, an iPhone app and an Android app are three front ends. Sharing one backend and one set of rules stops that from tripling the cost, but it is still more work than one.</p>
<h3>Offline working</h3>
<p>An app that must work with no signal needs a database on the phone, a sync process and rules for conflicts. For people on sites or in the field it is essential. For an office tool it usually isn't.</p>
<h3>Data you already have</h3>
<p>Moving years of spreadsheets, an old system or paper records into the new one is often underestimated. Messy data takes time to clean.</p>
<h2>What it costs to keep running</h2>
<p>After launch there is hosting, monitoring, backups, security updates, app store releases and the occasional fix when another service changes. Budget for running costs from the start.</p>
<h2>How to keep cost under control</h2>
<ul><li><b>Scope a first release</b> around the one job that costs you most today.</li><li><b>Fix the price</b> for that release from a written scope.</li><li><b>See it every two weeks</b>, so changes happen early, when they are cheap.</li><li><b>Leave the nice-to-haves</b> for the second release, when real use shows which ones matter.</li></ul>
<div class="callout"><b>Want to know what drives cost in your case?</b> <a href="https://flavoureak.co.uk/contact/">Send us a brief</a> and we'll tell you.</div>]]></content:encoded>
</item>
<item>
<title>SaaS security before your first enterprise customer: what buyers ask</title>
<link>https://flavoureak.co.uk/blog/saas-security-before-enterprise/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/saas-security-before-enterprise/</guid>
<pubDate>Wed, 14 Oct 2026 08:00:00 GMT</pubDate>
<category>SaaS</category>
<description>Large customers send a security questionnaire before they sign. Have MFA, tested backups, an incident plan, Cyber Essentials, a recent pen test and a data processing agreement ready first.</description>
<content:encoded><![CDATA[<ul><li>A security questionnaire is a set of questions a larger customer sends before buying, covering access control, encryption, backups, testing, incident response, suppliers and data protection.</li><li>Cyber Essentials is the UK government-backed certification for basic cyber security controls, and Cyber Essentials Plus adds a hands-on technical audit; UK central government requires it for some contracts.</li><li>ISO 27001 certifies an information security management system and is widely recognised in the UK and Europe, while SOC 2 is an auditor’s report on your controls that US buyers more often ask for.</li></ul><p>Before a large organisation buys a SaaS product, its security team sends a questionnaire and asks for evidence. A security questionnaire is a structured list of questions about how you protect the customer’s data, from staff access and encryption to backups, testing and incident response. A small company can answer one honestly if a few basics are in place, along with Cyber Essentials, a recent penetration test and a data processing agreement.</p>
<h2>What a security questionnaire asks</h2>
<p>Some customers use their own spreadsheet. Others use a standard set, such as the Cloud Security Alliance’s CAIQ or the Shared Assessments SIG. The topics are much the same.</p>
<ul><li><b>Access control.</b> Multi-factor authentication, single sign-on, least privilege, and how you handle joiners and leavers.</li><li><b>Data protection.</b> Encryption in transit and at rest, where data is stored and how tenants are kept apart.</li><li><b>Operations.</b> Backups, patching, logging and monitoring.</li><li><b>Assurance.</b> Certifications, penetration tests and written policies.</li><li><b>Incidents.</b> How you detect, respond to and report them.</li><li><b>Suppliers.</b> Your hosting provider and other sub-processors.</li><li><b>Privacy.</b> Your data processing agreement, retention, and deletion at the end of the contract.</li></ul>
<p>Keep a security pack ready with your policies, a library of approved answers, certificates, a pen test summary and your sub-processor list. Each questionnaire after the first then takes a fraction of the time.</p>
<h2>Cyber Essentials and Cyber Essentials Plus</h2>
<p>Cyber Essentials is a UK government-backed scheme, overseen by the National Cyber Security Centre and delivered through IASME, that certifies five basic technical controls. They are firewalls, secure configuration, security update management, user access control and malware protection. You complete a self-assessment questionnaire, which a certification body checks, and the certificate lasts a year.</p>
<p>Cyber Essentials Plus covers the same controls and adds a hands-on audit, in which an assessor tests a sample of your devices and accounts. The scope includes the company’s laptops, phones, servers and cloud services, so it is within reach of a small team that keeps its devices and accounts in order. UK central government requires Cyber Essentials for some contracts, such as certain contracts that involve handling personal information, and some private-sector buyers ask for it too. The requirements are updated from time to time, so check the current version before you apply.</p>
<h2>ISO 27001 and SOC 2 compared</h2>
<p>ISO/IEC 27001 is an international standard for an information security management system, meaning the policies, risk assessments and processes you use to manage security over time. An accredited certification body audits you and issues a certificate, followed by surveillance audits each year and recertification every three years.</p>
<p>SOC 2 is a US framework from the AICPA. A licensed CPA firm examines your controls against the Trust Services Criteria and writes a report. A Type 1 report covers how your controls are designed at one point in time. A Type 2 report covers how they operated over a period, usually several months.</p>
<table><thead><tr><th></th><th>Cyber Essentials</th><th>Cyber Essentials Plus</th><th>ISO 27001</th><th>SOC 2</th></tr></thead><tbody><tr><td>What it is</td><td>Self-assessed certification of five controls</td><td>The same controls, independently tested</td><td>Certified security management system</td><td>Auditor’s attestation report</td></tr><tr><td>Who checks</td><td>A certification body reviews your answers</td><td>An assessor tests your systems</td><td>An accredited certification body</td><td>A licensed CPA firm</td></tr><tr><td>Renewal</td><td>Every year</td><td>Every year</td><td>Three-year cycle with yearly audits</td><td>Buyers usually expect a new report each year</td></tr><tr><td>Most often asked for by</td><td>UK public sector and UK businesses</td><td>UK public sector and larger UK buyers</td><td>UK and European enterprises</td><td>US enterprises</td></tr><tr><td>Effort for a small SaaS</td><td>Low</td><td>Low to moderate</td><td>High</td><td>High</td></tr></tbody></table>
<p>For a UK SaaS company, Cyber Essentials is the sensible first step. ISO 27001 usually follows when large UK or European customers require it, and SOC 2 when US customers do. Many controls overlap, so the work for one makes the next easier.</p>
<h2>Penetration testing</h2>
<p>A penetration test is an authorised, simulated attack in which specialists look for weaknesses in your application and infrastructure. For a SaaS product the test should cover the web app, the API and any mobile apps, and above all tenant isolation. Testers should use two test tenants to check that one customer can never reach another’s data (<a class="link" href="https://flavoureak.co.uk/blog/multi-tenancy-explained/">multi-tenancy explained</a>).</p>
<p>Look for testers with recognised accreditation, such as CREST. Fix the findings, ask for a retest, and share a summary or letter of attestation with customers, keeping the full report for those who sign a non-disclosure agreement. Test again each year and after major changes.</p>
<h2>The basics to have in place first</h2>
<ol><li><b>MFA for all staff</b> on email, cloud hosting, code hosting, the billing provider and the domain registrar.</li><li><b>MFA and SSO for customers</b>, with <a class="link" href="https://flavoureak.co.uk/blog/sso-for-saas/">SAML or OpenID Connect</a> for larger ones.</li><li><b>Least privilege.</b> Production access only for the people who need it, with named accounts and no shared logins.</li><li><b>Encryption</b> in transit with TLS, and at rest for databases, files and backups.</li><li><b>Backups</b> that run automatically, sit in a separate account and are test-restored on a schedule. A backup that has never been restored is only a hope.</li><li><b>Patching</b> of servers, laptops and software dependencies, with automated alerts for known vulnerabilities.</li><li><b>Audit logs</b> of administrative and security actions (<a class="link" href="https://flavoureak.co.uk/blog/saas-audit-logs/">audit logs for SaaS</a>).</li><li><b>Joiners and leavers.</b> A checklist that removes every account on someone’s last day.</li></ol>
<h2>Incident response</h2>
<p>Write a short incident plan before you need one. It should say who leads, how to contain a problem, how to preserve evidence, who decides whether customers must be told and how they will be told. Practise it once with a realistic scenario.</p>
<p>For most customer data in a B2B SaaS, you are a processor acting for your customer, the controller. UK GDPR requires a processor to tell the controller about a personal data breach without undue delay. The controller must report a notifiable breach to the ICO within 72 hours of becoming aware of it where feasible. Your contract may set a fixed deadline for telling the customer, so know what you have signed.</p>
<h2>Data processing agreements</h2>
<p>Under Article 28 of UK GDPR, a controller that uses a processor must have a written contract with set terms. The processor must act only on the controller’s documented instructions, keep the data confidential, secure it, use sub-processors only with authorisation, help the controller meet its obligations, delete or return the data at the end and allow audits. The contract must also describe the processing, including the types of personal data and whose data it is.</p>
<p>Publish your own data processing agreement and sub-processor list, so customers can review them early. Large customers will often send their own instead. If personal data leaves the UK, for example to a hosting region abroad, the transfer also needs a lawful route, such as UK adequacy regulations for that country or the ICO’s International Data Transfer Agreement. Take legal advice on the agreement before your first enterprise contract.</p>
<div class="callout"><b>Getting ready for a larger customer?</b> <a href="https://flavoureak.co.uk/saas/">See how we build and run SaaS platforms</a> with these controls in place from the start.</div>
<h2>Questions people ask</h2>
<h3>Do I need Cyber Essentials to sell SaaS in the UK?</h3><p>It isn’t a legal requirement for selling software. UK central government requires it for some contracts, and some private-sector buyers ask for it in their questionnaires. It is also a practical checklist of basic controls, which makes it a sensible first certification for a small UK SaaS company.</p>
<h3>Should a UK SaaS choose ISO 27001 or SOC 2?</h3><p>Choose the one your customers ask for. UK and European enterprises more often recognise ISO 27001, and US enterprises more often ask for SOC 2. Many controls overlap, so a company that holds one can prepare for the other with less work than starting from scratch.</p>
<h3>How often should a SaaS product be penetration tested?</h3><p>Once a year is a common expectation in security questionnaires, plus a test after major changes such as a new API, a new mobile app or a change to how tenants are separated. Fix the findings and keep evidence of the retest to show customers.</p>
<h3>Can a small company answer a security questionnaire honestly?</h3><p>Yes. Answer what is true today, say which controls are planned, and give dates for them. Security teams review small suppliers regularly, and an honest gap with a plan is easier for them to accept than a claim that doesn’t hold up.</p>]]></content:encoded>
</item>
<item>
<title>Building event software that survives on-sale day</title>
<link>https://flavoureak.co.uk/blog/event-software-on-sale-day/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/event-software-on-sale-day/</guid>
<pubDate>Tue, 13 Oct 2026 08:00:00 GMT</pubDate>
<category>Events</category>
<description>Most event platforms are quiet for weeks, then face their busiest minute of the year. Here is how to build ticketing and check-in that hold up.</description>
<content:encoded><![CDATA[<ul><li>Put a waiting room in front of checkout, so buyers are let through at a pace the system can handle.</li><li>Hold tickets for a few minutes while people pay, and make sure a retried payment can never charge twice.</li><li>Download the guest list to scanners before doors open, so check-in keeps working when the venue Wi-Fi doesn’t.</li></ul><h2>The busiest minute of the year</h2>
<p>An event platform can sit quietly for weeks, then face tens of thousands of people in the minute tickets go on sale. Systems sized for an average day fall over, checkout slows, and buyers refresh the page, which makes it worse.</p>
<h2>A waiting room in front of checkout</h2>
<p>A virtual waiting room lets everyone in at once to a lightweight page, then releases buyers into checkout at a steady rate. It is fairer, because people are served in order, and it protects the parts of the system that take payments.</p>
<h2>Holds, not oversells</h2>
<p>When a buyer selects a ticket, it is held for a few minutes while they pay. If they abandon checkout, the hold expires and the ticket goes back on sale. Without holds you either oversell, or lose sales to people who never complete.</p>
<h2>Payments that can’t double-charge</h2>
<p>Under load, requests time out and get retried. Every payment request carries a unique key, so a retry is recognised as the same payment rather than a new one. Payment providers such as Stripe support this directly.</p>
<h2>Test the spike before the date</h2>
<p>Load testing with realistic traffic, a few weeks before a big on-sale, shows where the system bends. Static pages and cached content take most of the load off the parts that can’t be cached.</p>
<h2>Doors open, Wi-Fi down</h2>
<p>Venue Wi-Fi is often overloaded by the very crowd you are trying to check in. Scanners download the guest list before doors open and validate tickets on the device. Scans sync when a connection returns, and a ticket scanned at two entrances while both were offline is flagged as soon as they reconnect.</p>
<div class="callout"><b>Planning event software?</b> See <a href="https://flavoureak.co.uk/events/">what we build for events</a>.</div>]]></content:encoded>
</item>
<item>
<title>Multi-tenancy: how a SaaS keeps each customer’s data apart</title>
<link>https://flavoureak.co.uk/blog/multi-tenancy-explained/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/multi-tenancy-explained/</guid>
<pubDate>Sun, 11 Oct 2026 08:00:00 GMT</pubDate>
<category>SaaS</category>
<description>Every SaaS has to answer one question early: how is one customer’s data kept away from another’s? There are three common answers, each with different costs.</description>
<content:encoded><![CDATA[<ul><li>Most SaaS products share one database and tag every row with the customer it belongs to.</li><li>Separate schemas or separate databases give stronger isolation, at a higher running cost.</li><li>Whatever the model, enforce isolation in the database as well as in the application code.</li></ul><h2>The question</h2>
<p>A SaaS product serves many customers from one system. Each customer, usually called a tenant, must only ever see its own data. Getting this wrong is the most damaging bug a SaaS can have, so the approach should be decided before the first table is created.</p>
<h2>Three common models</h2>
<table><thead><tr><th>Model</th><th>How it works</th><th>Suits</th></tr></thead><tbody><tr><td>Shared tables</td><td>Every row carries a tenant ID</td><td>Most B2B SaaS, especially early on</td></tr><tr><td>Schema per tenant</td><td>Each customer gets its own set of tables in one database</td><td>Fewer, larger customers with custom needs</td></tr><tr><td>Database per tenant</td><td>Each customer gets a separate database</td><td>Strict data residency or contractual isolation</td></tr></tbody></table>
<h2>Shared tables, done safely</h2>
<p>Tagging rows with a tenant ID is cheap to run and simple to operate, but one forgotten filter in one query can leak data. We guard against that in two places: the application scopes every query to the current tenant, and the database enforces it too, for example with PostgreSQL row-level security, so a missed filter returns nothing instead of someone else’s records.</p>
<h2>When to isolate more</h2>
<p>Some customers require their data in a separate database, or in a specific region. Designing the data access layer so a tenant can be moved to its own database later keeps that option open without paying for it on day one.</p>
<h2>The parts people forget</h2>
<ul><li><b>Background jobs and exports</b> need the same tenant scoping as the web app.</li><li><b>Search indexes and caches</b> must be partitioned by tenant too.</li><li><b>File storage</b> should keep each tenant’s uploads under its own path, with access checked on every download.</li><li><b>Tests</b> should include a second tenant whose data must never appear.</li></ul>
<h2>Why decide early</h2>
<p>Changing the tenancy model after launch means migrating live customer data. Choosing deliberately at the start, and writing the reasons down, is one of the cheapest decisions in the whole build.</p>]]></content:encoded>
</item>
<item>
<title>What an AI feature costs per call, and why we work it out first</title>
<link>https://flavoureak.co.uk/blog/ai-cost-per-call/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/ai-cost-per-call/</guid>
<pubDate>Fri, 09 Oct 2026 08:00:00 GMT</pubDate>
<category>AI</category>
<description>A model call has a price you can calculate before you build anything. Here is the sum, where the money goes, and three ways to spend less.</description>
<content:encoded><![CDATA[<ul><li>Every model call costs (input tokens × input rate) + (output tokens × output rate).</li><li>The shape of the feature matters more than the rate: resending context and looping agents multiply cost.</li><li>Send less, route simple work to smaller models, and cache what repeats.</li></ul><h2>The sum</h2>
<p>Language models are billed by the token. A token is a piece of a word; in English text a token averages about four characters, so a thousand words is roughly 1,300 tokens. Every provider publishes two rates: one for the tokens you send, and a higher one for the tokens the model writes back.</p>
<div class="formula">cost per call = (input tokens × input rate) + (output tokens × output rate)</div>
<p>That is all there is to it, which means the cost of a feature can be worked out on paper before anyone writes a prompt.</p>
<h2>A worked example</h2>
<p>Take a feature that reads a 30-page tender and drafts a priced schedule. Thirty pages of dense text is around 15,000 words, or about 20,000 input tokens. The draft schedule it writes back might be 1,500 tokens. At example rates of £2 per million input tokens and £10 per million output tokens:</p>
<table><thead><tr><th></th><th>Tokens</th><th>Rate</th><th>Cost</th></tr></thead><tbody><tr><td>Input</td><td>20,000</td><td>£2 per million</td><td>£0.040</td></tr><tr><td>Output</td><td>1,500</td><td>£10 per million</td><td>£0.015</td></tr><tr><td><b>Per tender</b></td><td></td><td></td><td><b>£0.055</b></td></tr></tbody></table>
<p>At 400 tenders a month that is £22. If it saves an estimator two hours on each one, the model is the cheapest part of the feature. The rates here are examples; check your provider's current price list, because they change often.</p>
<h2>Where the money actually goes</h2>
<p>The rate is rarely the problem. The shape of the feature is.</p>
<ul><li><b>Resending context.</b> A chat assistant sends the whole conversation on every turn. By the twentieth message each call carries the nineteen before it, so the cost of a conversation grows much faster than its length.</li><li><b>Retrieval that over-fetches.</b> Pulling ten documents into every call to be safe can multiply input tokens tenfold for no better answer.</li><li><b>Loops.</b> An agent that plans, calls a tool, reads the result and plans again makes several model calls for one request. Each step is cheap. The loop is not.</li><li><b>Retries.</b> If an output fails validation and is generated again, you pay twice.</li></ul>
<h2>Three ways we bring it down</h2>
<h3>Send less</h3>
<p>Find the three pages of a tender that hold the bill of quantities and send those, not the whole document. Good retrieval is a cost control as much as a quality one.</p>
<h3>Route by difficulty</h3>
<p>Sorting an email into one of six categories doesn't need the largest model. A small, fast model handles routine work, and the large one is called only for drafting and reasoning.</p>
<h3>Cache what repeats</h3>
<p>Most providers discount input that repeats exactly between calls, such as a long set of instructions or a reference document. Put the fixed part of a prompt first and the changing part last.</p>
<h2>The cost of being wrong</h2>
<p>The last line in the sum is the one nobody writes down: what happens when the answer is wrong. A drafted quote that a person checks costs a few minutes. A drafted quote that goes straight to a customer costs whatever the mistake costs. We design the review screen at the same time as the prompt, because it decides whether the feature saves money at all.</p>
<div class="callout"><b>Try it with your own numbers.</b> The <a href="https://flavoureak.co.uk/ai/">AI guide</a> has a calculator for volumes and rates.</div>]]></content:encoded>
</item>
<item>
<title>How to design AI features for when the answer is wrong</title>
<link>https://flavoureak.co.uk/blog/designing-for-wrong-answers/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/designing-for-wrong-answers/</guid>
<pubDate>Thu, 08 Oct 2026 08:00:00 GMT</pubDate>
<category>AI</category>
<description>Every AI feature will sometimes be wrong. Design the screen for that day, with sources shown, outputs as editable drafts, facts checked in code and actions that can be undone.</description>
<content:encoded><![CDATA[<ul><li>A hallucination is when a language model produces a fluent, confident statement that is false or not supported by the information it was given.</li><li>Treat every AI output as a draft that a person can trace to its source, edit and reject, until testing shows a step is safe to automate.</li><li>Check numbers, prices, dates and permitted actions in code, so a model’s output is never the only thing between an error and a customer.</li></ul><p>Designing for wrong answers means building an AI feature so that when the model is mistaken, a person notices before it matters and can fix it in seconds. In practice that means showing where each answer came from, presenting outputs as editable drafts, checking key facts in code and making actions reversible. A hallucination is a fluent, confident output from a language model that is false or not supported by the information it was given.</p>
<h2>Why AI answers are sometimes wrong</h2>
<p>A language model writes by predicting likely next words, based on patterns learned in training and on whatever text it receives in the prompt. It doesn’t look facts up unless your software gives it a way to, and it has no reliable sense of when it doesn’t know. When the information it needs is missing, it tends to produce something that sounds right.</p>
<p>That is why a model asked to read a supplier invoice may return a plausible VAT number when the invoice doesn’t show one. Nothing in the output marks the invented value as different from the real ones beside it. Better models and better prompts reduce how often this happens. They don’t reduce it to zero, so the screen has to be designed for the cases that get through.</p>
<h2>Show where every answer came from</h2>
<p>The most useful single feature is a visible source. If an answer is built from documents or records, show the passage or field it used next to the claim, and let the person click through to it. This is one of the main practical benefits of retrieval, described in <a class="link" href="https://flavoureak.co.uk/blog/rag-explained/">RAG explained</a>.</p>
<p>Sources change how people review. A reviewer who can see that a £4,200 contract value came from page 3 of the uploaded tender can check it in two seconds. A reviewer shown only the number has to trust it or go hunting. Where an answer has no source, say so on the screen in words such as “not found in your documents” or “general answer, not from your records”.</p>
<p>For extraction features, a blank is better than a guess. If a field can’t be found in the source, leave it empty and highlight it. Code can enforce this by checking that each extracted value really appears in the source text before the field is filled.</p>
<h2>Review states and confidence</h2>
<p>Every AI output should carry a visible state, so nobody mistakes an unchecked draft for something a person has approved. A simple set of states covers most features.</p>
<table><thead><tr><th>State</th><th>What it means</th><th>What the screen does</th></tr></thead><tbody>
<tr><td>Draft</td><td>Generated by AI, not yet looked at</td><td>Labelled clearly as AI-generated; can’t be sent or posted</td></tr>
<tr><td>Needs attention</td><td>A check in code failed or a field is missing</td><td>Highlights the specific problem and where to fix it</td></tr>
<tr><td>Reviewed</td><td>A named person has checked and accepted it</td><td>Records who and when, and enables the next action</td></tr>
<tr><td>Sent or applied</td><td>It has had an effect outside the system</td><td>Shows the history and offers undo where possible</td></tr>
</tbody></table>
<p>Be careful with confidence scores. A model asked how sure it is will give a number, but that number is generated text like any other, and it is often a poor guide to accuracy. Better signals come from your own code. Did the value pass validation? Does it appear in the source? Do two independent methods agree? Do the line items add up to the total? Use those checks to decide what lands in “needs attention”.</p>
<h2>Editable drafts and undo</h2>
<p>The simplest protection is a person between the model and the outcome, with a screen that makes their job quick. An AI output should open as an editable draft, with editing as the obvious action and accept-and-send as a deliberate second step.</p>
<p>CurioKids works this way for lesson content. Content is generated for each child from their CEFR level, age and interests, and the teacher can change it. The model does the groundwork, and the teacher can change whatever doesn’t suit the child. The thinking behind that feature is in <a class="link" href="https://flavoureak.co.uk/blog/lesson-for-the-child/">how a lesson is shaped for each child</a>.</p>
<p>Undo matters most where AI acts directly. Keep previous versions of anything the AI changed, so one click restores the old text. Prefer soft deletes, which hide a record, to hard deletes, which destroy it. Where an action can’t be reversed, such as an email that has already been delivered, that is exactly where a human approval step belongs.</p>
<h2>Guardrails in code</h2>
<p>Guardrails are checks in your own software that the model can’t argue its way past. They sit between the model’s output and anything that happens next.</p>
<ul>
<li><b>Validate structure.</b> Require output in a defined format and reject anything that doesn’t match it.</li>
<li><b>Recalculate money.</b> Compute totals, VAT, CIS deductions and discounts in code from the line items, and never accept them from the model.</li>
<li><b>Check against your records.</b> A product code, customer or price the model mentions must exist in your database.</li>
<li><b>Allow-list actions.</b> The model can request only actions you have defined, and each request is checked against the user’s permissions.</li>
<li><b>Set limits.</b> Cap refund amounts, discount levels and message counts, so the worst case has a known size.</li>
<li><b>Filter what is shown.</b> Screen outputs for content that should never appear, especially in products used by children.</li>
</ul>
<h2>Feedback loops that improve the feature</h2>
<p>Thumbs up and thumbs down buttons are easy to add and easy to ignore. The richer signal is what people do with the output. Record whether each draft was accepted unchanged, edited or thrown away, and keep the edited version. Comparing the draft with what the person actually sent shows exactly where the model falls short, without asking anyone to fill in a form.</p>
<p>Ask for a reason only when someone rejects a draft, and make it one tap from a short list, such as “wrong facts”, “wrong tone” or “missing information”. Feed those cases into your evaluation set, so a fix for one failure is tested against everything that already worked. The process is set out in <a class="link" href="https://flavoureak.co.uk/blog/testing-ai-features/">how to test an AI feature</a>.</p>
<h2>How the screen should behave</h2>
<p>Pulling this together, an AI feature that handles wrong answers well follows a short list of rules.</p>
<ol>
<li>Label AI-generated content as AI-generated, in words.</li>
<li>Show the source next to each fact, or say plainly that there isn’t one.</li>
<li>Leave fields blank and highlighted when the source doesn’t contain them.</li>
<li>Keep the original input visible beside the output.</li>
<li>Make editing the obvious next step.</li>
<li>Never send, publish or charge without a deliberate action by a person.</li>
<li>Keep the history, and offer undo wherever the action allows it.</li>
</ol>
<div class="callout"><b>Adding AI to a product people rely on?</b> We design the review screen at the same time as the prompt, because it decides whether the feature can be trusted. <a class="link" href="https://flavoureak.co.uk/ai/">See how we build AI features</a>.</div>
<h2>Questions people ask</h2>
<h3>Can AI hallucinations be prevented completely?</h3>
<p>No current method removes them entirely. Better models, retrieval from your own documents and careful prompts make them less frequent. Design the feature so a wrong answer is visible, easy to correct and unable to cause harm on its own.</p>
<h3>Should we show a confidence score to users?</h3>
<p>Only if it comes from checks you can trust. A model’s own estimate of its confidence is often unreliable. Signals from code, such as whether a value was found in the source or whether totals reconcile, make better warnings.</p>
<h3>When is it safe to let AI act without review?</h3>
<p>When the action is low risk, easy to undo, and testing on real cases shows the feature is accurate enough for that job. Sorting emails into folders may qualify. Sending a quote or issuing a refund usually doesn’t.</p>
<h3>Who is responsible when an AI feature gets something wrong?</h3>
<p>In general, the business using the output remains responsible for what it tells customers or decides about them. That is a strong reason to keep a person in the loop for anything with legal, financial or safety consequences. Take legal advice where the stakes are high.</p>]]></content:encoded>
</item>
<item>
<title>What a B2B ordering portal does that a web shop can’t</title>
<link>https://flavoureak.co.uk/blog/b2b-ordering-portals/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/b2b-ordering-portals/</guid>
<pubDate>Wed, 07 Oct 2026 08:00:00 GMT</pubDate>
<category>E-commerce</category>
<description>Trade customers buy differently from consumers. A B2B portal gives them their own prices, terms and quick reordering, and takes the phone orders off your team.</description>
<content:encoded><![CDATA[<ul><li>Trade customers need their own price lists, credit terms and invoices on account, not a consumer checkout.</li><li>Quick reorder, bulk entry and approvals save time on both sides.</li><li>Some platforms now include B2B features. A custom portal makes sense when your pricing, terms or integrations go beyond them.</li></ul><h2>Trade customers buy differently</h2>
<p>A consumer finds a product, pays by card and leaves. A trade customer orders the same forty lines every fortnight, pays on 30-day terms, has prices agreed with your sales team, and may need a manager to approve anything over a set amount. Pushing them through a consumer checkout means phone calls, emails and manual re-entry.</p>
<h2>What a B2B portal includes</h2>
<ul><li><b>Customer-specific pricing.</b> Each account sees its own price list, discounts and minimum order quantities.</li><li><b>Credit terms.</b> Orders placed on account against a credit limit, with invoices instead of card payments.</li><li><b>Quick reorder.</b> Previous orders, saved lists, and bulk entry by product code or spreadsheet upload.</li><li><b>Approvals.</b> Buyers place orders; managers approve them above a set value.</li><li><b>Account management.</b> Several users per customer, delivery addresses, invoices and statements in one place.</li></ul>
<h2>Platform features or a custom portal?</h2>
<p>Some e-commerce platforms, such as Shopify Plus, now include B2B features like company accounts and price lists, and they are a good start. A custom portal earns its cost when your pricing rules are complex, when orders must flow into an ERP or accounts system in a specific way, or when the portal is part of a wider platform for reps, warehouses and deliveries.</p>
<h2>Connect it to the rest of the business</h2>
<p>A portal is only useful if stock, prices and orders stay in sync with the systems behind it. Integrations with accounting, stock and warehouse systems are where most of the work is, and where failures need handling: retries, alerts and a clear view of anything that needs a person.</p>
<h2>Start with your best customers</h2>
<p>Launch with a handful of trade accounts who order often, watch how they use it, and widen from there. Phone orders fall quickly once regular customers can reorder in a minute.</p>
<div class="callout"><b>Selling to trade customers?</b> See <a href="https://flavoureak.co.uk/ecommerce/">what we build for e-commerce</a>.</div>]]></content:encoded>
</item>
<item>
<title>Billing for a new SaaS: plans, trials and the edge cases</title>
<link>https://flavoureak.co.uk/blog/billing-for-saas/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/billing-for-saas/</guid>
<pubDate>Sun, 04 Oct 2026 08:00:00 GMT</pubDate>
<category>SaaS</category>
<description>Taking a card payment is easy. Running subscriptions is not: upgrades, failed payments, tax and cancellations all need deciding before launch.</description>
<content:encoded><![CDATA[<ul><li>Use a billing platform such as Stripe Billing rather than building subscription logic yourself.</li><li>Decide upgrades, downgrades, trials and failed-payment handling before launch, not after the first support ticket.</li><li>Keep your app’s record of each customer’s plan in sync with the billing platform through webhooks.</li></ul><h2>Don’t build billing from scratch</h2>
<p>Subscriptions involve proration, retries, invoices, receipts, tax and card updates. Platforms such as Stripe Billing handle these, along with a hosted customer portal where customers update cards and download invoices. Your product should own the plan logic and leave the money movement to the platform.</p>
<h2>Decisions to make before launch</h2>
<ul><li><b>Plans and limits.</b> What each plan includes, and what happens when a customer reaches a limit.</li><li><b>Trials.</b> Whether a card is needed to start, and what happens when the trial ends.</li><li><b>Upgrades and downgrades.</b> Whether changes take effect now with a prorated charge, or at the next renewal.</li><li><b>Failed payments.</b> How many retries, which emails go out, and when access is restricted.</li><li><b>Cancellations.</b> Whether access continues to the end of the paid period, and what happens to the data.</li><li><b>Tax.</b> VAT for UK customers, and the rules for selling abroad.</li></ul>
<h2>Keep the app and the billing platform in sync</h2>
<p>The billing platform is the source of truth for payments. Your app needs to know the result: which plan a workspace is on and whether it is paid up. That is done with webhooks, messages the platform sends when something changes, handled so that a repeated or late message never puts an account in the wrong state.</p>
<h2>Test the unhappy paths</h2>
<p>Test cards that decline, cards that expire mid-subscription, refunds, disputes, and a customer who upgrades and cancels on the same day. Billing bugs are expensive and very visible, so they get their own tests.</p>
<div class="callout"><b>This isn’t tax advice.</b> HMRC’s guidance is the authority on how VAT applies to what you sell.</div>]]></content:encoded>
</item>
<item>
<title>How to take over a stalled or failed software project</title>
<link>https://flavoureak.co.uk/blog/taking-over-a-stalled-project/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/taking-over-a-stalled-project/</guid>
<pubDate>Sat, 03 Oct 2026 08:00:00 GMT</pubDate>
<category>Working with us</category>
<description>Secure access to everything first, then commission a code and infrastructure review. Choose between stabilising, refactoring and rebuilding on what the review finds.</description>
<content:encoded><![CDATA[<ul><li>The first step in taking over a stalled project is securing access to the code, hosting, data, domains and app store accounts, before anything else changes.</li><li>A code and infrastructure review tells you whether the software builds, what state the data model and security are in, and what each option would cost.</li><li>Stabilise when the foundations are sound, refactor when the problems are local, and rebuild only when the data model or technology no longer fits the business.</li></ul><p>To take over a stalled or failed software project, first secure access to the code, infrastructure, data and accounts, then have an independent developer review what exists before you decide whether to stabilise, refactor or rebuild. A stalled project is one where development has stopped, or slowed so much that the software is not delivering what was paid for, often because a supplier has left, run out of budget or lost control of the code. The review is what turns a frustrating situation into a decision you can cost.</p>
<h2>Secure access before anything else</h2>
<p>When a relationship with a developer or agency ends badly, access is the most urgent problem and the easiest to lose. Ask for everything in writing, and change passwords and keys once you have control. You need all of the following.</p>
<ul><li>The source code repository, including its full commit history.</li><li>Hosting and cloud accounts, including the database, file storage and backups.</li><li>The domain registrar and DNS.</li><li>Apple Developer Program and Google Play Console access. For Android, find out who holds the upload key used to sign releases. If the app uses Play App Signing, Google holds the signing key and a lost upload key can be reset on request.</li><li>Third-party services such as payments, email sending, SMS, maps and AI providers.</li><li>Environment variables and secrets, the deployment pipeline and any servers someone logs into by hand.</li><li>Error monitoring, analytics, the issue tracker and design files.</li><li>The contract, the specification and any written agreement about ownership.</li></ul>
<p>If an account is in someone else’s name, ask for it to be transferred or for you to be made an owner. Check the contract to see whether the copyright in the code was assigned to you. Our guide to <a class="link" href="https://flavoureak.co.uk/blog/who-owns-the-code-uk/">who owns the code</a> explains why that matters.</p>
<h2>What a code and infrastructure review covers</h2>
<p>A review is an independent assessment of the software, its hosting and how changes reach users, written up with findings and costed options. It should answer the questions in this table.</p>
<table><thead><tr><th>Area</th><th>What the reviewer checks</th><th>What a good answer looks like</th></tr></thead><tbody><tr><td>Build</td><td>Whether the software builds and runs from a clean copy of the repository using the documentation</td><td>A new developer has it running locally on their first day</td></tr><tr><td>Completeness</td><td>What works against the specification, and what is missing or faked</td><td>Each feature marked as done, partial or missing</td></tr><tr><td>Data model</td><td>Whether the database structure matches how the business works</td><td>Tables and relationships that someone in the business would recognise</td></tr><tr><td>Code quality</td><td>Structure, duplication, consistency and tests</td><td>Clear modules, with tests on the rules that matter</td></tr><tr><td>Dependencies</td><td>Versions of the language, framework and packages, and known vulnerabilities</td><td>Supported versions, no critical known vulnerabilities</td></tr><tr><td>Security</td><td>Sign-in, permissions, secrets and the handling of personal data</td><td>Permissions checked on the server, no secrets in the code</td></tr><tr><td>Infrastructure</td><td>What is running, who pays for it, backups and environments</td><td>Infrastructure described in code, backups restored in a test</td></tr><tr><td>Delivery</td><td>How changes reach production</td><td>An automated pipeline that anyone on the team can run</td></tr></tbody></table>
<h2>Red flags to look for</h2>
<p>Some findings change the decision on their own. Price any of these into your options.</p>
<ul><li>The code cannot be built from the repository, or the repository has no history.</li><li>Passwords, API keys or private keys are committed in the code.</li><li>The production server has been edited directly, so the live system no longer matches the repository.</li><li>There are no backups, or nobody has ever restored one.</li><li>The framework or language version is past its end of life and no longer receives security fixes.</li><li>Permissions are enforced only in the user interface, so the API returns data to anyone who asks for it.</li><li>Business rules, such as pricing or tax, are copied into several places and disagree with each other.</li><li>Important accounts are tied to one person’s personal email address.</li></ul>
<h2>Stabilise, refactor or rebuild</h2>
<p>There are three broad options, and most takeovers use a mix of them.</p>
<h3>Stabilise</h3>
<p>Stabilising means making the existing software safe and dependable without changing its structure. Typical work includes fixing the build, adding monitoring and backups, patching vulnerabilities, writing tests around the most important behaviour and setting up automated deployment. It is the right first step in almost every takeover, because it reduces risk and buys time while you decide what else to do.</p>
<h3>Refactor</h3>
<p>Refactoring means improving the internal structure of code without changing what it does, one area at a time. It suits software whose foundations are sound but whose code has become hard to change. Users keep working throughout, and each improvement can ship on its own.</p>
<h3>Rebuild</h3>
<p>Rebuilding means writing a new system and moving users and data across to it. It makes sense when the data model is wrong for the business, when the technology is obsolete, or when so little works that there is nothing worth keeping. Rebuilds tend to take longer than expected, because the old system contains rules that nobody wrote down, and the data still has to be migrated. A common approach is to replace one part at a time, with the old and new systems running side by side until the old one can be switched off.</p>
<h2>How to decide</h2>
<p>Put the options side by side against the same questions. Working through them in order usually makes the answer clear.</p>
<ol><li><b>Does the data model fit the business?</b> If it does, the code on top of it can usually be improved. If it doesn’t, every new feature will fight it, and a rebuild of at least the data layer is likely.</li><li><b>Are the core rules correct?</b> Check pricing, tax, permissions and anything involving money against real examples from the business.</li><li><b>Is the technology supported?</b> A framework that still receives security updates can be upgraded in steps. One past its end of life sets a deadline.</li><li><b>Are there live users and data?</b> Live users make a rebuild slower and riskier, because data has to be migrated and nobody can be left without a working system.</li><li><b>What does the next year of changes cost under each option?</b> Compare the cost of building your planned features on the existing code with the cost of a rebuild plus those same features.</li></ol>
<p>Write the decision down with the reasons. If the situation changes, such as a larger customer arriving or a framework reaching end of life, you can revisit it with the same evidence.</p>
<h2>Protecting yourself next time</h2>
<p>Much of the damage in a failed project comes from losing visibility and control, and both can be designed in from the start. Keep every account in your company’s name and give the supplier access to it, instead of the other way round. Ask for access to the repository from the first week, so you can see work arriving. Pay against working software deployed to a test environment you can use, rather than against documents or estimates of progress. Agree a handover checklist in the contract, as described in <a class="link" href="https://flavoureak.co.uk/blog/you-get-the-repository/">what a proper handover includes</a>, and test it before the final payment.</p>
<p>When you choose the next supplier, ask how they report progress, who owns each account and what happens to the code if the relationship ends. Clear answers to those three questions prevent most of the problems in this article.</p>
<div class="callout"><b>Inherited a project that has stalled?</b> Secure the access and get a review before committing to a rebuild. <a class="link" href="https://flavoureak.co.uk/contact/">Talk to us about it</a>.</div>
<h2>Questions people ask</h2>
<h3>Should I rebuild a failed software project from scratch?</h3>
<p>Only when a review shows that the data model or technology no longer fits, or that too little works to be worth keeping. Rebuilds take longer than they appear to, because the old system holds rules nobody documented. Stabilising first, then refactoring or replacing one part at a time, is usually the lower-risk route.</p>
<h3>What if the previous developer will not hand over the code?</h3>
<p>Check your contract first, to see what it says about ownership and handover. If copyright was assigned to you, you are in a strong position to demand the code, and if it was not, your rights may be limited to an implied licence. Take legal advice before escalating, and secure every account you already control in the meantime.</p>
<h3>Can a new developer work with code written by someone else?</h3>
<p>Yes, and it is routine. Software written in a mainstream language and framework, with a repository and some documentation, can be taken over by any competent team. The review tells you how much work it will take to get comfortable with it.</p>
<h3>What is technical debt?</h3>
<p>Technical debt is the extra cost of future changes caused by shortcuts taken earlier, such as missing tests, duplicated logic or outdated dependencies. Like financial debt, a little can be a sensible way to ship sooner. A stalled project often carries so much of it that every change is slow and risky.</p>]]></content:encoded>
</item>
<item>
<title>Web app, mobile app, or both? How to decide</title>
<link>https://flavoureak.co.uk/blog/web-or-mobile-first/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/web-or-mobile-first/</guid>
<pubDate>Wed, 30 Sep 2026 08:00:00 GMT</pubDate>
<category>Working with us</category>
<description>Start with where the work happens and who does it. The answer is often a web app for the office and a mobile app for the people outside it.</description>
<content:encoded><![CDATA[<ul><li>Build for where the work happens: desks get a web app, sites and vans get a mobile app.</li><li>A mobile app earns its cost when people need the camera, offline working or notifications.</li><li>One backend can serve both, so starting with one doesn’t lock you out of the other.</li></ul><h2>Start with the work, not the device</h2>
<p>Ask who does the job, where they are when they do it, and what they need at that moment. The answer usually points to the right front end.</p>
<h2>When a web app is enough</h2>
<ul><li>People work at a desk, on a laptop or a large screen.</li><li>The job involves tables, reports, long forms or several things open at once.</li><li>You want changes live the moment you release them, with no app store review.</li></ul>
<p>A well-built web app also works in a phone's browser, which covers occasional use on the move.</p>
<h2>When you need a mobile app</h2>
<ul><li><b>No signal.</b> The work happens in basements, rural sites or vans, and must save offline.</li><li><b>Camera and files.</b> Photos, scans and signatures are part of the job.</li><li><b>Notifications.</b> People need to know about a new job or a change straight away.</li><li><b>Daily use.</b> An icon on the home screen gets used; a bookmarked web page often doesn't.</li></ul>
<h2>Why many businesses need both</h2>
<p>Estimark is a good example. The office plans, prices and invoices on a laptop. The site team records diaries, photos and safety checks on a phone. Both work on the same job record.</p>
<h2>One backend, two front ends</h2>
<p>The rules, the data and the permissions live in one place. The web app and the mobile apps are different windows onto it, so you can start with one and add the other later without rebuilding.</p>
<h2>Our usual advice</h2>
<p>Build first for the people whose work costs most to do by hand today. If that's the office, start with web. If it's the site, start with mobile.</p>]]></content:encoded>
</item>
<item>
<title>Audit logs for SaaS: what to record, how long to keep it and who sees it</title>
<link>https://flavoureak.co.uk/blog/saas-audit-logs/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/saas-audit-logs/</guid>
<pubDate>Sun, 27 Sep 2026 08:00:00 GMT</pubDate>
<category>SaaS</category>
<description>An audit log records who did what, to which record, when and from where, and can’t be changed afterwards. Security reviews ask for one, and it helps you show UK GDPR accountability.</description>
<content:encoded><![CDATA[<ul><li>An audit log is an append-only record of important actions in your product, showing who did it, what they did, to which record, when, from where and what changed.</li><li>Protect it from editing, keep it apart from day-to-day application data, and set a retention period you can justify, because audit logs contain personal data under UK GDPR.</li><li>Letting each customer’s admins search and export their own audit log answers common security questionnaire questions and supports UK GDPR’s accountability principle, which requires organisations to be able to demonstrate compliance.</li></ul><p>An audit log is a permanent, append-only record of the important actions taken in a software product, showing who acted, what they did, to which record, when, from where and what changed. In a B2B SaaS it lets a customer investigate a problem, lets you answer security reviews, and helps both of you show that personal data is handled properly under UK GDPR. It has to be designed in early, because actions that were never recorded can’t be reconstructed later.</p>
<h2>What an audit log is, and how it differs from other logs</h2>
<p>Products produce several kinds of record, and mixing them up causes problems. Application logs help developers debug and are usually deleted after days or weeks. Analytics events show how features are used. An audit log is evidence, so it is complete for the actions it covers and nobody can edit it.</p>
<table><thead><tr><th></th><th>Audit log</th><th>Application log</th><th>Product analytics</th></tr></thead><tbody><tr><td>Purpose</td><td>Evidence of who did what</td><td>Debugging and monitoring</td><td>Understanding how features are used</td></tr><tr><td>Read by</td><td>Customer admins and your security team</td><td>Your engineers</td><td>Your product team</td></tr><tr><td>Can entries change?</td><td>Never</td><td>Rarely edited, but often sampled or dropped</td><td>Often sampled or aggregated</td></tr><tr><td>Retention</td><td>Months to years, set deliberately</td><td>Days to weeks</td><td>Varies</td></tr><tr><td>Contains</td><td>Actor, action, target, time, source and changes</td><td>Errors, timings and stack traces</td><td>Events and their properties</td></tr></tbody></table>
<h2>What to record for each event</h2>
<ul><li><b>Who.</b> The user ID, their name and role at that moment, and the tenant. If the actor is an API key, an integration, a scheduled job or one of your own support staff acting on a customer’s behalf, record that too.</li><li><b>What.</b> A named action such as <code>invoice.approved</code>, <code>user.role_changed</code> or <code>export.created</code>, and whether it succeeded or was refused.</li><li><b>Which record.</b> Its type and ID, plus a readable label such as the invoice number.</li><li><b>When.</b> A server timestamp in UTC, never the device’s own clock.</li><li><b>Where.</b> The IP address, the device or browser, whether the action came through the web app, a mobile app or the API, and a request ID that links to the application logs.</li><li><b>Before and after.</b> The old and new values of each changed field.</li></ul>
<p>Store the name and role as they were at the time. If you only store a user ID and the person is later renamed or deleted, the entry stops making sense.</p>
<h2>Which actions to log</h2>
<p>Log actions that affect security, money, access or large amounts of data. Logging every page view makes the log too large to search and collects personal data you don’t need. A sensible starting list follows.</p>
<ol><li>Sign-ins, failed sign-ins, password resets and changes to multi-factor authentication.</li><li>Users invited, removed or given a different role, and records shared with people outside the customer’s company.</li><li>Changes to single sign-on, security settings, API keys and integrations.</li><li>Exports, bulk downloads and bulk deletions.</li><li>Approvals and changes involving money, such as invoices, payments and price changes.</li><li>Any access to the customer’s account by your own support staff.</li></ol>
<p>Viewing especially sensitive records, such as children’s details or health information, can be worth logging too, even though ordinary reads are usually left out.</p>
<h2>Making audit logs tamper-evident</h2>
<p>An audit log that an administrator can quietly edit is worth very little. These measures make it trustworthy.</p>
<ul><li><b>Append-only by design.</b> The code has no update or delete path, and the database account the application uses can only insert into the audit table.</li><li><b>Written with the change.</b> Record the audit entry in the same database transaction as the action it describes, so one can’t exist without the other.</li><li><b>Copied somewhere separate.</b> Send entries to separate storage with write-once settings, such as object storage with an object lock, held in a different account from the main application.</li><li><b>Chained.</b> Each entry can include a hash of the previous one, so any change or deletion breaks the chain and shows up when the chain is checked.</li></ul>
<h2>How long to keep audit logs</h2>
<p>For most businesses, no UK law sets a single retention period for audit logs. UK GDPR’s storage limitation principle says personal data should be kept no longer than necessary for its purpose, and audit logs contain personal data such as names and IP addresses. So choose a period, write down why, and apply it automatically.</p>
<p>One workable approach is to keep a period of entries searchable in the product, such as 12 months, and move older entries to cheaper archive storage for longer if customer contracts require it. Records with their own legal retention rules, such as the accounting records HMRC expects a business to keep, are a separate question from how long you keep the log of who edited them.</p>
<h2>Showing audit logs to customers</h2>
<p>Enterprise customers expect their admins to see their own audit log. Give them a screen that filters by person, action, record and date, an export to CSV and, for larger customers, an API or a feed into their own security monitoring tools. Viewing the log should be a permission of its own (<a class="link" href="https://flavoureak.co.uk/blog/saas-roles-permissions/">roles and permissions</a>), and every query must be limited to that customer’s tenant.</p>
<p>A history tab on individual records works well alongside it. Someone looking at an invoice can see who created, edited and approved it without opening the full log.</p>
<h2>How audit logs help with security reviews and UK GDPR</h2>
<p>Security questionnaires commonly ask whether you log administrative actions, how long you keep logs, who can access them and how they are protected from tampering. A properly designed audit log answers all four in a sentence each (<a class="link" href="https://flavoureak.co.uk/blog/saas-security-before-enterprise/">security before your first enterprise customer</a>).</p>
<p>Under Article 5(2) of UK GDPR, the accountability principle, a controller must be able to demonstrate compliance with the data protection principles. Under Article 32, both controllers and processors must have appropriate security measures. Audit logs support both, and they become essential after a personal data breach. A controller must report a notifiable breach to the ICO within 72 hours of becoming aware of it where feasible, and must document every breach, reported or not. The audit log is often the quickest way to work out which records were touched and by whom.</p>
<p>For most customer data in a B2B SaaS, you act as a processor for your customer, who is the controller. Your contract will usually require you to tell them about a breach without undue delay, and your audit log is what lets you tell them what happened.</p>
<div class="callout"><b>Adding audit logs to an existing product?</b> <a href="https://flavoureak.co.uk/contact/">Get in touch</a> and we’ll look at what your data model needs.</div>
<h2>Questions people ask</h2>
<h3>Does UK GDPR require an audit log?</h3><p>UK GDPR doesn’t name audit logs as a requirement. It does require appropriate security and the ability to demonstrate compliance, and audit logs are one of the most practical ways to do both. Customers’ contracts and security questionnaires often ask for them directly.</p>
<h3>Should an audit log include IP addresses?</h3><p>Usually yes, because the IP address helps show where an action came from during an investigation. IP addresses can be personal data, so mention them in your privacy information, protect them like the rest of the log and delete them when the retention period ends.</p>
<h3>Can a user ask for their audit log entries to be deleted?</h3><p>A person can ask for their personal data to be erased under UK GDPR, and the right has exceptions, for example where data is needed to comply with a legal obligation or for legal claims. Whether an exception applies to your audit log depends on its purpose and your contracts, so take legal advice and record your reasoning.</p>
<h3>Should reads be logged as well as changes?</h3><p>Logging every read makes the log enormous, so most teams log changes and security events and leave ordinary reads out. Log reads where the data is especially sensitive, or where a customer needs to know who viewed it, such as children’s records or exported reports.</p>]]></content:encoded>
</item>
<item>
<title>GEO: how to become the source an AI answer cites</title>
<link>https://flavoureak.co.uk/blog/geo-become-the-source/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/geo-become-the-source/</guid>
<pubDate>Fri, 25 Sep 2026 08:00:00 GMT</pubDate>
<category>GEO</category>
<description>AI answer engines search, read a handful of pages and quote passages from them. What makes a passage worth quoting is mostly clarity, specifics and corroboration.</description>
<content:encoded><![CDATA[<ul><li>Answer engines turn a question into searches, read a few results and cite the passages they used.</li><li>Pages that state facts plainly, early and specifically are easier to quote.</li><li>Being described the same way on other sites matters as much as your own pages.</li></ul><h2>What changed</h2>
<p>A growing share of searches end in a written answer instead of a list of links. ChatGPT search, Perplexity, Microsoft Copilot, Claude and Google's AI Overviews work in broadly the same way: they turn the question into searches, read some of the results and write an answer that cites a few of them.</p>
<p>Generative Engine Optimisation, or GEO, is the work of being one of the pages that gets read and cited. It overlaps heavily with SEO, because a page that isn't found can't be quoted. The difference is in what happens after the page is found.</p>
<h2>How an answer gets assembled</h2>
<ol><li><b>The question becomes searches.</b> "Which builders' software handles CIS?" might become three or four separate queries.</li><li><b>Pages are retrieved</b>, usually from a search index, sometimes by fetching them directly.</li><li><b>Passages are ranked.</b> The engine looks for the paragraphs that answer the question, not the page as a whole.</li><li><b>The answer is written</b> from those passages.</li><li><b>Sources are cited</b>, usually the passages that contributed most.</li></ol>
<p>The unit that matters is the passage. A page can rank well and still never be quoted if its answer is buried in the fourth paragraph behind a story.</p>
<h2>What makes a passage quotable</h2>
<ul><li><b>The answer comes first.</b> Open each section with the direct answer, then explain.</li><li><b>It stands alone.</b> A paragraph that only makes sense after the one before is hard to lift out.</li><li><b>It is specific.</b> "Fixed price after a free survey, most installs take two to four days" beats "competitive pricing". Dates, units, places and numbers survive summarising.</li><li><b>It says where facts come from.</b> A figure with a source is more likely to be repeated than one without.</li><li><b>It uses the words people ask with.</b> Write headings as the questions customers actually send you.</li></ul>
<h2>Be the same entity everywhere</h2>
<p>Answer engines need to know that a trading name, a registered company and a domain are the same thing. Use one name and one description consistently, mark the site up with Organization structured data, and link to the profiles that confirm you exist, such as your Companies House record.</p>
<h2>Corroboration</h2>
<p>Engines are cautious about repeating a claim that only one site makes about itself. Listings in the directories and trade bodies of your industry, reviews on independent sites and mentions in other people's articles all make your own claims safer to quote.</p>
<h2>Measuring it</h2>
<p>There is no rank to check, so measure directly. Keep a list of the twenty questions customers ask before they buy. Run them through the main answer engines every month and record who is cited. In analytics, watch referrals from chatgpt.com, perplexity.ai and copilot.microsoft.com.</p>
<div class="callout"><b>Nobody controls what an answer engine cites.</b> Anyone promising a guaranteed place in ChatGPT's answers is guessing. What you can control is whether your pages are easy to find, easy to understand and safe to quote.</div>]]></content:encoded>
</item>
<item>
<title>Google AI Overviews: what they are and what they change for your site</title>
<link>https://flavoureak.co.uk/blog/google-ai-overviews/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/google-ai-overviews/</guid>
<pubDate>Mon, 21 Sep 2026 08:00:00 GMT</pubDate>
<category>GEO</category>
<description>AI Overviews are generated summaries at the top of some Google results, with links to sources. Simple questions get answered on the page, and clear, trustworthy pages get cited.</description>
<content:encoded><![CDATA[<ul><li>Google AI Overviews are AI-generated summaries that appear above some search results and link to the web pages they draw on.</li><li>Google says there are no special requirements to appear in AI Overviews beyond being indexed and eligible to show a snippet in Google Search.</li><li>Questions with short factual answers are often answered on the results page itself, so business pages gain more from depth, first-hand detail and clear next steps than from quick definitions.</li></ul><p>Google AI Overviews are AI-generated summaries that appear at the top of some Google search results, answering the query in a few paragraphs with links to the web pages they used. For business sites they mean fewer clicks on simple factual questions and a chance to be cited as a source on others. To adapt, write pages that are worth citing and that give people a reason to visit.</p>
<h2>What AI Overviews are</h2>
<p>An AI Overview is a block of generated text that Google shows for some queries, usually above the traditional list of results. It summarises an answer drawn from several sources and shows links to those sources alongside or within the text. Google launched AI Overviews in the US in May 2024 and extended them to the UK and other countries later that year.</p>
<p>Google decides for each query whether to show an overview, and says it does so where it judges the summary adds something to the normal results. Overviews are typically seen on informational questions, and less often on searches for a particular website or brand. Google also offers AI Mode, a separate conversational search experience built on similar technology, where people can ask follow-up questions.</p>
<h2>How sources get chosen and shown</h2>
<p>Google’s guidance for site owners says that to be eligible as a supporting link in an AI Overview, a page must be indexed and eligible to be shown in Google Search with a snippet. There is no separate submission, markup or technical requirement. The usual SEO fundamentals apply. Google must be able to crawl the page, the content must be in text it can read, and the page must meet Google’s spam policies.</p>
<p>Google also describes a technique it calls query fan-out, in which AI Overviews and AI Mode may run several related searches across subtopics to build a response. A page can therefore be cited for one part of an answer even if it would not rank near the top for the original query. The links shown are the pages that support particular statements in the summary.</p>
<h3>Controls site owners have</h3>
<p>AI Overviews are part of Google Search, so they follow the same controls as normal snippets. The <code>nosnippet</code> robots directive, the <code>data-nosnippet</code> attribute and <code>max-snippet</code> limit what Google can show from a page, including in AI features. A page marked <code>noindex</code> cannot appear at all. The Google-Extended token in robots.txt has no effect on AI Overviews; it governs whether content is used for some of Google’s other AI products, such as Gemini models. Blocking content from snippets also reduces how it appears in ordinary results, so use these controls with care.</p>
<h2>What changes for business sites</h2>
<table>
<thead><tr><th>Type of query</th><th>What tends to happen</th><th>What to do</th></tr></thead>
<tbody>
<tr><td>Simple definition or fact</td><td>Answered in the overview, so fewer clicks</td><td>Answer it well anyway, because it builds citations and recognition</td></tr>
<tr><td>How-to with several steps</td><td>Summarised, with links for the detail</td><td>Give the full process, examples and edge cases on the page</td></tr>
<tr><td>Comparison or decision</td><td>Overview lists factors and cites sources</td><td>Publish clear comparisons, tables and criteria</td></tr>
<tr><td>Local or service search</td><td>Local results and business listings still play a large part</td><td>Keep your business profile and service pages accurate</td></tr>
<tr><td>Brand or product name</td><td>Usually leads straight to the official site</td><td>Make sure your own pages describe you clearly</td></tr>
</tbody>
</table>
<p>The broad pattern is that a question with a short, settled answer is less likely to send a visitor than before. A page that only restates a definition will lose clicks to the overview. A page that offers what the overview cannot, such as detailed price drivers, worked examples, tools, first-hand experience or a way to get a quote, still gives people a reason to visit.</p>
<p>Being cited also has value without a click. An overview that names your business or links to your guide puts you in front of the person asking, and some of them will search for you directly later. That effect doesn’t show up as a referral, which is why measurement needs more than one source.</p>
<h2>How to adapt your content</h2>
<ol>
<li><b>Answer first.</b> Start each page and section with a direct answer and a one-sentence definition, as set out in <a class="link" href="https://flavoureak.co.uk/blog/writing-for-ai-answers/">how to write pages that AI answers quote</a>.</li>
<li><b>Go further than the overview.</b> Add the detail a summary leaves out, such as worked examples, price drivers, local conditions, exceptions and what to do next.</li>
<li><b>Show first-hand experience.</b> Photos of your own work, real processes, named authors and specific cases are hard for a summary to replace, and they show the page comes from people who do the work.</li>
<li><b>Keep facts checkable.</b> Give units, dates and sources. Update pages when rules or prices change, and show the update date.</li>
<li><b>Make the next step obvious.</b> A visitor who clicks through from an overview has already read a summary. Take them straight to the detail, the calculator or the contact form.</li>
<li><b>Keep the technical basics right.</b> That means crawlable pages, fast loading, sensible internal links and structured data that matches the visible content.</li>
</ol>
<h2>What to avoid</h2>
<p>Don’t produce large numbers of thin pages in the hope of matching every possible question. Google’s spam policies cover scaled content created mainly to manipulate rankings, however it is produced. Don’t add structured data that describes content which isn’t on the page. And don’t rewrite every page around a guess about what the overview will say, because overviews change from day to day and differ between users.</p>
<p>Be careful with pages written only to be summarised. If a page contains nothing beyond what an overview can say in three sentences, there is little reason for Google to cite it or for a person to click it.</p>
<h2>Measuring the effect</h2>
<p>Google counts clicks and impressions from AI Overviews in Search Console’s Performance report under the Web search type, together with ordinary results, so there is no separate AI Overview figure to read. Watch for queries where impressions hold steady or rise while clicks fall, then search those queries yourself to see whether an overview appears and whether you are cited. Branded searches and enquiries that mention Google are other useful signals. <a class="link" href="https://flavoureak.co.uk/blog/measuring-ai-search-traffic/">Measuring traffic from AI search</a> sets out a monthly routine.</p>
<div class="callout"><b>Reviewing a site for AI search?</b> Our <a class="link" href="https://flavoureak.co.uk/seo/">SEO guide</a> covers the technical basics that decide whether Google can use your pages at all.</div>
<h2>Questions people ask</h2>
<h3>How do I get my website into Google AI Overviews?</h3>
<p>There is no way to opt in or submit a page to AI Overviews. Google says a page needs to be indexed and eligible to appear in Search with a snippet, and it may then be chosen as a supporting link. Clear, specific, trustworthy content on a technically sound site gives a page the best chance.</p>
<h3>Can I stop Google using my content in AI Overviews?</h3>
<p>Google applies its normal snippet controls to AI Overviews. The nosnippet directive, the data-nosnippet attribute and the max-snippet setting limit what can be shown, and noindex removes a page from Search entirely. These controls also affect how the page appears in ordinary results, so the trade-off needs thought.</p>
<h3>Do I need special structured data for AI Overviews?</h3>
<p>Google says no special structured data is needed to appear in AI Overviews. Structured data still helps search engines understand a page and can make it eligible for rich results. It should always match the content visible on the page.</p>
<h3>Are AI Overviews the same as AI Mode?</h3>
<p>AI Overviews appear within normal Google search results for some queries. AI Mode is a separate, conversational search experience where people can ask follow-up questions. Both draw on Google’s index, and Google’s guidance for site owners covers both.</p>]]></content:encoded>
</item>
<item>
<title>How to brief a software agency, and what to ask them</title>
<link>https://flavoureak.co.uk/blog/how-to-brief-an-agency/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/how-to-brief-an-agency/</guid>
<pubDate>Fri, 18 Sep 2026 08:00:00 GMT</pubDate>
<category>Working with us</category>
<description>A good brief describes the work, not the features. A good agency answers five questions without hesitating.</description>
<content:encoded><![CDATA[<ul><li>Describe the work done by hand today, who does it and what it costs, rather than a list of features.</li><li>Ask who will build it, who owns the code, and what happens after launch.</li><li>Ask to see software the agency runs, not only software it delivered.</li></ul><h2>What to put in the brief</h2>
<ul><li><b>The work.</b> What is done by hand today, step by step, and by whom.</li><li><b>The cost.</b> Hours a week, mistakes, missed payments, lost jobs. A rough number is enough.</li><li><b>The people.</b> Who will use the software, where they are, and what they use now.</li><li><b>The rules.</b> Approvals, tax, regulations, deadlines. Anything that must be right.</li><li><b>The systems.</b> Accounting, payments, email and anything else it must connect to.</li><li><b>The timing.</b> Any date that genuinely matters, and why.</li></ul>
<p>Leave out the feature list if you can. Agencies are good at turning problems into features, so give them the problem.</p>
<h2>Five questions to ask any agency</h2>
<ol><li><b>Who will actually build it?</b> Will you speak to them, or to an account manager?</li><li><b>Who owns the code and the accounts?</b> The answer should be you, from day one.</li><li><b>How often will we see working software?</b> Every week or two is healthy. A big reveal at the end is a risk.</li><li><b>What happens after launch?</b> Who fixes it at 8am on a Monday, and what does that cost?</li><li><b>What do you run yourselves?</b> Software an agency runs for paying customers tells you more than a portfolio of screenshots.</li></ol>
<h2>Warning signs</h2>
<ul><li>A price before anyone has asked how the work is done today.</li><li>No mention of data, permissions or what happens when something fails.</li><li>Code and app store accounts held in the agency's name.</li><li>AI added to the proposal with no cost per use and no review step.</li></ul>
<div class="callout"><b>Ready to write yours?</b> Our <a href="https://flavoureak.co.uk/contact/">brief builder</a> asks these questions for you.</div>]]></content:encoded>
</item>
<item>
<title>Internal linking for SEO: how to connect the pages on your website</title>
<link>https://flavoureak.co.uk/blog/internal-linking/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/internal-linking/</guid>
<pubDate>Tue, 15 Sep 2026 08:00:00 GMT</pubDate>
<category>SEO</category>
<description>Internal links help search engines find pages, understand them and judge which matter most. Group pages into hubs and spokes, use descriptive anchor text and leave no page orphaned.</description>
<content:encoded><![CDATA[<ul><li>Internal links are links from one page on a website to another page on the same website, and they are a main way search engines discover and understand pages.</li><li>A hub and spoke structure links a broad overview page to detailed pages on the same topic, and links those pages back to the hub and to each other.</li><li>Use descriptive anchor text that says what the linked page covers, and make sure every page you want found has links from other relevant pages.</li></ul><p>Internal linking is the practice of linking the pages of your own website to each other. It matters for SEO because links are how search engines find pages, learn what they are about and judge which ones are most important. An internal link is a link from one page to another on the same domain, and a deliberate structure of them is one of the few SEO improvements entirely in your own control.</p>
<h2>Why internal links matter for SEO</h2>
<p>Internal links do three separate jobs for search engines, and one more for readers.</p>
<h3>Discovery</h3>
<p>Search engines find pages mainly by following links. A page linked from your navigation or from several articles is found and recrawled quickly. A page listed only in a sitemap can still be found, but search engines have less reason to treat it as important.</p>
<h3>Context</h3>
<p>The words in and around a link tell search engines what the target page covers. A link reading “CIS deductions explained” on a page about construction invoicing says far more than any number of links reading “read more”. Google’s guidance on links asks for anchor text that is descriptive, reasonably concise and relevant to both the page it sits on and the page it points to.</p>
<h3>Importance</h3>
<p>Links pass signals of importance between pages, an idea that goes back to Google’s original PageRank. A page linked from the home page and from many relevant pages is likely to be treated as more important than one buried five clicks deep with a single link. You can’t control who links to you from other sites, but you decide entirely how links flow around your own.</p>
<p>For readers, good internal links answer the next question before they ask it, which keeps them on the site and moves them towards the pages that matter to the business.</p>
<h2>Hub and spoke structure</h2>
<p>A hub and spoke structure, sometimes called a topic cluster, groups pages by subject. The hub is a broad page covering the topic as a whole. The spokes are detailed pages on specific parts of it.</p>
<p>Take a builder’s website. A hub page on loft conversions gives an overview and links to spokes on planning permission, building regulations, dormer and roof-light options, typical costs and a gallery of finished projects. Each spoke links back to the hub. Spokes also link to each other where it helps the reader, so the planning permission page links to the building regulations page.</p>
<p>The structure gives search engines a clear picture of what the site covers in depth, and gives readers an obvious next step from any page. Each hub can then sit in the main navigation, so the most important pages are a single click from anywhere on the site.</p>
<h2>Types of internal link compared</h2>
<table><thead><tr><th>Link type</th><th>Where it appears</th><th>What it does well</th><th>Watch out for</th></tr></thead><tbody>
<tr><td>Main navigation</td><td>Every page</td><td>Signals your most important pages</td><td>Too many items dilute it and confuse visitors</td></tr>
<tr><td>Breadcrumbs</td><td>Top of inner pages</td><td>Shows hierarchy and links up to parent pages</td><td>Must match the real structure of the site</td></tr>
<tr><td>Contextual links</td><td>Inside body text</td><td>Carries specific context through its anchor text</td><td>Add them only where they help the reader</td></tr>
<tr><td>Related content</td><td>End of articles or product pages</td><td>Connects pages within the same topic</td><td>Automated lists can pick weak matches</td></tr>
<tr><td>Footer</td><td>Every page</td><td>Reaches contact, policy and legal pages</td><td>Long footers of keyword links look manipulative</td></tr>
</tbody></table>
<h2>Writing good anchor text</h2>
<p>Anchor text is the clickable text of a link. Write it so a person knows where the link goes before they click.</p>
<ul>
<li>Describe the destination, such as “our guide to CIS deductions”, and avoid “click here”.</li>
<li>Keep it short, usually a few words within a natural sentence.</li>
<li>Vary the wording naturally. The same exact keyword phrase in every link to a page looks forced.</li>
<li>Avoid linking the same words to different pages across the site, which sends mixed signals.</li>
<li>Make links real <code>a</code> elements with an <code>href</code>, so crawlers can follow them, as covered in <a class="link" href="https://flavoureak.co.uk/blog/seo-for-javascript-sites/">SEO for JavaScript sites</a>.</li>
</ul>
<h2>How many internal links a page should have</h2>
<p>There is no fixed number, and Google doesn’t set one. The practical limit is the reader’s attention. A link should exist because someone reading that paragraph might genuinely want to go there.</p>
<p>As a rule of thumb, a long article might link to between three and ten closely related pages in its body text, plus the site navigation. Important pages should be reachable within about three clicks of the home page. Pages that take six or seven clicks to reach tend to be crawled less often and look less important.</p>
<p>Don’t add <code>nofollow</code> to internal links to steer importance towards other pages. It doesn’t concentrate value on the remaining links, and it can make the target pages harder to discover. Keep <code>nofollow</code> for links you genuinely don’t vouch for, such as links in user comments.</p>
<h2>Finding orphan pages and auditing links</h2>
<p>An orphan page is a page that no other page on your site links to. It may still be in your sitemap, or receive visits from adverts or old emails, but search engines and visitors browsing the site can’t reach it. Old landing pages, products added without a category and blog posts never linked from anywhere are the usual cases. Here is how to find them and fix the rest of the structure at the same time.</p>
<ol>
<li>Crawl the site with a site crawler, starting from the home page. Every page it finds can be reached by links.</li>
<li>Export the URLs from your sitemaps, your analytics landing pages and Search Console’s Performance report.</li>
<li>Compare the lists. URLs in the second set that the crawl didn’t find are orphans.</li>
<li>For each orphan, link to it from its hub and from related pages, or remove it and redirect it if it no longer earns its place.</li>
<li>From the crawl, fix broken internal links and links that pass through redirects, so every link points straight at the final URL.</li>
<li>Check click depth and the number of internal links to each page. Important pages with few links or a deep position need more links from relevant pages.</li>
<li>Review the Links report in Search Console, which lists your most linked pages internally, and check the order matches your priorities.</li>
</ol>
<p>Repeat the audit every few months, and after any redesign, when links are often broken or lost. Link checks are part of <a class="link" href="https://flavoureak.co.uk/blog/redesign-without-losing-traffic/">redesigning a site without losing traffic</a>. Sites that generate many pages from data have the most to gain from a clear structure, as explained in <a class="link" href="https://flavoureak.co.uk/blog/programmatic-seo-saas/">programmatic SEO for SaaS</a>.</p>
<div class="callout"><b>Building or rebuilding a website?</b> We plan the page structure and the links between pages before the design, so every page has a place and a route to it. <a class="link" href="https://flavoureak.co.uk/seo/">See our SEO guide</a>.</div>
<h2>Questions people ask</h2>
<h3>What is the difference between internal and external links?</h3>
<p>Internal links point to pages on the same website. External links point to pages on other websites. You control your internal links completely, which makes them one of the simplest SEO improvements to make.</p>
<h3>Do internal links help SEO?</h3>
<p>Yes. They help search engines discover pages, understand what each page is about through the anchor text, and judge which pages are most important. A page with no internal links is hard for search engines to find and value.</p>
<h3>What is an orphan page?</h3>
<p>An orphan page is a page with no internal links pointing to it from elsewhere on the same site. Visitors browsing the site can’t reach it, and search engines may find it only through a sitemap or an external link. Fix it by linking to it from relevant pages, or remove it and redirect the URL.</p>
<h3>Should every page link to the home page?</h3>
<p>Most sites already do this through the logo and the main navigation, and that is enough. There is no need to add extra links to the home page in body text.</p>]]></content:encoded>
</item>
<item>
<title>Offline first: building a mobile app for a basement with no signal</title>
<link>https://flavoureak.co.uk/blog/offline-first-mobile/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/offline-first-mobile/</guid>
<pubDate>Fri, 11 Sep 2026 08:00:00 GMT</pubDate>
<category>Engineering</category>
<description>On a building site the signal drops exactly when someone needs to record something. Here is how an app keeps working, and keeps people’s trust, without a connection.</description>
<content:encoded><![CDATA[<ul><li>Save to the phone first, every time, and sync when a connection returns.</li><li>Create record IDs on the device, so work can exist before the server knows about it.</li><li>Decide how conflicts resolve for each type of data, and never drop someone’s work silently.</li></ul><h2>The basement</h2>
<p>An operative is in a plant room two floors underground. They photograph a weeping valve, note the time and add it to the site diary. There is no signal. If the app shows a spinner and then an error, they will write it on the back of a delivery note instead, and the software has lost.</p>
<p>Estimark is used in exactly these places, so working offline was never a later feature. These are the rules we build to.</p>
<h2>Save to the phone first</h2>
<p>Every change is written to a database on the device before anything else happens. The screen reads from that local database, so it is instant whether or not there is a connection. A separate process sends queued changes to the server when it can.</p>
<p>This reverses the usual order. Most apps ask the server and then show the result. An offline-first app shows the result and tells the server later.</p>
<h2>IDs made on the device</h2>
<p>If the server hands out IDs, nothing can be created offline, because nothing has an ID yet. We generate IDs on the phone, so a diary entry, its photos and the snag it raises can all refer to each other before any of them has been uploaded.</p>
<h2>Photos are the heavy part</h2>
<p>Text syncs in milliseconds; photos don't. They are compressed on the device, queued separately from the records that refer to them, uploaded in pieces that can resume, and shown with their own progress. A diary entry can appear in the office while its photos are still arriving.</p>
<h2>Conflicts are a product decision</h2>
<p>Two people edit the same record while both are offline. Something has to win, and the right answer depends on the data:</p>
<ul><li><b>Diary entries and photos</b> are append-only. Nothing is overwritten, so nothing conflicts.</li><li><b>Statuses</b> such as "snag closed" take the latest change, with both changes kept in the history.</li><li><b>Money</b> never merges automatically. A conflicting change to a valuation goes to a person.</li></ul>
<h2>Tell the person the truth</h2>
<p>"Saved" must mean saved. The app says "Saved to phone, 3 changes waiting to sync" when that is the case, and shows when they have reached the office. People trust an app that is honest about its connection far more than one that pretends.</p>
<h2>Test it like a site</h2>
<p>We test with the network switched off, with a connection that drops every few seconds, and by force-closing the app halfway through a sync. If it survives those, it survives a basement.</p>]]></content:encoded>
</item>
<item>
<title>How to release mobile app updates safely on iPhone and Android</title>
<link>https://flavoureak.co.uk/blog/releasing-app-updates/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/releasing-app-updates/</guid>
<pubDate>Thu, 10 Sep 2026 08:00:00 GMT</pubDate>
<category>Engineering</category>
<description>Roll out gradually, hide new features behind flags, keep the API compatible with old versions and watch crash reports, because an installed app cannot be rolled back.</description>
<content:encoded><![CDATA[<ul><li>You cannot roll back a mobile app that people have installed, so safe releases depend on staged rollouts, feature flags and an API that supports older versions.</li><li>The App Store’s phased release spreads automatic updates over seven days, and Google Play’s staged rollout lets you choose the percentage and halt it.</li><li>A minimum supported version, checked by the app when it starts, lets you require an update when an old version can no longer work safely.</li></ul><p>To release mobile app updates safely, roll each version out gradually with the stores’ staged release tools, hide risky changes behind feature flags, keep your API compatible with older app versions, and watch crash reports before widening the rollout. A staged rollout is a release that reaches a growing share of users over several days, so a problem affects a few people before it affects everyone. All of this matters because an app that is already on someone’s phone cannot be taken back.</p>
<h2>Why mobile releases are different from the web</h2>
<p>When a web release goes wrong, you deploy the previous version and every user has it within minutes. Mobile apps do not work that way. Each version is reviewed by the store, downloaded to each phone and updated when the person or their phone decides. At any moment your users are running several versions, some of them months old, and some people turn automatic updates off entirely.</p>
<p>Three principles follow. Limit how many people get a new version until you trust it. Control new behaviour from the server, where you can change it instantly. Keep the server working for every version still in use.</p>
<h2>Version numbers and build numbers</h2>
<p>Both platforms use two numbers. The version is what people see in the store, such as 2.4.1. The build number is an internal counter that must go up with every build you upload. On iOS these are the version string and the build number, and on Android they are versionName and versionCode, where versionCode is a whole number that must be higher for every release you upload to Google Play.</p>
<p>Most teams use semantic versioning for the visible number, with major, minor and patch parts. Raise the patch number for fixes, the minor number for new features and the major number for large changes. Keep the version the same on iPhone and Android for the same release, so support conversations are not confusing.</p>
<h2>Staged rollouts on the App Store and Google Play</h2>
<p>Both stores let you release to a share of users first. They work differently, and neither lets you take a version back.</p>
<table><thead><tr><th></th><th>App Store phased release</th><th>Google Play staged rollout</th></tr></thead><tbody><tr><td>How it works</td><td>Automatic updates reach a growing share of existing users over seven days</td><td>A percentage of users you choose is offered the update, and you raise it when ready</td></tr><tr><td>Schedule</td><td>Fixed at 1%, 2%, 5%, 10%, 20%, 50%, then 100%</td><td>Any percentage, at any pace</td></tr><tr><td>Pausing</td><td>Can be paused for up to 30 days in total, or released to everyone at once</td><td>Can be halted, then resumed or replaced by a new release</td></tr><tr><td>Who gets it early</td><td>New downloads, and anyone who updates manually from the store</td><td>A random selection of users, up to the percentage you set</td></tr><tr><td>Going back</td><td>Not possible; ship a new version with the fix</td><td>Not possible; halting stops more people getting it, and a fix needs a new release with a higher versionCode</td></tr></tbody></table>
<p>Use the time the rollout gives you. Watch crash reports, server error rates and support messages at each step, and agree in advance what would make you pause. On Google Play, a common pattern is to start small, wait a day or two, then widen. Apple offers expedited review for urgent fixes, but it is not guaranteed, so keep it out of the plan. The first submission has its own hurdles, covered in <a class="link" href="https://flavoureak.co.uk/blog/app-store-review-first-time/">passing app review first time</a>.</p>
<h2>Feature flags and kill switches</h2>
<p>A feature flag is a setting, held on your server or in a remote configuration service, that turns a piece of functionality on or off without a new release. It separates shipping code from switching it on. A new feature can go out in a release switched off, be turned on for staff, then for a small group of customers, then for everyone. If something goes wrong, it is switched off in seconds for every version that contains the flag.</p>
<p>A kill switch is a flag for something that already works, such as a payment method or a third-party integration, so it can be turned off if a provider has an outage. Keep flags tidy. Remove each one once a feature is fully live, or the code fills with branches nobody remembers.</p>
<h2>Forced updates and minimum versions</h2>
<p>Sometimes an old version has to stop working, because it has a security flaw or depends on an API you need to retire. Neither store forces updates on your behalf, so the app has to do it.</p>
<ol><li>When the app starts, it asks your server for the minimum supported version and, optionally, a recommended version.</li><li>If the installed version is below the minimum, the app shows a screen that explains why, links to the store and does nothing else.</li><li>If it is below the recommended version, the app shows a prompt that can be dismissed.</li><li>On Android, Google Play’s in-app updates API can download and install the update from inside the app, either straight away or in the background while the person carries on.</li></ol>
<p>Build the check into the first release, because a version that shipped without one can never be made to show it. Use forced updates rarely, since they interrupt people at a moment they did not choose, and test that the update screen itself works on the oldest supported version.</p>
<h2>Keeping the API working for old versions</h2>
<p>Every version still in use is a client of your API. A change that is safe for the latest app can break one that is three months old.</p>
<ul><li>Make changes additive. Add new fields and endpoints, and leave the old ones in place.</li><li>Never change the meaning of an existing field. Add a new one instead.</li><li>Remove something only after the minimum supported version no longer uses it.</li><li>Change the database in two steps, often called expand and contract. Add the new structure, move data and clients across, and remove the old structure in a later release.</li><li>Have the app send its version with each request, so the server can tell which versions are still active.</li></ul>
<p>Offline-first apps need extra care. Estimark saves to the phone first and syncs later, and an app built that way can reconnect after days with changes queued by an older version. The server has to accept those changes, and each new version has to migrate the database on the phone without losing anything that has not yet synced. Our guide to <a class="link" href="https://flavoureak.co.uk/blog/offline-first-mobile/">offline-first mobile apps</a> covers the sync side.</p>
<h2>Crash monitoring and release checks</h2>
<p>Crash reporting tools, such as Firebase Crashlytics or Sentry, capture each crash with the device, operating system and app version. Apple’s Xcode Organizer and Google Play’s Android vitals report crashes and freezes too. Google Play publishes thresholds for crash and “app not responding” rates, and apps above them may be shown less prominently in the store.</p>
<p>Compare the crash-free rate of the new version with the previous one at each stage of the rollout. A short checklist keeps releases routine.</p>
<ol><li>Test the release build on real devices, including the oldest supported operating system.</li><li>Upgrade from the previous version with real data on the device, as well as testing fresh installs.</li><li>Confirm that new features are behind flags and switched off.</li><li>Confirm that the API changes are already live and work with the previous app version.</li><li>Release to a small percentage and watch crashes, errors and support messages.</li><li>Widen the rollout in steps, then switch features on gradually.</li></ol>
<p>Each release also adds to the yearly cost of running an app, which is covered in <a class="link" href="https://flavoureak.co.uk/blog/software-maintenance-costs/">software maintenance costs</a>.</p>
<div class="callout"><b>Shipping an app on both stores?</b> Plan the versioning, feature flags and minimum-version check before the first release. <a class="link" href="https://flavoureak.co.uk/mobile/">See how we approach mobile apps</a>.</div>
<h2>Questions people ask</h2>
<h3>Can I roll back an app update on the App Store or Google Play?</h3>
<p>No. Neither store lets you restore a previous version for people who have already updated. To undo a bad release, you ship a new version, often built from the previous code, with a higher build number. Feature flags and staged rollouts reduce how often you need to.</p>
<h3>How long does Apple’s phased release take?</h3>
<p>Seven days, if it is not paused. Automatic updates reach 1%, 2%, 5%, 10%, 20% and 50% of existing users on successive days, then everyone on day seven. You can pause it for up to 30 days in total, or release to everyone at any point.</p>
<h3>How do I force users to update my app?</h3>
<p>Build a minimum version check into the app. When it starts, it asks your server which versions are still supported and shows an update screen if the installed version is too old. On Android, Google Play’s in-app updates API can run the update from inside the app.</p>
<h3>What is a feature flag?</h3>
<p>A feature flag is a setting controlled from your server that turns a feature on or off without releasing a new version. It lets you ship code switched off, turn it on for a small group first and switch it off instantly if something goes wrong.</p>]]></content:encoded>
</item>
<item>
<title>Native or cross-platform? How we choose for a mobile app</title>
<link>https://flavoureak.co.uk/blog/native-or-cross-platform/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/native-or-cross-platform/</guid>
<pubDate>Fri, 04 Sep 2026 08:00:00 GMT</pubDate>
<category>Engineering</category>
<description>Cross-platform frameworks have closed most of the gap with native apps. The choice now comes down to what the app needs from the phone, and who will maintain it.</description>
<content:encoded><![CDATA[<ul><li>Cross-platform frameworks such as React Native and Flutter share one codebase across iPhone and Android.</li><li>Native Swift and Kotlin still win for heavy device features, complex animation or platform-specific design.</li><li>Choose from the app’s needs and from who will maintain it after launch.</li></ul><h2>The two approaches</h2>
<p><b>Native</b> means writing the iPhone app in Swift and the Android app in Kotlin: two codebases, each using the platform's own tools. <b>Cross-platform</b> means one codebase, in a framework such as React Native or Flutter, that produces both apps.</p>
<h2>Where cross-platform fits</h2>
<ul><li>Business apps built from forms, lists, photos and sync.</li><li>Teams that want one codebase to maintain, often alongside a web app in the same language.</li><li>Features that need to reach iPhone and Android users at the same time.</li></ul>
<h2>Where native earns its cost</h2>
<ul><li>Heavy use of the camera, Bluetooth, background location or sensors.</li><li>Complex animation, graphics, or real-time audio and video.</li><li>Apps that must feel exactly like the platform, or adopt new OS features on day one.</li></ul>
<h2>What matters more than the framework</h2>
<p>Offline storage, sync, error reporting, release automation and a sensible app store process decide whether a mobile app works in the field. That work is the same in any framework.</p>
<h2>How we decide</h2>
<p>We list what the app needs from the phone, who will maintain it, and what the web side is written in. Then we pick the approach that keeps the app reliable and the maintenance simple, and write the reasons down so you can check them later.</p>]]></content:encoded>
</item>
<item>
<title>AI agents in business software: what they can and can’t do reliably</title>
<link>https://flavoureak.co.uk/blog/ai-agents-in-business-software/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/ai-agents-in-business-software/</guid>
<pubDate>Fri, 04 Sep 2026 08:00:00 GMT</pubDate>
<category>AI</category>
<description>An AI agent is a model that chooses tools in a loop until a task is done. Narrow agents with few tools and an approval step work well today. Open-ended ones still fail in costly ways.</description>
<content:encoded><![CDATA[<ul><li>An AI agent is a language model that repeatedly chooses a tool, reads the result and decides the next step until it finishes or reaches a limit.</li><li>Agents are reliable today on narrow, checkable tasks with a few tools, such as looking up records, drafting replies and proposing schedules for a person to approve.</li><li>Give an agent its own limited permissions, require human approval before it spends money or contacts a customer, and cap its steps, time and cost.</li></ul><p>An AI agent is a language model that works in a loop, choosing a tool to use, reading the result and deciding what to do next until the task is finished. In business software, agents are reliable today when the job is narrow, the tools are few and a person approves anything that matters. They are unreliable when the goal is vague, the chain of steps is long, or a mistake can’t be undone.</p>
<h2>What an AI agent actually is</h2>
<p>A tool, in this sense, is a function your software makes available to the model, with a name, a description and defined inputs. “Look up a customer by email”, “list free slots for an engineer next week” and “create a draft invoice” are all tools. The model can’t run them itself. It asks your code to run one, your code checks the request and runs it, and the result goes back to the model.</p>
<p>An agent repeats that cycle. Asked to find out why order 4471 hasn’t shipped and draft a reply to the customer, it might look up the order, then the stock level, then the courier status, then write the draft. Nobody wrote those four steps in advance. The model chose them as it went, based on what each step returned.</p>
<p>That freedom is what separates an agent from the two simpler patterns most AI features use.</p>
<table><thead><tr><th>Pattern</th><th>Who decides the steps</th><th>Good for</th><th>Main risk</th></tr></thead><tbody>
<tr><td>Single model call</td><td>Your code, one step</td><td>Summarising, classifying, drafting from given text</td><td>A wrong answer</td></tr>
<tr><td>Workflow</td><td>Your code, fixed steps with the model inside some of them</td><td>Repeatable processes such as reading a document and filing it</td><td>Rigid when inputs vary a lot</td></tr>
<tr><td>Agent</td><td>The model, one step at a time</td><td>Tasks where the next step depends on what was found</td><td>Wrong actions, loops, cost and time</td></tr>
</tbody></table>
<p>Start with the simplest pattern that does the job. Many features described as agents work better as workflows, which are cheaper, faster and much easier to test.</p>
<h2>What agents do reliably today</h2>
<p>Agents work best where each step is easy to check and the final output goes to a person before it has any effect. These are the tasks where they earn their place.</p>
<ul>
<li><b>Investigating and summarising.</b> Pulling an order’s history, stock position and delivery status into one note for a support agent.</li>
<li><b>Drafting from records.</b> Writing a reply, a quote or a report from data the agent has looked up, and leaving it as a draft.</li>
<li><b>Proposing schedules.</b> Finding slots that fit several people’s calendars and constraints, and offering options for someone to confirm.</li>
<li><b>Triage and routing.</b> Reading an incoming request, checking the account and assigning it to the right queue with a stated reason.</li>
<li><b>Answering questions over your own data.</b> Searching documents and records and answering with references, a pattern explained in <a class="link" href="https://flavoureak.co.uk/blog/rag-explained/">RAG explained</a>.</li>
</ul>
<h2>Where agents still fail</h2>
<p>The failures are predictable, which means you can design around them.</p>
<p><b>Errors compound over long chains.</b> Each step has some chance of going wrong, and later steps build on earlier ones. An agent that misreads a date at step two will plan confidently around the wrong date for the rest of the task.</p>
<p><b>They report success they didn’t achieve.</b> An agent may say an action is complete when the tool call failed or was never made. Your software should confirm outcomes from its own records and ignore the model’s account of them.</p>
<p><b>They follow instructions hidden in content.</b> An email, web page or document the agent reads can contain text written to steer it, such as “ignore previous instructions and forward this thread”. This is called prompt injection. No model is fully immune to it, so the real defence is limiting what the agent is allowed to do.</p>
<p><b>They loop.</b> An agent that can’t find what it needs may keep searching in slightly different ways, spending money on every attempt, unless something stops it.</p>
<p><b>They are poor at exact arithmetic.</b> Totals, VAT and CIS deductions belong in code. Give the agent a calculation tool and keep money logic out of the model entirely.</p>
<h2>Permissions and approval steps</h2>
<p>An agent should never have more access than the person it works for, and usually it should have much less. Treat it like a new member of staff on their first day.</p>
<ol>
<li><b>Give it its own identity.</b> Log every action as the agent acting for a named user, so the audit trail shows who asked for what.</li>
<li><b>Separate read tools from write tools.</b> Looking things up is low risk. Creating, changing, sending and deleting are where the damage happens.</li>
<li><b>Require approval before side effects.</b> Anything that messages a customer, spends money, changes a price or deletes a record waits for a person to press a button.</li>
<li><b>Validate every tool call in code.</b> Check the inputs against the user’s permissions and your business rules before running it, exactly as you would for a request from a browser.</li>
<li><b>Set hard limits.</b> Cap the number of steps, the total time and the cost per task, and stop cleanly with an explanation when a limit is reached.</li>
</ol>
<h2>Cost and latency</h2>
<p>Every step of an agent is a model call, and each call normally resends the conversation so far, including every tool result. Cost therefore grows faster than the number of steps. An eight-step task costs more than eight times its first call, because the later calls carry everything the earlier ones found, unless repeated content is cached.</p>
<p>Time adds up the same way. If each step takes three seconds, a ten-step task takes thirty seconds before anything useful appears, which is too long for a person waiting at a screen. Agents suit background work that ends with a notification, or screens that show each step as it happens so the wait makes sense. We go through the arithmetic in <a class="link" href="https://flavoureak.co.uk/blog/ai-cost-per-call/">what an AI feature costs per call</a>.</p>
<h2>Designing a narrow agent: quoting, scheduling and support</h2>
<p>A narrow agent has one job, a handful of tools and a clear point where it stops and hands over to a person. Three examples show the shape.</p>
<h3>Quoting</h3>
<p>A builder receives an enquiry for a bathroom refit. A quoting agent reads it, looks up the relevant items in the price book and the labour rates, checks similar past jobs and produces a draft quote with every line traced to its source. It has no tool for sending. A person reviews the quantities, adjusts the margin and sends it. Construction software such as Estimark already holds a price book and labour rates, which is the kind of structured data an agent like this depends on.</p>
<h3>Scheduling</h3>
<p>A scheduling agent receives a request to book a site visit or a lesson. It checks availability, travel, existing commitments and any rules about notice, then offers two or three options. A customer or a member of staff makes the booking by choosing one. Time zone conversion belongs in code. CurioKids shows bookings in each family’s own time zone, and that kind of conversion should never be left to a model.</p>
<h3>Support</h3>
<p>A support agent reads a ticket, identifies the customer and order, checks delivery and payment status, and drafts a reply using the facts it found. It can prepare a refund for approval, only up to a limit set in code. Tickets it can’t place go to a person with a summary of what it checked.</p>
<p>Before any of these go live, run them against a set of real past cases and check both the final output and the steps taken. The method is set out in <a class="link" href="https://flavoureak.co.uk/blog/testing-ai-features/">how to test an AI feature</a>.</p>
<div class="callout"><b>Considering an agent for your product?</b> We’d start by checking whether a fixed workflow would do the job, and add agent steps only where they pay for themselves. <a class="link" href="https://flavoureak.co.uk/ai/">See how we approach AI features</a>.</div>
<h2>Questions people ask</h2>
<h3>What is the difference between an AI agent and a chatbot?</h3>
<p>A chatbot answers messages in a conversation. An agent uses tools to take steps, such as looking up records or creating drafts, and decides which step to take next. A chatbot can have an agent behind it, but many chatbots only generate text.</p>
<h3>Are AI agents safe to use with customer data?</h3>
<p>They can be, if the agent has limited permissions, every tool call is checked in code and actions with real effects need human approval. The same UK GDPR duties apply as for any processing of personal data, including a processor contract with the AI provider.</p>
<h3>How long does an AI agent take to complete a task?</h3>
<p>It depends on the number of steps, because each step is a separate model call. A simple lookup and draft might take a few seconds, and a long investigation can take a minute or more. Long-running agents work best in the background, with a notification when they finish.</p>
<h3>Can an AI agent replace a member of staff?</h3>
<p>Not for a whole role today. Agents handle narrow, well-defined tasks well and struggle with ambiguity, judgement and long chains of work. They are most useful taking repetitive groundwork off a person who then makes the decision.</p>]]></content:encoded>
</item>
<item>
<title>What it costs to maintain software after launch, and how to budget</title>
<link>https://flavoureak.co.uk/blog/software-maintenance-costs/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/software-maintenance-costs/</guid>
<pubDate>Sat, 29 Aug 2026 08:00:00 GMT</pubDate>
<category>Working with us</category>
<description>Software keeps costing money after it goes live, for hosting, services, security updates, upgrades, support and small improvements. Here is what drives the bill and how to plan for it.</description>
<content:encoded><![CDATA[<ul><li>Running costs after launch come from hosting, third-party services, security and dependency updates, operating system and app store changes, monitoring, support and small improvements.</li><li>A common rule of thumb is to budget 15 to 20 per cent of the original build cost each year, but a bottom-up estimate for your own system is more reliable.</li><li>Mobile apps add a yearly cycle of work, because Apple and Google raise their technical requirements and each new operating system changes how apps behave.</li></ul><p>Software maintenance is the work and spending needed to keep an application running, secure and useful after it launches, and it starts on the day the software goes live. The main costs are hosting, third-party services, security and dependency updates, operating system and app store changes, monitoring, support and small improvements. A common rule of thumb puts the yearly total at 15 to 20 per cent of the original build cost, but your own figure depends on what the software does and how it was built.</p>
<h2>Why software costs money after launch</h2>
<p>Software sits on top of things that keep moving. The framework it is written in releases new versions and eventually stops supporting old ones. Browsers, iOS and Android change every year. Security flaws are found in packages that were safe when you launched. Third-party services change their APIs and retire old versions. None of this is visible to your users until something breaks.</p>
<p>The other half is the business itself. Once people use the software, they find things it should do differently, and processes change. A system that nobody changes for two years is usually a system people have started working around.</p>
<p>Much of the long-term cost is set before launch, by choices about platforms and features. Our guide to <a class="link" href="https://flavoureak.co.uk/blog/what-drives-app-cost/">what drives app cost</a> covers the build side.</p>
<h2>The cost drivers</h2>
<p>This table sets out where maintenance money goes, what makes each item larger and how often it comes up.</p>
<table><thead><tr><th>Cost driver</th><th>What it covers</th><th>What makes it bigger</th><th>How often</th></tr></thead><tbody><tr><td>Hosting and infrastructure</td><td>Servers or serverless functions, databases, file storage, backups, content delivery</td><td>Traffic, stored files such as photos, uptime requirements, extra environments for testing</td><td>Monthly</td></tr><tr><td>Third-party services</td><td>Payments, email and SMS sending, maps, search, AI models, video</td><td>Usage-based pricing that grows with your customers, per-seat tools</td><td>Monthly</td></tr><tr><td>Security updates</td><td>Patching known vulnerabilities in packages and operating systems</td><td>Number of dependencies, how far behind they are</td><td>Weekly to monthly</td></tr><tr><td>Dependency and framework upgrades</td><td>Moving to supported versions of languages, frameworks and libraries</td><td>Large version jumps left too long</td><td>Several times a year</td></tr><tr><td>Operating system and app store changes</td><td>New iOS and Android versions, new store requirements</td><td>Number of platforms, use of device features</td><td>Yearly, with deadlines</td></tr><tr><td>Monitoring and incident response</td><td>Error tracking, uptime checks, logs, someone responding to alerts</td><td>Out-of-hours cover, strict response times</td><td>Continuous</td></tr><tr><td>Support</td><td>Answering users, investigating bugs, correcting data</td><td>Number of users, complexity of the workflows</td><td>Weekly</td></tr><tr><td>Small improvements</td><td>Changes users ask for, new reports, adjusted business rules</td><td>How fast the business changes</td><td>Monthly</td></tr><tr><td>Renewals and compliance</td><td>Domains, developer programme memberships, data protection requests</td><td>Number of accounts, products and markets</td><td>Yearly</td></tr></tbody></table>
<h2>Mobile apps and the yearly platform cycle</h2>
<p>Mobile apps carry a cost that websites mostly avoid. Apple and Google release a major operating system version every year, and each one can change permissions, background behaviour or how screens are laid out. Both stores also raise their minimum technical requirements on a schedule.</p>
<ul><li><b>Google Play</b> requires new apps and updates to target a recent Android API level, and the required level rises each year. Apps that fall too far behind stop being available to new users on newer devices.</li><li><b>Apple</b> periodically requires apps submitted to the App Store to be built with a recent version of Xcode and the iOS SDK. An app that cannot be built with current tools cannot ship an update, including an urgent fix.</li><li><b>Store policies</b> change too. Requirements for in-app account deletion, privacy labels and Google’s Data safety form have all arrived over the past few years, and existing apps had to comply.</li></ul>
<p>Estimark runs on the web, iPhone and Android, so every one of these cycles applies to it. For any product with apps in both stores, plan for platform work every year whether or not you add a feature.</p>
<h2>How to budget for maintenance</h2>
<h3>Start with the rule of thumb</h3>
<p>The figure you will hear most often is 15 to 20 per cent of the original build cost each year. It is a rule of thumb passed around the industry. No study stands behind it, and it can be wrong in both directions. It is useful as a sense check. If your software cost £80,000 to build, the rule suggests £12,000 to £16,000 a year, or about £1,000 to £1,333 a month.</p>
<h3>Then build it from the bottom up</h3>
<p>A better estimate adds up the drivers for your own system.</p>
<ol><li>List every paid service with its current monthly bill, then estimate how each grows as you add customers.</li><li>Count the engineering days needed each year for security patches, upgrades and platform changes. A web app with few dependencies needs fewer than a product with apps in both stores.</li><li>Estimate support time from the number of users and how often they get in touch.</li><li>Decide how much improvement you want each month, in days.</li><li>Add a contingency for the unexpected, such as a vulnerability that needs an urgent fix or a provider retiring an API you depend on.</li></ol>
<p>Compare the total with the rule of thumb. If yours is far lower, check whether you have left out improvements.</p>
<h3>Separate keeping the lights on from improving the product</h3>
<p>Split the budget in two. Keeping the lights on covers hosting, services, security, upgrades, monitoring and bug fixes, and it is not optional. Improvement covers new reports, changes to workflows and new features. It is where the software keeps earning its place, and it is the part most easily cut when money is tight. A budget that covers only the first tends to fall short, because users always ask for changes.</p>
<h2>What makes maintenance cheaper</h2>
<p>Most maintenance cost is decided during the build. These choices reduce it for years afterwards.</p>
<ul><li><b>Fewer dependencies.</b> Every package is something to update. Add one only when it saves real work.</li><li><b>Automated tests.</b> An upgrade is cheap when a test suite tells you within minutes whether anything broke, and expensive when someone has to click through every screen.</li><li><b>Small, regular upgrades.</b> Moving one version at a time every few months is far easier than jumping several versions after years of neglect.</li><li><b>Managed services.</b> A managed database with automatic backups usually costs more per month than a server you run yourself, and much less in engineering time.</li><li><b>Monitoring from day one.</b> Error tracking tells you about a problem before a customer does, while it is still small.</li><li><b>Ownership and documentation.</b> If the code, accounts and runbook are yours, you can change who maintains the software without paying someone to rediscover how it works.</li></ul>
<h2>Ways to pay for maintenance</h2>
<p>Maintenance is usually paid for in one of three ways. A monthly retainer buys a set amount of time and often an agreed response time for problems. Pay-as-you-go work is billed when you need it, which suits stable software but gives no guarantee that anyone is available when something breaks. An in-house developer gives continuity, though one person may not cover security, infrastructure and both mobile platforms. Whichever you choose, write down who responds to an outage, how quickly and at what hours. The trade-offs are similar to those in <a class="link" href="https://flavoureak.co.uk/blog/agency-freelancer-or-in-house/">choosing between an agency, a freelancer or an in-house team</a>.</p>
<div class="callout"><b>Planning a build?</b> Ask for an estimate of the yearly running costs at the same time as the build price. <a class="link" href="https://flavoureak.co.uk/businesses/">See how we plan software for businesses</a>.</div>
<h2>Questions people ask</h2>
<h3>How much does it cost to maintain an app each year?</h3>
<p>There is no single figure, because it depends on hosting, services, platforms and how much you want to change. A common rule of thumb is 15 to 20 per cent of the build cost each year. Use it as a sense check, and build a bottom-up estimate from your own services and engineering time.</p>
<h3>Can I stop paying for maintenance once the software is finished?</h3>
<p>You can stop adding features, but you cannot safely stop hosting, security updates or platform changes. Unpatched software becomes a security risk, and a mobile app that falls behind store requirements can no longer be updated and may disappear for new users. Software that people use is never finished in that sense.</p>
<h3>Why does a mobile app need updating when nothing has changed?</h3>
<p>The platforms it runs on have changed. Apple and Google release new operating systems every year and raise the minimum requirements for apps in their stores. An app that is not updated gradually stops meeting them, and eventually cannot be updated at all without a larger piece of work.</p>
<h3>Is hosting the biggest maintenance cost?</h3>
<p>Often it is not. For many business applications, engineering time for updates, support and improvements costs more than the servers. Hosting becomes significant when an application stores large files, processes video or calls AI models at high volume.</p>]]></content:encoded>
</item>
<item>
<title>CIS, reverse charge and payment notices in construction software</title>
<link>https://flavoureak.co.uk/blog/construction-rules/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/construction-rules/</guid>
<pubDate>Fri, 28 Aug 2026 08:00:00 GMT</pubDate>
<category>Construction</category>
<description>Most of the work in Estimark is rules, not screens. Three of them change what every payment in UK construction has to say.</description>
<content:encoded><![CDATA[<ul><li>Under CIS, contractors deduct 20% (registered) or 30% (unregistered) from the labour part of a subcontractor’s payment, or nothing for gross payment status.</li><li>Since 1 March 2021, the VAT domestic reverse charge moves VAT accounting to the customer for many construction services.</li><li>Under the Construction Act, payment notices and their dates decide what is owed, so the software has to get them right.</li></ul><h2>Why rules, not screens</h2>
<p>Anyone can draw an invoice screen. What makes a builder trust software with money is that the invoice is correct: the right deduction, the right VAT treatment, the right dates. In UK construction, three sets of rules shape almost every payment.</p>
<h2>The Construction Industry Scheme</h2>
<p>Under CIS, a contractor paying a subcontractor deducts money from the payment and passes it to HMRC as an advance on the subcontractor's tax. The rate depends on the subcontractor's status:</p>
<table><thead><tr><th>Subcontractor status</th><th>Deduction</th></tr></thead><tbody><tr><td>Registered for CIS</td><td>20%</td></tr><tr><td>Not registered</td><td>30%</td></tr><tr><td>Gross payment status</td><td>0%</td></tr></tbody></table>
<p>The deduction applies to labour, not to materials the subcontractor supplied, so every invoice has to separate the two. The contractor verifies each subcontractor with HMRC, files a monthly return and gives the subcontractor a statement of what was deducted. Software that gets this wrong creates a reconciliation problem every month.</p>
<h2>The VAT domestic reverse charge</h2>
<p>Since 1 March 2021, for many construction services supplied between VAT-registered businesses that report under CIS, the customer accounts for the VAT instead of the supplier. The invoice states that the reverse charge applies, and the supplier doesn't charge the VAT.</p>
<p>The exception that causes the most mistakes is the end user. If the customer is the final client and not a contractor, normal VAT applies, so the software has to know where each customer sits in the supply chain.</p>
<h2>Payment notices and the Construction Act</h2>
<p>The Housing Grants, Construction and Regeneration Act 1996, as amended in 2009, sets out how payments under construction contracts are notified. A payment notice states what will be paid. A pay less notice has to be given in time if the payer intends to pay less. If the notices are missing or late, the sum the payee applied for can become the sum that is due.</p>
<p>For software, that means dates are calculated from the contract terms, reminders go out before deadlines, and every notice is stored with the time it was sent.</p>
<h2>What this means for your project</h2>
<p>Every industry has its own version of these rules. In ours they are CIS and the Construction Act. In yours they might be care regulations, food safety records or financial conduct rules. They are where projects quietly fail, because they rarely appear on a wireframe. We write them as tested code, checked against real documents from the business, before we build the screens around them.</p>
<div class="callout"><b>This isn't tax advice.</b> It's a summary of how the rules shape software. HMRC's guidance is the authority on how they apply to you.</div>]]></content:encoded>
</item>
<item>
<title>SaaS metrics to track early: MRR, churn, NRR, activation and CAC payback</title>
<link>https://flavoureak.co.uk/blog/saas-metrics/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/saas-metrics/</guid>
<pubDate>Sun, 23 Aug 2026 08:00:00 GMT</pubDate>
<category>SaaS</category>
<description>Track MRR and how it moves, logo and revenue churn, net revenue retention, activation and CAC payback. Here are the formulas, one worked example and what to record from day one.</description>
<content:encoded><![CDATA[<ul><li>Monthly recurring revenue (MRR) is the monthly value of all active subscriptions, with annual plans divided by 12 and one-off fees and VAT left out; annual recurring revenue (ARR) is MRR × 12.</li><li>Logo churn counts the customers you lose, revenue churn counts the recurring revenue you lose, and net revenue retention shows whether existing customers pay you more or less over time after expansion, contraction and churn.</li><li>CAC payback is customer acquisition cost divided by the monthly gross margin a new customer brings, and gives the number of months it takes to earn back the cost of winning that customer.</li></ul><p>The metrics an early SaaS should track are monthly recurring revenue (MRR), churn, net revenue retention, activation, and customer acquisition cost with its payback period. MRR is the normalised monthly value of every active subscription, and the other metrics explain why it grows or shrinks. Each one is a simple formula, and the work that matters most is recording clean data from the first paying customer.</p>
<h2>MRR and ARR</h2>
<p>Monthly recurring revenue is the subscription revenue you would receive in a month if nothing changed. Count what customers actually pay after discounts. Divide annual plans by 12, so a £1,800 annual subscription adds £150 to MRR. Leave out VAT, set-up fees, one-off services and any usage charges that don’t recur predictably.</p>
<p>Annual recurring revenue is MRR × 12. Businesses that sell mainly annual contracts tend to report ARR, and it describes the same recurring revenue in a different unit.</p>
<p>MRR moves for five reasons, and tracking each one separately tells you far more than the total does.</p>
<ul><li><b>New</b> MRR comes from customers who started paying this month.</li><li><b>Expansion</b> MRR comes from existing customers paying more, through extra seats or an upgrade.</li><li><b>Contraction</b> MRR is lost when customers who stay pay less, such as after a downgrade.</li><li><b>Churned</b> MRR is lost when customers cancel.</li><li><b>Reactivation</b> MRR comes from former customers who return.</li></ul>
<p><code>Closing MRR = opening MRR + new + expansion + reactivation − contraction − churned</code></p>
<h2>Churn and net revenue retention</h2>
<h3>Logo churn and revenue churn</h3>
<p>Churn measures what you lose. Logo churn counts customers and revenue churn counts money. The two can tell different stories, because losing ten small customers is a very different month from losing one large one.</p>
<p><code>Logo churn rate = customers lost in the period ÷ customers at the start of the period</code></p>
<p><code>Gross revenue churn rate = (churned MRR + contraction MRR) ÷ opening MRR</code></p>
<p>Only count customers who were paying at the start of the period. Customers who sign up and cancel within the same month are better studied as an onboarding problem.</p>
<h3>Net revenue retention</h3>
<p>Net revenue retention (NRR) measures what the customers you had at the start of a period pay you now. It includes their expansion, contraction and churn, and it ignores new customers.</p>
<p><code>NRR = (opening MRR + expansion − contraction − churned) ÷ opening MRR</code></p>
<p>An NRR above 100% means your existing customers pay you more over time, before you win anyone new. It is usually reported over 12 months. In the early days a monthly figure is more useful, as long as you remember how far one customer can move it.</p>
<h2>A worked example</h2>
<p>A SaaS starts the month with 50 customers paying £10,000 MRR. During the month it wins 6 new customers worth £1,200 MRR between them. Three existing customers add seats worth £600, one downgrades by £200, and 3 customers worth £450 cancel. It spent £4,800 on sales and marketing.</p>
<table><thead><tr><th>Metric</th><th>Calculation</th><th>Result</th></tr></thead><tbody><tr><td>Closing MRR</td><td>£10,000 + £1,200 + £600 − £200 − £450</td><td>£11,150</td></tr><tr><td>ARR</td><td>£11,150 × 12</td><td>£133,800</td></tr><tr><td>Customers at month end</td><td>50 + 6 − 3</td><td>53</td></tr><tr><td>Logo churn</td><td>3 ÷ 50</td><td>6%</td></tr><tr><td>Gross revenue churn</td><td>(£450 + £200) ÷ £10,000</td><td>6.5%</td></tr><tr><td>Net revenue retention</td><td>(£10,000 + £600 − £200 − £450) ÷ £10,000</td><td>99.5%</td></tr><tr><td>CAC</td><td>£4,800 ÷ 6</td><td>£800</td></tr><tr><td>CAC payback</td><td>£800 ÷ (£200 × 80%)</td><td>5 months</td></tr></tbody></table>
<p>MRR grew by 11.5% in the month. NRR just under 100% shows that the existing customer base shrank slightly, and new sales covered the gap. Both facts matter, and the total alone hides the second one.</p>
<h2>Activation</h2>
<p>Activation is the share of new accounts that reach a first useful outcome, such as sending a first quote, within a set number of days. It is the earliest of these metrics you can influence, because it moves weeks before churn does. Define one activation event, record it for every account and report it by monthly sign-up cohort. <a class="link" href="https://flavoureak.co.uk/blog/saas-onboarding/">SaaS onboarding</a> covers how to choose the event and improve the rate.</p>
<h2>CAC and payback</h2>
<p>Customer acquisition cost (CAC) is what you spend to win a customer.</p>
<p><code>CAC = sales and marketing spend in the period ÷ new customers in the period</code></p>
<p>Include advertising, sales salaries and commission, tools and events. CAC payback then tells you how long a new customer takes to repay that cost from gross margin. Gross margin is revenue minus the cost of serving the customer, such as hosting, third-party services, payment fees, support and, for AI features, model calls (<a class="link" href="https://flavoureak.co.uk/blog/ai-cost-per-call/">AI cost per call</a>).</p>
<p><code>CAC payback in months = CAC ÷ (new customer MRR × gross margin %)</code></p>
<p>In the example, £4,800 of spend won 6 customers, so CAC is £800. The average new customer pays £200 a month. At an 80% gross margin each one contributes £160 a month and repays its CAC in 5 months. The 80% is an assumption for the example, so work out your own figure from real costs.</p>
<p>Customer lifetime value (LTV) is often quoted next to CAC. A simple version is monthly gross margin per customer divided by the monthly customer churn rate. In the example, £160 ÷ 6% gives about £2,667. With one fewer cancellation, churn would be 4% and LTV £4,000. With few customers the figure swings this much from month to month, so treat it as a rough guide until you have a long run of data.</p>
<h2>What to instrument from day one</h2>
<ol><li><b>Subscription events from your billing provider</b>, stored in your own database with plan, price, currency, interval, discounts and the dates of every start, change and cancellation (<a class="link" href="https://flavoureak.co.uk/blog/billing-for-saas/">billing for SaaS</a>).</li><li><b>Acquisition source</b>, captured at sign-up from the referrer, campaign tags and a “how did you hear about us” question.</li><li><b>The activation event</b>, with a timestamp for each account.</li><li><b>A few core usage events</b> per account, such as jobs created or lessons booked, so you can spot accounts going quiet before they cancel.</li><li><b>Cancellation reasons</b>, asked in one question when someone cancels.</li><li><b>Cost per account</b> where it varies, such as storage, messages or AI usage.</li></ol>
<h2>Mistakes with early SaaS metrics</h2>
<ul><li>Counting an annual payment as one month’s MRR, which makes that month look twelve times better than it was.</li><li>Including trials, free accounts or VAT in MRR.</li><li>Quoting percentages without counts. With 50 customers, one cancellation is 2% churn.</li><li>Changing a definition partway through the year without restating earlier months.</li><li>Working metrics out by hand from a billing dashboard each month instead of from stored events.</li></ul>
<div class="callout"><b>Building reporting into your product?</b> <a href="https://flavoureak.co.uk/saas/">See how we build SaaS platforms</a> with billing events and product analytics recorded from the first release.</div>
<h2>Questions people ask</h2>
<h3>What is a good churn rate for an early SaaS?</h3><p>There is no single good figure, because churn depends on who your customers are and how they buy. Customers on monthly plans can leave in any month, while annual contracts only come up at renewal, so the two churn very differently. Compare your own rate over time, by customer size and by sign-up cohort, and read the reasons people give when they leave.</p>
<h3>Is ARR just MRR multiplied by 12?</h3><p>For a subscription business, yes. ARR is MRR × 12, and like MRR it excludes one-off fees, services and VAT. Companies selling mostly annual contracts tend to talk in ARR, and those selling monthly plans in MRR.</p>
<h3>Is MRR the same as revenue in my accounts?</h3><p>No. MRR is an operating metric based on active subscriptions. Revenue in your statutory accounts follows accounting standards, which decide when income is recognised, and cash follows when customers actually pay. Ask your accountant how to treat annual prepayments and set-up fees.</p>
<h3>Should free users count in SaaS metrics?</h3><p>Leave free users out of MRR, churn and NRR, which are about paying customers. Track them separately, with the rate at which they move to paid plans and how long that takes.</p>]]></content:encoded>
</item>
<item>
<title>Subscription e-commerce: billing, dunning, pausing and easy cancellation</title>
<link>https://flavoureak.co.uk/blog/subscription-ecommerce/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/subscription-ecommerce/</guid>
<pubDate>Tue, 18 Aug 2026 08:00:00 GMT</pubDate>
<category>E-commerce</category>
<description>A subscription shop is won or lost after the first order, in renewals, failed payment recovery, skips and pauses, and a cancellation route as easy as signing up.</description>
<content:encoded><![CDATA[<ul><li>Subscription e-commerce sells physical products on a repeating schedule, such as subscribe-and-save replenishment, curated boxes or memberships, with the customer charged automatically each cycle.</li><li>Some subscribers leave only because a card payment failed, so build automatic retries, card updating and dunning emails before launch.</li><li>UK consumer law is moving towards subscriptions that are as easy to leave as to join, so design cancellation that takes a few steps online, with clear reminders before renewals.</li></ul><p>Subscription e-commerce is the sale of physical products on a repeating schedule, where the customer signs up once and is then charged and sent an order automatically each cycle. The common models are subscribe-and-save replenishment, curated boxes and paid memberships. Whether a subscription business keeps its customers depends on what happens after the first order: billing cycles, failed payments, skipping and pausing, and cancellation that is as easy as signing up.</p>
<h2>The three subscription models</h2>
<table><thead><tr><th>Model</th><th>Example</th><th>What drives the software</th></tr></thead><tbody>
<tr><td>Replenishment, or subscribe and save</td><td>Coffee, pet food or contact lenses every 2, 4 or 6 weeks</td><td>Flexible frequency, a repeat-order discount, easy changes to quantity and date</td></tr>
<tr><td>Curated box</td><td>A monthly box whose contents the business chooses</td><td>A fixed cut-off date, contents planning and forecasting how many boxes to pack</td></tr>
<tr><td>Membership</td><td>Paid access to member prices, free delivery or early releases</td><td>Entitlements checked at checkout, plus renewal and expiry</td></tr>
</tbody></table>
<p>Many shops mix them. A coffee roaster might sell a replenishment plan for a customer’s favourite blend alongside a monthly discovery box. Each model needs the same billing foundations and different screens on top.</p>
<h2>Billing cycles and cut-off dates</h2>
<p>Every subscription needs a billing date, a fulfilment date and a rule linking the two. For replenishment, customers are usually charged a few days before dispatch, so a failed payment can be fixed before the parcel goes out. For a curated box, everyone is billed on or before a common cut-off date, after which the contents and quantities are fixed. Decide these points before you build:</p>
<ul>
<li>whether cycles run from each customer’s sign-up date or align to one date for everyone;</li>
<li>whether a customer who joins after the cut-off gets this month’s box or next month’s;</li>
<li>how price changes reach existing subscribers, with notice before the next charge;</li>
<li>whether a change of plan takes effect now, with a pro-rata adjustment, or at the next renewal.</li>
</ul>
<p>Charges after the first are merchant-initiated transactions. Authenticate the first payment under Strong Customer Authentication and record the customer’s agreement to future charges, and later charges can then usually go through without the customer being present. Direct Debit is the other common route in the UK, with the Direct Debit Guarantee protecting customers. On Shopify, subscriptions run through subscription apps built on its subscription APIs; on a custom build, billing usually runs on your payment provider’s subscription features. We compare the two routes in <a class="link" href="https://flavoureak.co.uk/blog/shopify-or-custom/">Shopify or a custom build</a>.</p>
<h2>Failed payments and dunning</h2>
<p>Dunning is the process of recovering a failed subscription payment by retrying the charge, asking the customer to update their card and, if nothing works, pausing the subscription. Cards fail for ordinary reasons, such as expiry, a replacement card after a fraud alert, low funds on the day or a bank’s risk check. A customer whose card failed did not choose to leave, so recovering them deserves real effort. A sound sequence:</p>
<ol>
<li>Retry automatically on a schedule, for example after one, three and seven days. Many payment providers offer retries timed by their own models.</li>
<li>Use card updater services through your provider, which can refresh expired or replaced card details from the card networks.</li>
<li>Email the customer straight after the first failure, with a secure one-click link to update their card.</li>
<li>Send one or two reminders, each saying what happens next and when.</li>
<li>Hold the order until payment succeeds, and tell the customer it is waiting.</li>
<li>After the last retry, pause the subscription and keep the customer’s details, so they can restart with one click.</li>
</ol>
<p>The mechanics are the same as recurring billing in software products, which we cover in <a class="link" href="https://flavoureak.co.uk/blog/billing-for-saas/">billing for SaaS</a>.</p>
<h2>Pausing and skipping</h2>
<p>Pausing and skipping are useful retention tools, because a customer with too much coffee in the cupboard otherwise has only one button to press, and it says cancel. Let customers do all of these without contacting support:</p>
<ul>
<li>skip the next delivery;</li>
<li>pause for a set number of cycles, with the restart date shown;</li>
<li>move the next delivery earlier or later;</li>
<li>change frequency or quantity.</li>
</ul>
<p>Every skip and pause also tells you something. A customer who skips three times in a row is probably oversupplied and could be offered a longer gap between deliveries before they decide to cancel.</p>
<h2>Cancellation that is easy</h2>
<p>Making cancellation hard keeps a few subscribers for a short while and creates complaints, chargebacks and legal risk. UK law is moving firmly in the other direction. The Digital Markets, Competition and Consumers Act 2024 includes a new regime for subscription contracts. As passed, it requires clear information before sign-up, reminders before renewals and before a free or discounted trial ends, and a straightforward way for customers to leave. The subscription provisions are being brought into force separately from the rest of the Act, with detail set by regulations and guidance, so check the commencement date and the final rules before you design your flows.</p>
<p>Existing rules already apply. Under the Consumer Contracts Regulations 2013, a customer who buys goods online usually has 14 days to cancel, and for regular deliveries of goods that period runs from delivery of the first goods. Perishable goods are among the exceptions. Rules on unfair terms and misleading practices also apply to subscription traps. In practice, design cancellation the way you design sign-up:</p>
<ul>
<li>a cancel option in the account, reachable in a few taps;</li>
<li>at most one retention offer, such as a pause or a discount, with “Cancel anyway” just as visible;</li>
<li>an immediate confirmation email stating the last charge and the last delivery;</li>
<li>renewal and trial-end reminders sent before the charge, each with a link to cancel.</li>
</ul>
<h2>Forecasting stock from subscriptions</h2>
<p>Subscriptions give you advance notice of demand, which an ordinary shop does not have. For a curated box, start with active subscriptions and subtract the skips and cancellations you expect before cut-off, based on your own recent history. With 1,200 active subscriptions, and an average over the last three cycles of 90 skips and 40 cancellations before cut-off, plan for 1,200 − 90 − 40 = 1,070 boxes. Add the new sign-ups you expect before cut-off, and a small buffer for damaged parcels and replacements.</p>
<p>For replenishment, add up the next delivery dates for each product over the coming weeks, then adjust each product by its own skip rate. Share that forecast with purchasing every week, because a supplier’s lead time is often longer than one subscription cycle.</p>
<h2>The metrics that matter</h2>
<ul>
<li><b>Monthly recurring revenue (MRR):</b> the value of active subscriptions, converted to a monthly amount.</li>
<li><b>Churn:</b> the share of subscribers who leave each month, split into voluntary, where they cancelled, and involuntary, where a payment failed.</li>
<li><b>Recovery rate:</b> the share of failed payments that dunning recovers.</li>
<li><b>Skip and pause rates:</b> early warnings of churn.</li>
<li><b>Gross margin per order:</b> after product, packing and delivery.</li>
<li><b>Customer lifetime value:</b> gross margin per month divided by monthly churn.</li>
</ul>
<p>Take a £25 box with a 40% gross margin, which earns £10 a month. At 5% monthly churn, the average subscriber stays 1 ÷ 0.05 = 20 months, so lifetime margin is about £200. If winning a subscriber costs more than that, the business loses money on each one, however fast it grows. Our article on <a class="link" href="https://flavoureak.co.uk/blog/saas-metrics/">SaaS metrics</a> covers the same measures in software.</p>
<div class="callout"><b>Launching or rebuilding a subscription shop?</b> <a href="https://flavoureak.co.uk/ecommerce/">See how we build e-commerce</a>.</div>
<h2>Questions people ask</h2>
<h3>What is dunning in subscription e-commerce?</h3>
<p>Dunning is the process of recovering a failed subscription payment. It combines automatic retries, card updater services and emails asking the customer to update their payment details. If payment still fails, the subscription is paused so the customer can restart easily.</p>
<h3>How easy does subscription cancellation have to be in the UK?</h3>
<p>The Digital Markets, Competition and Consumers Act 2024 contains a subscription regime that requires a straightforward way to leave, along with reminders before renewals and trial ends. Its provisions are commencing separately from the rest of the Act, so check the current position. Existing consumer law already treats subscription traps as a risk.</p>
<h3>How do you calculate subscription churn?</h3>
<p>Monthly churn is the number of subscribers who left during the month divided by the number at the start of the month. With 1,000 subscribers at the start and 50 cancellations, churn is 50 ÷ 1,000 = 5%. Track voluntary and involuntary churn separately, because they have different fixes.</p>
<h3>Do subscription payments need Strong Customer Authentication every time?</h3>
<p>Usually not. The first payment should be authenticated, and later charges that the merchant starts under the customer’s agreement are generally treated as merchant-initiated transactions. The card issuer can still ask for authentication, so the software needs a way to bring the customer back to approve a payment.</p>]]></content:encoded>
</item>
<item>
<title>Core Web Vitals for a marketing site: why ours is a static export</title>
<link>https://flavoureak.co.uk/blog/core-web-vitals-static/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/core-web-vitals-static/</guid>
<pubDate>Fri, 14 Aug 2026 08:00:00 GMT</pubDate>
<category>SEO</category>
<description>Google measures loading, responsiveness and visual stability from real visitors. A static site served from the edge passes most of that by default.</description>
<content:encoded><![CDATA[<ul><li>The three Core Web Vitals are LCP (good at 2.5 s or less), INP (200 ms or less) and CLS (0.1 or less), measured at the 75th percentile of real visits.</li><li>A static page served from a CDN edge has no server render to wait for, which takes care of most loading problems.</li><li>What still goes wrong is usually fonts, images without dimensions, and third-party scripts.</li></ul><h2>The three numbers</h2>
<table><thead><tr><th>Metric</th><th>What it measures</th><th>Good</th><th>Poor</th></tr></thead><tbody><tr><td>LCP, Largest Contentful Paint</td><td>When the main content has loaded</td><td>≤ 2.5 s</td><td>&gt; 4.0 s</td></tr><tr><td>INP, Interaction to Next Paint</td><td>How quickly the page responds to taps and clicks</td><td>≤ 200 ms</td><td>&gt; 500 ms</td></tr><tr><td>CLS, Cumulative Layout Shift</td><td>How much the page jumps while loading</td><td>≤ 0.1</td><td>&gt; 0.25</td></tr></tbody></table>
<p>Google assesses each one at the 75th percentile of real visits, so three quarters of your visitors need a good experience for the page to pass. INP replaced First Input Delay as a Core Web Vital in March 2024.</p>
<h2>Why we built ours static</h2>
<p>flavoureak.co.uk is a Next.js site exported to plain HTML and served by Cloudflare Pages. Every page exists before anyone asks for it. There is no server rendering on request, no database query and nothing to warm up, so the HTML arrives from a server near the visitor almost immediately.</p>
<p>A marketing site rarely needs more. Its pages are the same for everyone, so building them on each request only adds time and moving parts.</p>
<h2>What still goes wrong</h2>
<h3>Fonts</h3>
<p>A web font that arrives late swaps in and shifts the text. We self-host fonts and use fallback metrics that match the real face, so the swap doesn't move anything.</p>
<h3>Images</h3>
<p>An image without a width and height leaves no space for itself, then pushes everything down when it arrives. Every image gets explicit dimensions, a modern format and a size that matches where it is shown.</p>
<h3>Third-party scripts</h3>
<p>Chat widgets, tag managers and analytics are the usual cause of poor INP, because they compete with the page for the main thread. Each one should earn its place, load after the page, and be removed when nobody reads its reports.</p>
<h3>Animation</h3>
<p>Our scroll reveals use CSS scroll-driven animations, not JavaScript, so they cost the main thread nothing and leave the page fully readable if they don't run.</p>
<h2>Measure the field, not the lab</h2>
<p>A Lighthouse score is a test run on one machine. Google uses field data from real Chrome users, which you can see in Search Console's Core Web Vitals report. Fix what the field data shows, and use the lab to find out why.</p>
<div class="callout"><b>Check your own numbers</b> against the thresholds in the <a href="https://flavoureak.co.uk/seo/">technical SEO guide</a>.</div>]]></content:encoded>
</item>
<item>
<title>Live video for online lessons: WebRTC, media servers and reliability</title>
<link>https://flavoureak.co.uk/blog/live-video-for-lessons/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/live-video-for-lessons/</guid>
<pubDate>Wed, 12 Aug 2026 08:00:00 GMT</pubDate>
<category>Education</category>
<description>Most live lesson video runs on WebRTC. One-to-one lessons can connect peer to peer, while recording, group classes and poor connections push you towards a media server or a video API.</description>
<content:encoded><![CDATA[<ul><li>WebRTC is an open standard for real-time audio, video and data, built into all major browsers and available for mobile apps, and it underpins most live lesson software.</li><li>Peer-to-peer connections suit one-to-one lessons, while a selective forwarding unit (SFU) media server suits group classes, recording and learners on mixed connections.</li><li>On a poor connection, protect audio and the whiteboard and lower video quality first, because a lesson can carry on without video and cannot carry on without sound.</li></ul><p>Live video for online lessons means two-way audio and video between teacher and learner, with a delay short enough for natural conversation, usually alongside a shared whiteboard. Almost all of it runs on WebRTC, an open standard for real-time audio, video and data that is built into every major browser and available as libraries for iPhone and Android apps. The big decisions are whether media travels directly between devices or through a server, how lessons cope with poor connections, and whether you build on WebRTC yourself or use a video API.</p>
<h2>How WebRTC works</h2>
<p>A WebRTC call has four moving parts.</p>
<ul>
<li><b>Signalling.</b> Before two devices can connect, they exchange details of what media they can send and how to reach each other. WebRTC does not define how this happens, so you run your own signalling server, usually over WebSockets.</li>
<li><b>STUN.</b> Most devices sit behind a router that hides their address. A STUN server tells each device its public address so the two can try to connect directly.</li>
<li><b>TURN.</b> Some networks, such as strict school or office firewalls, block direct connections. A TURN server relays the media instead. It always works, but all the traffic passes through your server and you pay for the bandwidth.</li>
<li><b>Media.</b> Audio and video are compressed with codecs, such as Opus for audio and VP8, VP9 or H.264 for video, and WebRTC always encrypts them in transit.</li>
</ul>
<p>A process called ICE ties these together by trying the possible routes between two devices and choosing one that works. A reliable lesson platform needs TURN servers that also accept connections over TLS on port 443, because some networks allow nothing else.</p>
<h2>Peer to peer, SFU or MCU</h2>
<p>In a peer-to-peer call, each device sends its media straight to the other. With a media server, each device sends its media to the server, which passes it on. There are two kinds of media server.</p>
<table><thead><tr><th></th><th>Peer to peer</th><th>SFU (selective forwarding unit)</th><th>MCU (multipoint control unit)</th></tr></thead><tbody>
<tr><td>How it works</td><td>Devices send media directly to each other</td><td>Server forwards each stream to the others without decoding it</td><td>Server decodes every stream and mixes them into one</td></tr>
<tr><td>Best for</td><td>One-to-one lessons</td><td>Group classes, recording, mixed connections</td><td>Very low-powered devices and older systems</td></tr>
<tr><td>Main server cost</td><td>Low: signalling, plus TURN when needed</td><td>Bandwidth</td><td>Processing power</td></tr>
<tr><td>Upload per device</td><td>One stream for each other participant</td><td>One stream, or a few quality layers</td><td>One stream</td></tr>
<tr><td>Recording</td><td>Harder: on a device, or by adding a server</td><td>Straightforward on the server</td><td>Straightforward, as media is already mixed</td></tr>
<tr><td>Adapting to each viewer</td><td>Limited</td><td>Good, using simulcast</td><td>Good, at high processing cost</td></tr>
</tbody></table>
<p>Peer to peer is fine for one teacher and one learner, but the upload each device needs grows with every extra person. In a group of six, each device would upload five copies of its video. An SFU solves this, because each device uploads once and the server forwards. With simulcast, a device sends the same video at two or three qualities and the SFU picks one for each viewer, so a learner on a weak connection gets the low-quality copy without spoiling the call for everyone else.</p>
<h2>Latency and bandwidth</h2>
<p>Latency is the delay between someone speaking and the other person hearing them. The ITU’s G.114 recommendation treats one-way delay of up to 150 milliseconds as acceptable for most conversation, and as delay rises beyond that, people start to talk over each other. WebRTC is designed for this range. Broadcast-style streaming such as HLS adds a delay measured in seconds, which suits a lecture watched by thousands and makes a conversation with a seven-year-old impossible.</p>
<p>Bandwidth is arithmetic. At an example video bitrate of 1 megabit per second, a 45-minute lesson is 2,700 seconds, so each direction carries 2,700 megabits, or about 340 megabytes. If a one-to-one lesson has to be relayed through a TURN server, the server sends out both directions, about 675 megabytes per lesson. At 1,000 relayed lessons a week, that is about 675 gigabytes of outbound traffic. That is why it pays to connect directly whenever the network allows, and to estimate TURN usage before launch.</p>
<h2>Recording and consent</h2>
<p>Recording a lesson creates a file of personal data, often about a child, so it needs a purpose, a lawful basis under UK GDPR and a clear notice. Settle these before building the feature:</p>
<ol>
<li>Why you record: safeguarding, quality review or a replay for the learner. Each purpose shapes who may watch and how long you keep the file.</li>
<li>Your lawful basis. If it is consent and the learner is under 13, the consent must come from a parent.</li>
<li>How people are told: in the booking terms, in the privacy notice and with a visible indicator in the classroom whenever recording is on.</li>
<li>Who can open recordings, with every access logged.</li>
<li>When recordings are deleted, enforced automatically.</li>
</ol>
<p>For products used by children, the ICO’s Children’s code adds expectations about defaults and transparency. See <a class="link" href="https://flavoureak.co.uk/blog/childrens-code-for-edtech/">the Children’s code for edtech</a>.</p>
<h2>Staying reliable on poor connections</h2>
<p>Lessons happen on home Wi-Fi, on tablets in kitchens and on mobile data. A lesson platform has to keep going when the connection does not.</p>
<ul>
<li><b>Protect audio first.</b> When bandwidth drops, lower video resolution and frame rate, then pause video, before audio suffers.</li>
<li><b>Reconnect automatically.</b> When the network changes, from Wi-Fi to mobile data for example, restart the connection without the user doing anything, and show “Reconnecting” while it happens.</li>
<li><b>Check before the lesson.</b> A short test of camera, microphone and connection catches problems while there is still time to fix them.</li>
<li><b>Show quality honestly.</b> A small connection indicator tells the teacher why the learner’s video is blurred.</li>
<li><b>Fall back to TURN over TLS</b> when direct connections fail.</li>
</ul>
<p>In CurioKids, the classroom opens 15 minutes before a lesson, so a family can sort out a camera or microphone before the teacher needs them.</p>
<h2>Whiteboards and shared state</h2>
<p>A shared whiteboard is the other half of a lesson. Strokes, text and images need to appear on both screens within a fraction of a second, survive a reconnection and look the same to someone who joins late. Video can drop a frame; the whiteboard cannot drop a stroke.</p>
<p>The usual approach is to send whiteboard changes as small messages, over a WebRTC data channel or a WebSocket, and to keep the authoritative state on a server so that anyone reconnecting gets the current board. When both people can draw at once, conflict-free replicated data types (CRDTs) let each device apply changes in any order and still end up with the same board. Open-source libraries such as Yjs and Automerge implement them. Save the board when the lesson ends, so it can be shared with parents or picked up next time. CurioKids runs its lessons on live video with a shared whiteboard, and the two have to stay in step.</p>
<h2>Build on WebRTC or use a video API</h2>
<p>There are three routes, and most teams should start with one of the first two.</p>
<h3>A hosted video API</h3>
<p>Several providers sell video calling as a service, usually priced per participant minute. You get SDKs, TURN servers, recording and servers around the world. You give up some control, and the cost per minute matters more as volume grows.</p>
<h3>An open-source SFU</h3>
<p>Projects such as LiveKit, mediasoup and Janus give you the media server to host yourself, and some are also offered as a hosted service. You control cost and where data is processed, but you run the servers and keep them updated.</p>
<h3>Raw WebRTC, peer to peer</h3>
<p>This is fine for a one-to-one prototype with your own signalling and TURN. It gets harder once you need recording, group lessons or detailed quality data. Whichever route you choose, keep the video layer behind your own interface in the code, so changing provider later means replacing one module. For the rest of the product, see <a class="link" href="https://flavoureak.co.uk/blog/building-a-tutoring-platform/">building a tutoring platform</a>.</p>
<div class="callout"><b>Building a live lessons product?</b> <a href="https://flavoureak.co.uk/curiokids/">See how CurioKids runs live lessons</a>.</div>
<h2>Questions people ask</h2>
<h3>What is WebRTC?</h3>
<p>WebRTC is an open standard for sending real-time audio, video and data between devices. It is built into all major browsers and is available as libraries for mobile apps. Most online lesson and video call products are built on it.</p>
<h3>Do I need a media server for one-to-one lessons?</h3>
<p>Not necessarily. A one-to-one lesson can run peer to peer, with a TURN server for networks that block direct connections. A media server becomes worthwhile when you add group lessons, server-side recording or need better control over quality.</p>
<h3>How much bandwidth does a video lesson use?</h3>
<p>It depends on the video bitrate. At an example rate of 1 megabit per second, a 45-minute lesson carries about 340 megabytes in each direction. Bitrates fall automatically on poor connections, and audio needs far less than video.</p>
<h3>Can I record online lessons with children?</h3>
<p>Yes, with a clear purpose, a lawful basis under UK GDPR and a visible notice when recording is on. If you rely on consent and the child is under 13, the consent must come from a parent. Limit who can watch recordings and delete them on a fixed schedule.</p>]]></content:encoded>
</item>
<item>
<title>Ticket refunds, transfers and resale rules for ticketing software</title>
<link>https://flavoureak.co.uk/blog/ticket-refunds-and-transfers/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/ticket-refunds-and-transfers/</guid>
<pubDate>Thu, 06 Aug 2026 08:00:00 GMT</pubDate>
<category>Events</category>
<description>Ticketing software has to handle cancellations, postponements, name changes and resale within UK consumer law. Here is what to build, and where to take legal advice.</description>
<content:encoded><![CDATA[<ul><li>Event tickets bought online have no 14-day cooling-off period in the UK, because the Consumer Contracts Regulations 2013 exclude leisure services booked for a specific date.</li><li>When an event is cancelled, ticket buyers are generally entitled to their money back, so ticketing software should be able to refund every order for an event in one action.</li><li>Under the Consumer Rights Act 2015, a ticket resold online must come with its face value, its seat or standing area where relevant and any restriction on who can use it.</li></ul><p>Ticketing software has to decide what happens to money and tickets when plans change: when an event is cancelled or postponed, when a buyer asks for a refund, when a ticket changes hands and when a payment is disputed. A refund policy is the set of rules an organiser publishes about when money is returned, and the software’s job is to apply those rules consistently and within UK consumer law. This is a planning guide and does not replace legal advice.</p>
<h2>There is no cooling-off period for event tickets</h2>
<p>Most things bought online in the UK come with a 14-day right to cancel under the Consumer Contracts (Information, Cancellation and Additional Charges) Regulations 2013. Event tickets usually do not. The regulations exclude services related to leisure activities where the contract provides for a specific date or period, which covers most concerts, conferences, festivals and matches.</p>
<p>So an organiser can lawfully refuse a refund when a buyer simply changes their mind. The software should still let organisers offer refunds up to a set date, transfers or official resale, because a buyer with a legitimate way out has less reason to raise a chargeback.</p>
<p>The terms themselves must be fair and transparent under Part 2 of the Consumer Rights Act 2015. A term that let an organiser keep the money after cancelling the event is likely to be challenged as unfair. Show the refund policy before payment and store the version each buyer accepted with their order. Our guide to <a class="link" href="https://flavoureak.co.uk/blog/event-registration-uk-gdpr/">event registration and UK GDPR</a> covers the personal data side of the same checkout.</p>
<h2>Refund and transfer scenarios at a glance</h2>
<table><thead><tr><th>Situation</th><th>Usual position</th><th>What the software should do</th></tr></thead><tbody>
<tr><td>Buyer changes their mind</td><td>No legal right to cancel; the organiser’s policy applies</td><td>Apply the published policy and offer a transfer or resale if allowed</td></tr>
<tr><td>Event cancelled</td><td>Buyer is generally entitled to a refund</td><td>Refund every order to the original payment method, showing progress and failures</td></tr>
<tr><td>Event postponed</td><td>Tickets usually stay valid, with refunds for those who cannot attend</td><td>Move the date, email holders and collect refund requests up to a deadline</td></tr>
<tr><td>Significant change, such as a new venue</td><td>Depends on the terms and how significant the change is; take advice</td><td>Let the organiser open a refund window for one event</td></tr>
<tr><td>Ticket passed to a friend</td><td>Set by the organiser’s terms</td><td>Cancel the old barcode, issue a new one and log the transfer</td></tr>
<tr><td>Ticket resold online</td><td>Consumer Rights Act 2015 information rules apply to the listing</td><td>Show face value, seat details and restrictions, then reissue the ticket</td></tr>
</tbody></table>
<h2>Cancellations and postponements</h2>
<p>When an organiser cancels an event, the buyer is generally entitled to their money back, so build for that as the default. Whether booking and delivery fees are refunded as well depends on the circumstances, so take advice and state the position in the terms. Since April 2025, the Digital Markets, Competition and Consumers Act 2024 has required mandatory fees, such as booking fees, to be included in the headline price. Record the price breakdown for each ticket as well as the order total, so that any refund of fees can be worked out from the record.</p>
<p>A postponement is different, because the event will still happen on another date. Common practice is to keep tickets valid for the new date and offer refunds to anyone who cannot attend. Build it as a single guided action for the organiser:</p>
<ol>
<li>Change the event date and keep every existing ticket valid.</li>
<li>Email every ticket holder with the new date and a deadline for requesting a refund.</li>
<li>Give each order a one-click refund request link that needs no login.</li>
<li>Process approved refunds in batches through the payment provider, with a record of each one.</li>
<li>After the deadline, close requests and report what was refunded.</li>
</ol>
<p>A cancellation runs the same machinery without the request step. A large event can mean thousands of refunds at once, so run them as a queue that can pause, retry and report. The load looks a lot like an on-sale, which we cover in <a class="link" href="https://flavoureak.co.uk/blog/event-software-on-sale-day/">event software on sale day</a>.</p>
<h2>Refunds through the payment provider</h2>
<p>Card refunds go back to the card that paid. The money leaves the organiser’s balance straight away but can take several working days to reach the buyer, so say so in the confirmation email. Settle these questions with your payment provider before the first event:</p>
<ul>
<li>Whether the provider returns its processing fee on a refund. Some do not, in which case every full refund costs the organiser money.</li>
<li>What happens when the organiser’s balance cannot cover a batch of refunds.</li>
<li>Whether the provider limits how long after a payment a refund can be made.</li>
</ul>
<h2>Name changes and transfers</h2>
<p>A transfer moves a ticket from the buyer to someone else. A name change edits the name on a named ticket. Organisers restrict both to discourage touting, so the software should let them set the rules for each event.</p>
<p>Build transfers so that the original barcode is cancelled and a new one is issued to the recipient. If both barcodes stay valid, the same ticket can be used twice, and whoever reaches the gate first gets in. Record who transferred to whom and when, allow a cap on transfers per ticket, and send the new holder their own ticket and wallet pass. For named tickets, decide whether name changes are free, charged or closed after a set date, and show that rule at the point of purchase.</p>
<p>Some platforms use barcodes that rotate every few seconds inside an app, which stops screenshots being sold on. The cost is that every attendee then needs the app and a charged phone at the door.</p>
<h2>Resale and the Consumer Rights Act 2015</h2>
<p>The Consumer Rights Act 2015 has a chapter on secondary ticketing, meaning tickets resold through an online platform. When a ticket is resold that way, the buyer must be given:</p>
<ul>
<li>the information needed to identify the seat or standing area, such as the block, row and seat number, where that applies;</li>
<li>any restriction on who can use the ticket, such as a minimum age;</li>
<li>the face value of the ticket;</li>
<li>the unique ticket number, where the organiser issued one.</li>
</ul>
<p>Further disclosures apply where the seller is connected to the platform or to the event organiser, and general consumer law stops business sellers presenting themselves as private individuals. The Act also limits an organiser’s power to cancel a ticket just because it was resold. That is only possible where the original terms said so and the term was fair.</p>
<p>If you build a resale or fan-to-fan exchange, the cleanest pattern is to resell the returned ticket through the organiser’s own system, cancel the old barcode and issue a new one. Using software to get round an event’s purchase limit in order to resell at a profit is a criminal offence under the Breaching of Limits on Ticket Sales Regulations 2018, so enforce limits per person and per card as well as per order. The government has also consulted on capping resale prices, so check the current rules before designing a resale feature.</p>
<h2>Fraud and chargebacks</h2>
<p>A chargeback happens when a cardholder asks their bank to reverse a payment. The payment provider takes the money back, usually adds a dispute fee and gives the organiser a window to submit evidence. Ticketing sees both genuine fraud with stolen card details and disputes from real buyers about tickets they could have used. Software can help with both.</p>
<ul>
<li>Use Strong Customer Authentication through 3D Secure. For an authenticated payment, liability for a fraud chargeback generally moves to the card issuer.</li>
<li>Record scan data: the time and gate at which each ticket was used.</li>
<li>Keep evidence that the buyer accepted the terms, received the tickets and was offered a refund route.</li>
<li>Watch for card testing, such as many small purchases or failed attempts from one device in a short time.</li>
<li>Use a clear statement descriptor, so the charge on a bank statement names the event.</li>
</ul>
<p>Buyers who pay by credit card also have rights under section 75 of the Consumer Credit Act 1974 where a single item costs more than £100 and not more than £30,000, which lets them claim against the card issuer if the event does not go ahead. A quick, visible refund process for cancelled events keeps those claims to a minimum.</p>
<div class="callout"><b>Building or replacing a ticketing system?</b> <a href="https://flavoureak.co.uk/events/">See how we approach event software</a>.</div>
<h2>Questions people ask</h2>
<h3>Do UK event tickets have a 14-day cooling-off period?</h3>
<p>No. The Consumer Contracts Regulations 2013 exclude leisure services booked for a specific date or period, which covers most event tickets. Organisers can still choose to offer refunds, transfers or resale, and the terms they use must be fair under the Consumer Rights Act 2015.</p>
<h3>Does an organiser have to refund tickets if an event is cancelled?</h3>
<p>Generally, yes, because the service the buyer paid for will not be provided. Whether booking fees are refunded too depends on the circumstances and the terms, so take legal advice. Ticketing software should be able to refund every order for a cancelled event in one action.</p>
<h3>What information must a resold ticket listing show?</h3>
<p>Under the Consumer Rights Act 2015, the buyer must be told the face value, the seat or standing area where relevant, any restriction on who can use the ticket and the unique ticket number where the organiser issued one. Extra disclosures apply where the seller is connected to the platform or the organiser.</p>
]]></content:encoded>
</item>
<item>
<title>Generating a lesson for the child in front of the teacher</title>
<link>https://flavoureak.co.uk/blog/lesson-for-the-child/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/lesson-for-the-child/</guid>
<pubDate>Fri, 31 Jul 2026 08:00:00 GMT</pubDate>
<category>Education</category>
<description>In CurioKids, lesson content is generated for each child. The interesting work is everything around the model: what it starts from, what it may produce, and who decides.</description>
<content:encoded><![CDATA[<ul><li>Generated content should start from what is known about the child: level, interests and what they covered last time.</li><li>Rules about what suits a child belong in code and review, not only in a prompt.</li><li>The teacher runs the lesson. Generated material is something to teach from, not a script.</li></ul><h2>The problem</h2>
<p>A teacher running a live English class for a seven-year-old wants material at the right level, about something the child cares about, ready before the call starts. Writing that by hand for every child doesn't scale, and generic worksheets don't hold a child's attention.</p>
<p>CurioKids generates lesson content for the child in front of the teacher. This is how we think about building that kind of feature.</p>
<h2>Start from what is known</h2>
<p>The model should get structured inputs, not a vague request: the child's level on the CEFR scale, from Pre-A1 for a beginner to A2 for an elementary learner; their age; what they're interested in; and what was covered last time. The output follows a fixed shape the whiteboard understands, such as target words, a warm-up, an activity and a review.</p>
<h2>Constrain the output</h2>
<p>A lesson for a child has rules: vocabulary that suits the level, sentences of a sensible length, and topics that are appropriate. Some of that can be asked for in the prompt. All of it should also be checked afterwards in code, because a prompt is a request and a check is a guarantee.</p>
<h2>The teacher stays in charge</h2>
<p>Generated material works best as a draft the teacher can change before the lesson. During the class the teacher decides what to use. Software that tries to run the lesson, instead of supporting the person running it, makes the teacher's job harder.</p>
<h2>Cost and speed</h2>
<p>Material that is prepared before a class can use a slower, more careful model without making anyone wait. The cost per lesson is known in advance, and it is small next to the cost of the teacher's time.</p>
<h2>What carries over to other products</h2>
<p>The same pattern works anywhere a professional needs a first draft tailored to one case: a quote for one job, a report for one site visit, a reply to one customer. Structured inputs, a fixed output shape, checks in code and a person who signs it off.</p>]]></content:encoded>
</item>
<item>
<title>How to choose estimating and quoting software for a UK building firm</title>
<link>https://flavoureak.co.uk/blog/choosing-estimating-software/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/choosing-estimating-software/</guid>
<pubDate>Fri, 31 Jul 2026 08:00:00 GMT</pubDate>
<category>Construction</category>
<description>Good estimating software for builders prices from your own price book and labour rates, handles VAT and CIS correctly, turns an accepted quote into a job and works on site.</description>
<content:encoded><![CDATA[<ul><li>Estimating software for builders should build quotes from your own price book, labour rates and templates so that no quote starts from scratch.</li><li>A UK building quote must show VAT correctly, flag the domestic reverse charge where it applies, and separate labour from materials where CIS will apply.</li><li>Quote-to-job conversion is often the most valuable feature, because the accepted estimate becomes the job budget without re-keying.</li></ul><p>To choose estimating and quoting software for a UK building business, check that it prices from your own price book and labour rates, supports takeoffs and templates, handles VAT, the domestic reverse charge and CIS correctly, turns an accepted quote into a job, works on a phone and connects to your accounting software. Estimating software is a tool that calculates the cost and price of building work from quantities, rates and margins, and produces a quote for the customer. The right choice depends on the kind of work you price, how many people quote and what happens after the customer says yes.</p>
<h2>What estimating software does</h2>
<p>An estimate works out what a job will cost you. A quote is the price you offer the customer, built from the estimate plus overheads and margin, with your terms. Estimating software does both. At the simple end, it is a quote builder with a list of items and prices. At the other end, it measures from drawings, prices from a large database of rates and produces a bill of quantities.</p>
<p>Most small and medium builders sit in the middle. They need fast, consistent quotes for repeat types of work, accurate enough to make money and presentable enough to win the job.</p>
<h2>The features that matter</h2>
<table>
<thead><tr><th>Feature</th><th>What it does</th><th>Why it matters</th></tr></thead>
<tbody>
<tr><td>Price book</td><td>Your own items, materials and costs, kept up to date</td><td>Every quote starts from current prices</td></tr>
<tr><td>Labour rates</td><td>Cost and charge rates by trade or grade</td><td>Labour is priced at its true cost</td></tr>
<tr><td>Templates and assemblies</td><td>Saved build-ups, such as a metre of stud wall with all its parts</td><td>Repeat work is quoted quickly and consistently</td></tr>
<tr><td>Takeoffs</td><td>Measuring quantities from drawings</td><td>Fewer missed items and arithmetic errors</td></tr>
<tr><td>Markup and margin</td><td>Overheads and profit applied visibly</td><td>You know what each quote should earn</td></tr>
<tr><td>Options and exclusions</td><td>Alternatives, provisional sums, assumptions</td><td>Fewer disputes about what was included</td></tr>
<tr><td>VAT and CIS</td><td>Correct tax treatment on the quote</td><td>Quotes and invoices follow HMRC’s rules</td></tr>
<tr><td>Quote to job</td><td>The accepted quote becomes the job budget</td><td>No re-keying, and job costing starts at once</td></tr>
<tr><td>Mobile use</td><td>Survey, measure and quote on site</td><td>Quotes go out while the customer still remembers you</td></tr>
<tr><td>Accounting integration</td><td>Contacts and invoices flow to your accounts</td><td>One set of figures and less admin</td></tr>
</tbody>
</table>
<h3>Price book and labour rates</h3>
<p>The price book is the heart of the system. It should hold your items with their unit, cost, supplier and the date the price was last checked, and let you update prices in bulk when a supplier list changes. Labour rates should hold the full cost of each grade, including employer’s on-costs, and the rate you charge. Quotes built from a well-kept price book are consistent from one estimator to the next.</p>
<h3>Takeoffs</h3>
<p>A takeoff is the process of measuring quantities, such as areas of floor, lengths of wall or numbers of doors, from drawings. Some tools let you measure directly on a PDF drawing and send the quantities into the estimate. If most of your work is priced from drawings, this saves hours. If you mainly price from site visits, a good mobile survey form may matter more.</p>
<h2>VAT, the reverse charge and CIS on quotes</h2>
<p>UK construction tax rules affect what a quote has to show, and software that gets them wrong creates problems that only surface at invoicing.</p>
<ul>
<li><b>VAT.</b> Quotes to homeowners should make the full price including VAT clear. Quotes to businesses usually show VAT separately. Some work has a reduced or zero rate in specific circumstances, such as certain new dwellings and some energy-saving installations, so the software needs to let you set the rate for each line.</li>
<li><b>Domestic reverse charge.</b> For many construction services between VAT-registered businesses that report under CIS, the customer accounts for the VAT. The quote and invoice must say the reverse charge applies. End users and intermediary suppliers are outside it, so the software must know which kind of customer you are quoting.</li>
<li><b>CIS.</b> When you work as a subcontractor, the contractor deducts CIS from the labour part of your payment at 20%, 30% or 0%, depending on your status. Materials are not subject to the deduction, so quotes and invoices should show labour and materials separately.</li>
</ul>
<p>These rules are explained in <a class="link" href="https://flavoureak.co.uk/blog/construction-rules/">CIS, reverse charge and payment notices</a>. This is general information, and HMRC’s guidance is the authority on how they apply to your business.</p>
<h2>From quote to job</h2>
<p>The point where many setups fall down is what happens after the customer accepts. If the quote lives in one tool and the job in another, someone retypes the scope and the budget, and the link between what was priced and what is spent is lost. Look for software where the accepted quote becomes the job record, with its line items carried across as the budget for <a class="link" href="https://flavoureak.co.uk/blog/job-costing-for-builders/">job costing</a>, and where later variations are priced the same way as the original quote.</p>
<p>In Estimark, tenders and quotes sit alongside jobs, scheduling and invoices in one record per job, with a price book and labour rates behind them. That is the model we would recommend to any builder who wants to know whether the price they quoted became the margin they made.</p>
<h2>Mobile use and accounting integration</h2>
<h3>Working on site</h3>
<p>Many quotes start on site with measuring a room, photographing the existing condition and noting access problems. Software that works on a phone lets an estimator capture this once, in the right place, instead of on paper to be typed up later. Check whether the mobile app can build and send a quote or only view one, and whether it works without signal. A survey in a cellar or a rural property is a poor time to discover the app needs a connection.</p>
<h3>Connecting to your accounts</h3>
<p>Ask which accounting packages the software connects to, such as Xero, QuickBooks or Sage, and exactly what passes between them, including contacts, invoices, payments, CIS deductions and VAT codes. A connection that only sends invoices still leaves payments and CIS to be reconciled by hand. Make sure tax codes map correctly, especially for the reverse charge, before relying on the link.</p>
<h2>A checklist of questions to ask suppliers</h2>
<ol>
<li>Can we import our existing price list, and update prices in bulk?</li>
<li>How are labour rates stored, and can cost and charge rates differ?</li>
<li>Can we build templates for our common types of work?</li>
<li>Does the quote handle VAT per line, the domestic reverse charge and the labour and materials split for CIS?</li>
<li>Does an accepted quote become a job, with its budget carried across?</li>
<li>How are variations priced and approved?</li>
<li>Can we create and send a quote on a phone, and does the app work offline?</li>
<li>Which accounting systems does it connect to, and what data passes each way?</li>
<li>Can we export all our data in a usable format if we leave?</li>
<li>Where is our data stored, and who in your company can see it?</li>
<li>What does support look like, and who answers when we call?</li>
<li>How are price changes and contract terms handled at renewal?</li>
</ol>
<p>Run a trial on a real job you have already priced. If the software reaches the same answer faster, and the quote looks like something you would send, that is a good sign. If it takes an afternoon to price a bathroom, it won’t be used.</p>
<h2>When off-the-shelf isn’t enough</h2>
<p>Off-the-shelf estimating software suits most builders. A firm with an unusual pricing method, a specialist trade or a process that ties estimating to manufacturing or maintenance contracts may find that no product fits without workarounds. In that case, building around your own process can make sense, provided you own the code and the data. <a class="link" href="https://flavoureak.co.uk/blog/how-to-brief-an-agency/">How to brief an agency</a> explains how to start that conversation.</p>
<div class="callout"><b>Comparing options?</b> See how <a class="link" href="https://flavoureak.co.uk/estimark/">Estimark</a> handles quotes, price books and labour rates, or read about our <a class="link" href="https://flavoureak.co.uk/construction/">construction software work</a>.</div>
<h2>Questions people ask</h2>
<h3>What is the difference between an estimate and a quote?</h3>
<p>An estimate is a calculation of what a job will cost, often given to a customer as an approximate price. A quote is a firm offer to do defined work for a stated price. Make clear on the document which one you are giving, because a customer may treat an accepted quote as a fixed price.</p>
<h3>Do builders need to show VAT on quotes?</h3>
<p>VAT-registered builders should show the VAT treatment on every quote. For homeowners, the total including VAT should be clear. Where the domestic reverse charge applies between businesses, the quote should say so and should not add VAT to the amount the customer pays the builder.</p>
<h3>Can estimating software do takeoffs from PDF drawings?</h3>
<p>Some estimating software lets you measure areas, lengths and counts directly on PDF drawings and send the quantities into the estimate. Other products rely on quantities entered by hand. If you price mainly from drawings, test the takeoff tool with your own drawings during a trial.</p>
<h3>Should estimating software connect to my accounts?</h3>
<p>Connecting estimating software to your accounting system avoids retyping contacts and invoices and keeps figures consistent. Check exactly what data passes between them, including payments, CIS deductions and VAT codes. A partial connection can create as much reconciliation work as it saves.</p>]]></content:encoded>
</item>
<item>
<title>Entities and brand mentions: how search and AI recognise your business</title>
<link>https://flavoureak.co.uk/blog/entities-and-brand-mentions/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/entities-and-brand-mentions/</guid>
<pubDate>Sun, 26 Jul 2026 08:00:00 GMT</pubDate>
<category>GEO</category>
<description>Search engines and AI systems treat a business as an entity built from consistent facts across the web. One name, Organization markup, official records and independent mentions make it recognisable.</description>
<content:encoded><![CDATA[<ul><li>An entity is a distinct thing, such as a company, person or product, that search engines and AI systems recognise as one identity across many sources.</li><li>Using one name, one description and one set of contact details everywhere helps machines connect your website, your Companies House record and your listings.</li><li>Organization structured data with sameAs links tells search engines which profiles on other sites describe the same business.</li></ul><p>Search engines and AI systems recognise a business as an entity when many independent sources describe it with the same name, details and description. An entity is a distinct, identifiable thing, such as a company, person, product or place, that a system can treat as one identity wherever it is mentioned. For a business, becoming a clear entity means consistent details across your own site, official registers, directories, reviews and other people’s websites.</p>
<h2>What an entity is and why it matters</h2>
<p>Take a fictional firm called Acme Build. A search engine that only matched words would treat “Acme Build”, “Acme Build Ltd” and “Acme Builders” as three unrelated strings. Modern search engines go further. Google keeps a Knowledge Graph of entities and the facts that connect them, and other search engines and AI systems build similar models from the web. When a system is confident that a name refers to one particular company, it can attach facts to it, such as what it does, where it is, who runs it and what people say about it.</p>
<p>That confidence matters most for AI answers. An answer engine that is unsure which “Acme” a page refers to is less likely to name it, and more likely to confuse it with another business. A business that is easy to identify is safer to mention.</p>
<h2>Where the signals come from</h2>
<table>
<thead><tr><th>Source</th><th>What it confirms</th><th>Who controls it</th></tr></thead>
<tbody>
<tr><td>Your website</td><td>Name, description, services, contact details</td><td>You</td></tr>
<tr><td>Organization structured data</td><td>The same facts in machine-readable form, plus links to profiles</td><td>You</td></tr>
<tr><td>Companies House</td><td>Registered name, number, registered office, officers</td><td>Public register, filed by you</td></tr>
<tr><td>Google Business Profile and map listings</td><td>Location, opening hours, category, reviews</td><td>You, with checks by the platform</td></tr>
<tr><td>Trade bodies and directories</td><td>Membership, sector, location</td><td>The directory</td></tr>
<tr><td>Review sites</td><td>That real customers have used you</td><td>Customers and the platform</td></tr>
<tr><td>Articles and mentions on other sites</td><td>What independent sources say about you</td><td>Other people</td></tr>
</tbody>
</table>
<p>The sources you control state your claims. The sources you don’t control corroborate them. Answer engines are cautious about repeating a claim that only one site makes about itself, as explained in <a class="link" href="https://flavoureak.co.uk/blog/geo-become-the-source/">how to become the source an AI answer cites</a>, so both kinds are needed.</p>
<h2>Use one name and one description everywhere</h2>
<p>Pick the name you trade under and use it exactly the same way on your site, social profiles, directory listings and email signatures. If your registered company name differs from your trading name, state the connection once in plain words, for example “Acme Build is the trading name of Acme Construction Services Ltd”, and keep both versions consistent from then on.</p>
<p>Do the same with a one-sentence description of what you do, your address and your phone number. Small differences add up. An old address on one directory, a former name on another and a different phone number on a third each give a system a reason to doubt that the records describe the same business.</p>
<ul>
<li>Write a standard name, a one-sentence description, an address and a phone number, and keep them in one document.</li>
<li>Search for your business name, your old names and your phone number, and list every profile you find.</li>
<li>Update or close the profiles that are out of date, starting with the ones that rank highest.</li>
<li>Check the list again whenever you move, rebrand or change number.</li>
</ul>
<h2>Organization structured data and sameAs</h2>
<p>Structured data is code on a page that describes its content in a standard vocabulary, usually schema.org written as JSON-LD. Organization markup on your home page or about page tells search engines your official name, legal name, logo, address, contact details and identifiers. Google’s documentation for Organization markup lists properties including <code>legalName</code>, <code>logo</code>, <code>address</code>, <code>vatID</code> and <code>sameAs</code>.</p>
<p>The <code>sameAs</code> property lists the addresses of other pages that describe the same organisation, such as your Companies House record, your LinkedIn page and your profiles with trade bodies. It is a direct statement that these profiles are you. Only list profiles you control or that are authoritative records of your company, and keep the markup consistent with what the page visibly says. The guide to <a class="link" href="https://flavoureak.co.uk/blog/structured-data-for-business/">structured data for business sites</a> covers the markup in more detail.</p>
<h2>Companies House and official listings</h2>
<p>For a UK limited company, the Companies House register is the authoritative public record of its registered name, company number, registered office and officers. It is free to search and widely trusted, which makes it one of the most useful links in your <code>sameAs</code> list.</p>
<p>Under the UK’s trading disclosure rules, a company’s website must show its registered name, its registered number, the part of the UK where it is registered and its registered office address. Putting these in the footer or on a legal page meets the requirement and gives search engines a clear link between the site and the register. Keep your Companies House filings up to date, because an old registered office or a late confirmation statement leaves a public record that contradicts your site.</p>
<h2>Reviews</h2>
<p>Reviews on independent platforms confirm that real customers have used a business, and their text describes what the business does in other people’s words. Ask every customer at the end of a job, make it easy with a direct link, and reply to reviews, including critical ones, with facts and courtesy.</p>
<p>Never buy, invent or selectively hide reviews. Since April 2025, the Digital Markets, Competition and Consumers Act 2024 has made it a banned practice in the UK to commission or publish fake consumer reviews, or to present reviews in a misleading way. Beyond the legal risk, a pattern of suspicious reviews damages trust with platforms and customers alike.</p>
<h2>Earning mentions on other sites</h2>
<p>A brand mention is any reference to your business on a site you don’t own, with or without a link. Mentions from relevant, credible sources turn your own claims into corroborated facts. They are earned by being useful to the people who run those sites.</p>
<ol>
<li><b>Join the bodies you qualify for.</b> Trade associations, accreditation schemes and local business groups usually list their members.</li>
<li><b>Get listed by suppliers and partners.</b> Software you integrate with, manufacturers who approve you and firms you work alongside often keep partner pages.</li>
<li><b>Publish something others want to cite.</b> A clear guide, a template or original figures from your own work give writers a reason to reference you.</li>
<li><b>Speak and contribute.</b> Talks, podcasts, guest articles and answers in trade press put your name next to your subject.</li>
<li><b>Fix mentions that are wrong.</b> Ask for corrections to old names, wrong addresses and broken links.</li>
</ol>
<p>Be wary of paid link schemes and mass directory submissions. They add noise without adding trust, and Google’s spam policies treat buying links for ranking purposes as link spam.</p>
<p>Wikipedia and Wikidata feed many entity systems, and a business with an entry there is easier to identify. Both have their own rules on notability and conflicts of interest, and writing about your own company is strongly discouraged. Most small businesses don’t qualify and don’t need to. Consistent details on your own site, in official records and from independent sources do the same job at a smaller scale.</p>
<div class="callout"><b>Launching a new product name?</b> We set up structured data, consistent naming and launch pages from the first release. See our <a class="link" href="https://flavoureak.co.uk/seo/">SEO guide</a> and <a class="link" href="https://flavoureak.co.uk/geo/">GEO guide</a>.</div>
<h2>Questions people ask</h2>
<h3>What is an entity in SEO?</h3>
<p>An entity in SEO is a distinct thing, such as a business, person, product or place, that a search engine recognises as a single identity. Search engines connect facts to entities as well as to keywords. A business becomes a clear entity when many sources describe it consistently.</p>
<h3>What should I put in sameAs?</h3>
<p>The sameAs property should list the addresses of authoritative profiles that describe the same organisation. For a UK company that usually means its Companies House page, its main social profiles and its listings with relevant trade bodies. Only include profiles you control or that are official records of your business.</p>
<h3>Do unlinked brand mentions help?</h3>
<p>Brand mentions without a link still tell search engines and AI systems what others say about a business, particularly when the name, location and services match your own description. Links carry additional value for search ranking. Both are worth earning from relevant, credible sources.</p>
<h3>Does my trading name need to match my company name?</h3>
<p>A trading name does not have to match the registered company name. The website must still show the registered name and the other details required by the trading disclosure rules. Stating clearly that the trading name belongs to the registered company helps search engines connect the two.</p>]]></content:encoded>
</item>
<item>
<title>Programmatic SEO for SaaS: landing pages at scale without thin content</title>
<link>https://flavoureak.co.uk/blog/programmatic-seo-saas/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/programmatic-seo-saas/</guid>
<pubDate>Mon, 20 Jul 2026 08:00:00 GMT</pubDate>
<category>SEO</category>
<description>Programmatic SEO generates many pages from structured data. It works for SaaS when each page answers a real search with information unique to it, and fails when pages only swap a keyword.</description>
<content:encoded><![CDATA[<ul><li>Programmatic SEO is creating many landing pages from a template and a dataset, such as one page per integration, template or use case.</li><li>Each generated page needs real information the other pages don’t have, or it risks being treated as thin, duplicate or scaled content.</li><li>Google’s spam policies cover scaled content abuse, which means producing many pages mainly to manipulate rankings, however the pages are made.</li></ul><p>Programmatic SEO is the practice of generating many landing pages from a template and a structured dataset, each targeting a specific search, such as “scheduling software that integrates with Xero” or “job sheet template for electricians”. For SaaS it works when every page carries real, page-specific information that answers the search. It fails, and can harm the whole site, when pages differ only by the keyword swapped into them.</p>
<h2>Page types that suit SaaS</h2>
<p>Good programmatic pages exist because each item in a dataset genuinely deserves its own page. For SaaS products, these are the common patterns and what each one needs.</p>
<table><thead><tr><th>Page type</th><th>Example search</th><th>Data each page needs</th><th>Risk</th></tr></thead><tbody>
<tr><td>Integrations</td><td>“scheduling software that integrates with Xero”</td><td>What syncs, in which direction, how often, setup steps and limits</td><td>Low, if each integration exists and is documented</td></tr>
<tr><td>Templates</td><td>“free RAMS template”</td><td>The template itself, ready to use, with guidance on filling it in</td><td>Low to medium, depending on template quality</td></tr>
<tr><td>Use cases or trades</td><td>“job management software for roofers”</td><td>Workflows, terms and rules specific to that trade</td><td>Medium, because pages drift towards swapping the trade name</td></tr>
<tr><td>Comparisons</td><td>“[product] alternatives”</td><td>Accurate, current differences in features and pricing</td><td>Medium, as they must be fair and kept up to date</td></tr>
<tr><td>Locations</td><td>“builders’ software Manchester”</td><td>Something genuinely local, which software rarely has</td><td>High, close to doorway pages</td></tr>
</tbody></table>
<p>Integration and template pages usually make the best start. The data already exists inside the product, and the searches show clear intent. Someone searching for an integration already uses the other tool and wants to know whether yours works with it, which is a question one well-built page can answer completely.</p>
<h2>What Google’s spam policies say</h2>
<p>Google’s spam policies include scaled content abuse, added in March 2024. It covers generating many pages where the main purpose is manipulating search rankings and the pages don’t help users, whether they are made by automation, generative AI, people or a combination. Google’s examples include using AI tools to produce many pages that add no value, and stitching together content from other pages without adding anything.</p>
<p>The doorway abuse policy is just as relevant. Doorways are pages created to rank for similar searches that lead people to much the same destination, such as one near-identical page for each town. Location pages for a software product that works the same everywhere fit that description closely.</p>
<p>Breaking these policies can lead to pages ranking lower or not appearing at all, or to a manual action. A manual action can apply to part of a site or to all of it, so a careless batch of generated pages can put the rest of the site at risk.</p>
<h2>Data quality decides page quality</h2>
<p>A programmatic page is only as good as the record behind it. If the integrations table holds a name and a logo for each partner, you get forty pages with a name and a logo. If it holds the fields that sync, the direction, the frequency, the setup steps and the known limits, you get forty useful pages.</p>
<p>Some arithmetic shows how quickly page counts outrun information. A product with 40 integrations can justify 40 integration pages. A page for every pair of integrations gives 40 × 39 ÷ 2 = 780 pages. Crossing the 40 integrations with 50 towns gives 40 × 50 = 2,000 pages. The useful information hasn’t grown at all. Only the page count has.</p>
<p>Set rules before generating anything:</p>
<ol>
<li>Define the minimum fields a record needs before its page can be published.</li>
<li>Write the template so that missing data hides a section, and never fills it with filler text.</li>
<li>Give each page at least one element that exists nowhere else on the site, such as real screenshots, a worked example or data from your own product.</li>
<li>Have a person review a sample of pages from every batch before it goes live.</li>
<li>Assign an owner to keep the data current, because an integration page that describes a feature you have removed is worse than no page.</li>
</ol>
<h2>Unique value on every page</h2>
<p>A useful test is to imagine search engines didn’t exist. The page should still be one you would build for a customer who asked about that subject. An integration page passes easily, because customers ask how an integration works before they buy. A page about job management software for plasterers passes only if it says something specific to plasterers, such as how the product handles the materials, measurements or payment terms they deal with.</p>
<p>Generative AI can help draft the variable parts of a template, but text written to fill space is exactly what the scaled content policy describes. Use it to turn real data into readable sentences, and have people check the results. Don’t use it to produce content with no data behind it.</p>
<p>Comparison pages need particular care. They should be accurate on the date they are published, say when they were last checked, and treat competitors fairly. Out-of-date or one-sided comparisons damage trust with readers, and misleading claims about a competitor can also create legal risk.</p>
<h2>Internal links and indexing control</h2>
<p>Generated pages need a place in the site’s structure. Build a hub page for each set, such as an integrations directory, that links to every page in it. Link related pages to each other, so an integration page links to the use cases where it matters, and add breadcrumbs. Pages reachable only through a sitemap are harder for search engines to find and judge. The principles are set out in <a class="link" href="https://flavoureak.co.uk/blog/internal-linking/">internal linking for SEO</a>.</p>
<p>Generate the pages as static or server-rendered HTML, so crawlers get the full content on their first request, as explained in <a class="link" href="https://flavoureak.co.uk/blog/seo-for-javascript-sites/">SEO for JavaScript sites</a>. Then control what gets indexed:</p>
<ul>
<li>Launch in batches, starting with the pages that have the best data, and see how Google treats them before generating more.</li>
<li>Add a noindex tag to pages that don’t yet meet your minimum data rules, instead of publishing them half-empty.</li>
<li>Where two generated pages are near-duplicates, merge them, or point a canonical tag at the stronger one.</li>
<li>Keep a separate sitemap for each page type, so Search Console shows indexing for each type on its own.</li>
<li>Check pages from each new batch with the URL Inspection tool to confirm Google sees the full content.</li>
<li>Watch the Page indexing report for “Crawled - currently not indexed” and “Discovered - currently not indexed”. Large numbers of either for one page type suggest Google doesn’t see enough value in those pages.</li>
</ul>
<p>Measure each page type separately, by clicks in Search Console and by the sign-ups or demo requests it leads to. Improve or remove page types that attract neither, because a large set of pages nobody visits adds crawling work and no value.</p>
<div class="callout"><b>Planning integration, template or use-case pages for your SaaS?</b> We build them from the same data as the product, so they stay accurate as the product changes. <a class="link" href="https://flavoureak.co.uk/seo/">See our SEO guide</a>.</div>
<h2>Questions people ask</h2>
<h3>Is programmatic SEO against Google’s guidelines?</h3>
<p>No. Generating pages from data is allowed, and many large, well-ranked sites are built that way. It breaks Google’s spam policies when the main purpose is manipulating rankings and the pages offer little value, which Google calls scaled content abuse.</p>
<h3>How many programmatic pages should we launch at once?</h3>
<p>Start with a small batch of the pages with the best data, such as the twenty integrations customers use most. Watch how they are indexed and whether they attract traffic before generating the rest.</p>
<h3>Can we use AI to write programmatic SEO pages?</h3>
<p>You can use AI to turn real data into readable text, with people reviewing the output. Using AI to produce large numbers of pages with no real information behind them is the pattern Google’s scaled content abuse policy describes.</p>
<h3>What is the difference between programmatic SEO and doorway pages?</h3>
<p>Programmatic pages each answer a distinct need with information specific to that page. Doorway pages are near-identical pages made to rank for similar searches, such as one per town, that lead to the same place. The template can be identical in both cases, and the difference lies in whether each page has its own substance.</p>]]></content:encoded>
</item>
<item>
<title>You get the repository: what a proper handover includes</title>
<link>https://flavoureak.co.uk/blog/you-get-the-repository/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/you-get-the-repository/</guid>
<pubDate>Fri, 17 Jul 2026 08:00:00 GMT</pubDate>
<category>Engineering</category>
<description>If only your agency can deploy your software, you don’t fully own it. This is the checklist we hand over against.</description>
<content:encoded><![CDATA[<ul><li>Domains, cloud and app store accounts should be in your company’s name from day one.</li><li>A handover is a checklist you can verify, not a meeting.</li><li>Test it: someone outside the project should be able to deploy from the documentation alone.</li></ul><h2>Why this matters</h2>
<p>Businesses get stuck with software they paid for but can't change. The code lives in the agency's account, the app is published under the agency's developer name, and the only person who knew how to deploy has left. Moving to anyone else means starting again.</p>
<p>We work the other way round. Everything is yours from the start, and the handover is a list you can check.</p>
<h2>Accounts in your name</h2>
<ul><li>Domain registration and DNS</li><li>Cloud hosting and the database</li><li>Apple Developer Program and Google Play Console, under your organisation</li><li>Payment provider, email sending and any AI provider</li><li>Error monitoring and analytics</li></ul>
<p>We are added as users on your accounts, and removed when we're done.</p>
<h2>The repository</h2>
<ul><li>The full history, not a zip file of the final version</li><li>A README that gets a new developer from clone to running locally</li><li>Every environment variable listed with what it's for, and no secrets in the code</li><li>Tests that run with one command</li></ul>
<h2>Deploys you can repeat</h2>
<p>A deploy is one command or one merge, run by a pipeline, not by someone's laptop. The pipeline lives in the repository, so it moves with the code.</p>
<h2>A runbook</h2>
<p>A short document for when things go wrong: how to roll back a release, restore the database, rotate a leaked key, and who to contact at each provider.</p>
<h2>Data you can get back</h2>
<p>Backups run automatically, and a restore has been tested. A backup that has never been restored is a hope.</p>
<h2>Test the handover</h2>
<p>The test is simple. Could another developer, with only the repository and the runbook, deploy a change on their first day? If not, the handover isn't finished.</p>]]></content:encoded>
</item>
<item>
<title>Mobile app security basics: what every business app should get right</title>
<link>https://flavoureak.co.uk/blog/mobile-app-security/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/mobile-app-security/</guid>
<pubDate>Tue, 14 Jul 2026 08:00:00 GMT</pubDate>
<category>Engineering</category>
<description>Keep secrets out of the app, store tokens in the Keychain or with the Android Keystore, check every permission on the server, and test against OWASP’s mobile guidance.</description>
<content:encoded><![CDATA[<ul><li>Store tokens and keys in the iOS Keychain, or encrypt them with a key held in the Android Keystore, and never in plain files or preferences.</li><li>Treat anything shipped inside the app as public, and enforce every permission on the server, because an attacker can call your API without the app.</li><li>Certificate pinning, root detection and app attestation add protection with trade-offs attached, and none of them replaces a secure API.</li></ul><p>Mobile app security is the set of measures that protect an app’s data, its users’ accounts and the systems it talks to, on a device you do not control. The basics are to keep secrets out of the app, store tokens in the platform’s secure storage, use short-lived tokens, enforce every permission on the server and test against a recognised standard. Many serious weaknesses sit in the API behind an app, so security work starts there as much as on the phone.</p>
<h2>Why the device cannot be trusted</h2>
<p>An app runs on a phone that someone else owns. Anyone can download it, unpack it, read its code, watch its network traffic and run it on a modified device. Two rules follow from that.</p>
<ul><li><b>Anything shipped in the app is public.</b> API keys, secrets and hidden URLs in the app bundle can be extracted. Keys that must stay secret, such as a payment provider’s secret key or an AI provider’s API key, belong on your server, and the app calls your server instead.</li><li><b>The app is not the security boundary.</b> Checks in the app improve the experience. Checks on the server are what protect the data.</li></ul>
<h2>Secure storage with the Keychain and Keystore</h2>
<p>Each platform provides secure storage, and tokens, keys and other sensitive values belong there.</p>
<h3>iOS Keychain</h3>
<p>The Keychain is an encrypted store managed by the operating system. Items can be limited to the current device, so they are not copied to a new phone from a backup, and they can require a fresh Face ID or Touch ID check before they are read.</p>
<h3>Android Keystore</h3>
<p>The Android Keystore holds cryptographic keys in a way that makes them hard to extract, using secure hardware on many devices. Apps do not usually put tokens in the Keystore itself. They encrypt the tokens with a key held in the Keystore and store the encrypted result. Keys can also be set to require the person to authenticate before each use.</p>
<p>Plain storage, such as UserDefaults on iOS or SharedPreferences on Android, should only hold data you would not mind someone reading. The same thinking applies to the local database in an offline app. Estimark saves to the phone first so people can work without signal, and any app built that way has to decide what is stored on the device, for how long and how it is protected if the phone is lost. Hide sensitive screens from the app switcher, keep tokens and personal data out of logs, and exclude sensitive files from device backups.</p>
<h2>Tokens and sessions</h2>
<p>Most apps sign people in once and then use tokens. A short-lived access token is sent with each API request, and a longer-lived refresh token is used to get new access tokens.</p>
<ul><li>Keep access tokens short-lived, measured in minutes rather than days.</li><li>Issue a new refresh token each time one is used, and revoke the whole chain if an old one is presented again, because that suggests it was stolen.</li><li>Let people see and end their sessions on other devices, and end them all when the password changes.</li><li>For sign-in through an external identity provider, use the system browser with OAuth 2.0 and PKCE, the approach recommended for native apps, instead of an embedded web view that could capture passwords.</li><li>Revoke tokens on the server when someone signs out. Deleting them from the phone is not enough.</li></ul>
<h2>Certificate pinning and its trade-offs</h2>
<p>Every connection should use HTTPS, and both platforms push apps that way by default. App Transport Security on iOS blocks insecure connections unless an exception is declared, and Android blocks plain HTTP by default for apps that target Android 9 or later.</p>
<p>Certificate pinning goes a step further. The app accepts only a specific public key or certificate for your server, which protects against interception by someone holding a certificate from a compromised or misused certificate authority. The risk is operational. If your certificate or hosting provider changes and the app was not prepared, every installed copy stops connecting until people update.</p>
<p>If you pin, pin public keys instead of whole certificates, include at least one backup key, and plan how keys will be rotated. iOS supports pinning in the app’s configuration through NSPinnedDomains, and Android supports it through its network security configuration, which can set an expiry date on pins.</p>
<h2>Biometrics, jailbreaks and attestation</h2>
<h3>Biometrics</h3>
<p>Face ID, Touch ID and Android’s biometric prompt let a person confirm it is them without typing a password. The app never sees the face or fingerprint, only a result. A simple yes or no can be bypassed on a modified device, so the stronger pattern ties the biometric check to a key in the Keychain or Keystore. The key, and the token it protects, can then be used only after a successful check.</p>
<h3>Jailbroken and rooted devices</h3>
<p>Jailbreaking an iPhone or rooting an Android phone removes protections the operating system provides. Apps can look for signs of it, but a determined attacker can always hide them. Treat detection as a signal to record, or a reason to limit high-risk actions, and never as a guarantee.</p>
<h3>App attestation</h3>
<p>Apple’s App Attest and Google’s Play Integrity API let your server check that a request comes from a genuine copy of your app on a genuine device. The server verifies a token signed by Apple or Google, which is much harder to fake than anything the app reports about itself. It helps against scripted abuse of your API, such as fake sign-ups and scraping.</p>
<h2>Securing the API behind the app</h2>
<p>An attacker does not need your app to call your API. They can replay its requests with any tool. One of the most damaging flaws is broken object-level authorisation, where changing an ID in a request returns someone else’s data, and OWASP’s API security guidance treats it as a leading risk.</p>
<ul><li>Check on the server, on every request, that the signed-in user may access that specific record, within their own organisation’s data.</li><li>Validate all input on the server, whatever the app has already checked.</li><li>Rate-limit sign-in, password reset and expensive endpoints.</li><li>Return only the fields the screen needs, so a response never carries data the person should not see.</li><li>Log security events, such as failed sign-ins and refused requests, without logging tokens or passwords.</li></ul>
<p>Our guide to <a class="link" href="https://flavoureak.co.uk/blog/saas-roles-permissions/">roles and permissions</a> covers how to model who may do what.</p>
<h2>What to test</h2>
<p>OWASP publishes the Mobile Top 10 alongside the Mobile Application Security Verification Standard (MASVS) and a testing guide that explains how to check each requirement. Use them as a checklist, in proportion to the data your app holds.</p>
<table><thead><tr><th>Area</th><th>How to test it</th><th>When</th></tr></thead><tbody><tr><td>Data stored on the device</td><td>Inspect the app’s files, databases and backups on a test device</td><td>Before launch and after storage changes</td></tr><tr><td>Network traffic</td><td>Route the app through an intercepting proxy and read every request and response</td><td>Before launch and after major changes</td></tr><tr><td>API authorisation</td><td>Replay requests as a different user and change record IDs</td><td>Every release, with automated tests</td></tr><tr><td>Secrets in the app</td><td>Unpack the release build and search it for keys and URLs</td><td>Every release</td></tr><tr><td>Dependencies and SDKs</td><td>Scan for known vulnerabilities and review what each SDK collects</td><td>Continuously, in the build pipeline</td></tr><tr><td>Sessions</td><td>Sign out, change the password, revoke a device and reuse an old refresh token</td><td>Before launch and after sign-in changes</td></tr><tr><td>Modified devices</td><td>Run the app on a rooted or jailbroken test device</td><td>Before launch</td></tr></tbody></table>
<p>For apps that handle payments, health information, children’s data or large amounts of personal data, commission an independent penetration test before launch. UK GDPR requires appropriate technical and organisational measures to keep personal data secure, and a documented test helps show you have taken them. The checks larger customers ask for are covered in <a class="link" href="https://flavoureak.co.uk/blog/saas-security-before-enterprise/">security before enterprise customers</a>.</p>
<div class="callout"><b>Building an app that holds customer data?</b> Design the API permissions and on-device storage first, then the screens. <a class="link" href="https://flavoureak.co.uk/mobile/">See how we approach mobile apps</a>.</div>
<h2>Questions people ask</h2>
<h3>Is it safe to put an API key in a mobile app?</h3>
<p>Only if the key is meant to be public, such as a key restricted to your app that can only do harmless things. Any key that grants access to paid services or private data can be extracted from the app and should live on your server instead. The app calls your server, and your server calls the provider.</p>
<h3>Should my app use certificate pinning?</h3>
<p>It depends on the risk. Pinning protects against some interception attacks, but it can stop every installed copy connecting if certificates change unexpectedly. Apps handling payments or sensitive data often use it with backup keys and a rotation plan, while many business apps gain more from a well-secured API.</p>
<h3>Should my app block jailbroken or rooted phones?</h3>
<p>Blocking them outright can turn away legitimate users and is easy for attackers to get around. A better approach is to check attestation from Apple and Google on the server, and to limit high-risk actions when a device fails the check. Decide based on what the app protects.</p>
<h3>What is the OWASP Mobile Top 10?</h3>
<p>The OWASP Mobile Top 10 is a list of the most significant security risks for mobile apps, published by the Open Worldwide Application Security Project. It is updated periodically and is widely used as a starting point for design reviews and testing, alongside OWASP’s more detailed verification standard.</p>]]></content:encoded>
</item>
<item>
<title>Using AI with customer data under UK GDPR: what businesses need to do</title>
<link>https://flavoureak.co.uk/blog/ai-and-uk-gdpr/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/ai-and-uk-gdpr/</guid>
<pubDate>Thu, 09 Jul 2026 08:00:00 GMT</pubDate>
<category>AI</category>
<description>UK GDPR applies to AI like any other use of personal data. You need a lawful basis, an honest privacy notice, a processor contract with the AI provider and, usually, a DPIA.</description>
<content:encoded><![CDATA[<ul><li>Sending customer data to an AI model is processing under UK GDPR, so you need a lawful basis, transparency, data minimisation and appropriate security.</li><li>An AI provider you call through an API is usually your processor, which means you need an Article 28 contract and must know where it stores data and for how long.</li><li>The ICO’s guidance on AI and data protection says most uses of AI with personal data are likely to need a data protection impact assessment.</li></ul><p>Using AI with customer data is processing personal data, so UK GDPR and the Data Protection Act 2018 apply in full. In practice that means a lawful basis for the use, a privacy notice that explains it, sending the model no more data than it needs, a written contract with the AI provider, and usually a data protection impact assessment before you go live. Personal data is any information relating to an identified or identifiable living person, and that includes what a model writes about them.</p>
<p>This is a practical summary for people building or buying software. Take legal advice for anything with real risk.</p>
<h2>Choosing a lawful basis for AI processing</h2>
<p>Every use of personal data needs a lawful basis under Article 6 of UK GDPR. For AI features in business software, two bases come up most often.</p>
<ul>
<li><b>Contract.</b> The processing is necessary to deliver a service the customer has asked for. An AI feature that is part of the product the customer signed up to, such as a tool that drafts quotes from their own enquiries, may fit here.</li>
<li><b>Legitimate interests.</b> You have a genuine interest, the processing is necessary for it, and it doesn’t override the person’s interests and rights. This needs a written legitimate interests assessment, kept on file.</li>
</ul>
<p>Consent is possible but awkward for AI, because it must be freely given, specific, informed and as easy to withdraw as to give.</p>
<p>Special category data, such as health information, needs an additional condition under Article 9. Children’s data needs extra care, and online services likely to be accessed by children must also meet the ICO’s Age Appropriate Design Code, which we cover in <a class="link" href="https://flavoureak.co.uk/blog/childrens-code-for-edtech/">the Children’s code for edtech</a>.</p>
<h2>Telling people what the AI does</h2>
<p>Transparency means people know their data is used by AI, for what purpose and who receives it. Your privacy notice should name the purpose in plain words, such as “we use an AI model to summarise your support requests so an agent can reply faster”. It should name the categories of recipient, including the AI provider, and say where data goes and how long it is kept.</p>
<p>Where AI output affects someone directly, say so on the screen where it happens as well as in the policy.</p>
<p>People keep their rights over data processed by AI. They can ask for a copy of their personal data, including AI outputs about them, and you normally have one month to respond. Build the feature so those outputs can be found and exported.</p>
<h2>Data minimisation and the AI prompt</h2>
<p>Data minimisation means using only personal data that is adequate, relevant and limited to what the purpose needs. With AI, the place to apply it is the prompt. Developers often send a whole customer record because it is convenient. A model summarising a complaint usually needs the complaint text and the order details. It doesn’t need the customer’s date of birth, phone number or payment history.</p>
<ul>
<li>Choose the fields each task needs in code, and send only those.</li>
<li>Replace names and identifiers with placeholders where the task allows, and restore them afterwards.</li>
<li>Strip signatures, forwarded chains and attachments that aren’t relevant to the task.</li>
<li>Never paste whole exports or databases into a prompt to save time.</li>
</ul>
<h2>AI providers as processors and sub-processors</h2>
<p>When you send data to an AI provider through its API to do a job for you, the provider is usually your processor. You are the controller, which means you decide why and how the data is used, and you stay responsible for it. Article 28 requires a written contract with every processor. Most providers offer a data processing addendum to business and API customers. Read it and keep a copy.</p>
<p>The provider will also use sub-processors, typically the cloud companies that host its models. The contract should list them, or say where the list is published, and explain how you will hear about changes. Ask these questions before you sign.</p>
<table><thead><tr><th>Question to ask the provider</th><th>Why it matters under UK GDPR</th></tr></thead><tbody>
<tr><td>Is there a data processing addendum that meets Article 28?</td><td>A written processor contract is a legal requirement</td></tr>
<tr><td>Is our data used to train or improve your models?</td><td>If so, the provider is using it for its own purposes, which a processor arrangement doesn’t cover</td></tr>
<tr><td>How long are inputs and outputs kept, and why?</td><td>Retention must be limited and explained in your privacy notice</td></tr>
<tr><td>In which countries is data processed and stored?</td><td>Transfers outside the UK need a valid transfer mechanism</td></tr>
<tr><td>Who are your sub-processors, and how are changes notified?</td><td>You need to know everyone who handles the data</td></tr>
<tr><td>What security controls and certifications do you have?</td><td>Article 32 requires security appropriate to the risk</td></tr>
</tbody></table>
<h2>International transfers and retention by AI providers</h2>
<p>Many AI models run in data centres outside the UK. Sending personal data abroad is a restricted transfer under UK GDPR and needs a lawful route. The usual routes are UK adequacy regulations, which cover the EEA and some other countries, the UK extension to the EU-US Data Privacy Framework for US companies certified under it, and the ICO’s International Data Transfer Agreement or the UK Addendum to the EU standard contractual clauses. The contractual routes also need a documented risk check. Some providers offer processing in a UK or European region, which simplifies this.</p>
<p>Retention is the other question. Many providers keep inputs and outputs for a period, often to detect misuse, and some offer shorter or zero retention for business customers on request. Whatever the answer, record it in your records of processing and reflect it in your privacy notice. Your own logs of AI calls count too, so give them a retention period and delete on schedule. Logging and monitoring are covered in <a class="link" href="https://flavoureak.co.uk/blog/testing-ai-features/">how to test an AI feature</a>.</p>
<h2>When you need a DPIA</h2>
<p>A data protection impact assessment, or DPIA, is a documented assessment of the risks a type of processing poses to people and of how you will reduce them. Article 35 makes a DPIA mandatory where processing is likely to result in a high risk to people’s rights and freedoms. The ICO’s guidance on AI and data protection says that most uses of AI with personal data are likely to meet that test, and the ICO lists the use of innovative technology, including AI, among the factors that point to a DPIA.</p>
<p>Do it before you build, while the design can still change. Describe the flow of data from the screen to the provider and back, the lawful basis, what could go wrong and the controls that will prevent it. Typical risks are wrong outputs about a person, data leaking through logs, and outputs that treat one group worse than another. If a high risk remains after your controls, you must consult the ICO before going ahead.</p>
<h2>Automated decisions and Article 22</h2>
<p>Article 22 of UK GDPR has long restricted decisions based solely on automated processing that have a legal or similarly significant effect on someone, such as refusing credit online or rejecting a job application with no person involved. “Solely” means no meaningful human involvement. A person who approves every output without looking doesn’t count.</p>
<p>The Data (Use and Access) Act 2025 replaces Article 22 with a new set of rules. In broad terms, it allows more solely automated significant decisions as long as safeguards are in place, and keeps tighter limits where special category data is involved. The safeguards include telling people about such decisions, letting them make representations, and giving them a way to obtain human intervention and contest the outcome. Check which rules apply on the date you go live, read the ICO’s current guidance and take legal advice before automating any decision of this kind.</p>
<p>Most AI features in business software avoid the question by design. The model drafts, sorts or suggests, and a person decides. That is good practice in any case, as we explain in <a class="link" href="https://flavoureak.co.uk/blog/designing-for-wrong-answers/">designing AI features for when the answer is wrong</a>.</p>
<div class="callout"><b>Adding AI to software that holds customer data?</b> We map the data flow and check the provider’s terms before any code is written. <a class="link" href="https://flavoureak.co.uk/ai/">See how we build AI features</a>.</div>
<h2>Questions people ask</h2>
<h3>Is it legal to use AI on customer data in the UK?</h3>
<p>Yes, if you meet UK GDPR as you would for any other processing, with a lawful basis, a clear privacy notice, data minimisation, a processor contract, a lawful route for transfers abroad and appropriate security. Take legal advice where the risks are high.</p>
<h3>Can staff paste customer emails into an AI chatbot?</h3>
<p>Only if the tool is covered by a proper processor contract and your privacy notice and lawful basis cover the use. Consumer chat apps on personal accounts usually aren’t covered. Give staff an approved tool and a short written rule about what may be pasted into it.</p>
<h3>Do we need consent to use AI on customer data?</h3>
<p>Not necessarily. Consent is one of several lawful bases, and contract or legitimate interests often fit AI features in business software better. Whichever you choose, record why it applies and explain the use in your privacy notice.</p>
<h3>Does UK GDPR apply to anonymised data in prompts?</h3>
<p>Truly anonymised data, where nobody can be identified by any means reasonably likely to be used, is outside UK GDPR. Swapping names for placeholders is usually pseudonymisation, and pseudonymised data is still personal data. Treat it as personal data unless you are confident it can’t be linked back to a person.</p>]]></content:encoded>
</item>
<item>
<title>robots.txt, llms.txt and the AI crawlers: what to allow</title>
<link>https://flavoureak.co.uk/blog/ai-crawlers-robots/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/ai-crawlers-robots/</guid>
<pubDate>Fri, 03 Jul 2026 08:00:00 GMT</pubDate>
<category>GEO</category>
<description>AI companies run separate crawlers for training and for live answers. You can block one without disappearing from the other.</description>
<content:encoded><![CDATA[<ul><li>OpenAI, Anthropic and Perplexity each publish separate user agents for training, search and user-requested fetches.</li><li>Google-Extended controls use of your content in Gemini. It doesn’t affect Google Search or AI Overviews.</li><li>llms.txt is a proposed convention. It’s cheap to add, but don’t expect it to change results on its own.</li></ul><h2>Who is crawling</h2>
<p>The main AI companies now separate their crawlers by purpose and document them. These are the user agents you write rules for in robots.txt.</p>
<table><thead><tr><th>Company</th><th>User agent</th><th>Used for</th></tr></thead><tbody>
<tr><td>OpenAI</td><td>GPTBot</td><td>Training</td></tr>
<tr><td>OpenAI</td><td>OAI-SearchBot</td><td>ChatGPT search results</td></tr>
<tr><td>OpenAI</td><td>ChatGPT-User</td><td>Pages a user asks ChatGPT to open</td></tr>
<tr><td>Anthropic</td><td>ClaudeBot</td><td>Training</td></tr>
<tr><td>Anthropic</td><td>Claude-SearchBot</td><td>Search results in Claude</td></tr>
<tr><td>Anthropic</td><td>Claude-User</td><td>Pages a user asks Claude to open</td></tr>
<tr><td>Perplexity</td><td>PerplexityBot</td><td>Perplexity's search index</td></tr>
<tr><td>Perplexity</td><td>Perplexity-User</td><td>Pages a user asks for</td></tr>
<tr><td>Google</td><td>Google-Extended</td><td>Use in Gemini models (a control token, not a separate crawler)</td></tr>
<tr><td>Common Crawl</td><td>CCBot</td><td>An open dataset widely used for training</td></tr>
</tbody></table>
<p>Check each company's documentation before relying on this list, because names and behaviour change.</p>
<h2>A sensible default</h2>
<p>For most businesses that want customers to find them, the useful position is to allow the search and user-fetch agents, so you can appear and be cited in AI answers, and to decide separately about training.</p>
<pre><code>User-agent: GPTBot
Disallow: /

User-agent: ClaudeBot
Disallow: /

User-agent: CCBot
Disallow: /

User-agent: *
Allow: /

Sitemap: https://your-business.co.uk/sitemap.xml</code></pre>
<p>Blocking Google-Extended doesn't remove you from Google Search or AI Overviews, which use the normal Googlebot. Block Googlebot and you disappear from both.</p>
<h2>What about llms.txt?</h2>
<p>llms.txt is a convention proposed in 2024: a Markdown file at the root of your site that lists your most useful pages, each with a one-line description, written for a language model to read. It's quick to write and does no harm. No major answer engine has committed to using it, so treat it as tidy housekeeping rather than a ranking lever.</p>
<h2>robots.txt is a request</h2>
<p>Well-behaved crawlers follow robots.txt, and nothing forces the rest to. If content must be kept out of AI systems, it shouldn't be public. For everything else, robots.txt is how you state your preference, and the main companies honour it.</p>
<div class="callout"><b>Build yours</b> with the robots.txt builder in the <a href="https://flavoureak.co.uk/geo/">GEO guide</a>.</div>]]></content:encoded>
</item>
<item>
<title>Who owns the code when you commission software in the UK</title>
<link>https://flavoureak.co.uk/blog/who-owns-the-code-uk/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/who-owns-the-code-uk/</guid>
<pubDate>Fri, 03 Jul 2026 08:00:00 GMT</pubDate>
<category>Working with us</category>
<description>Under UK law the developer, or the agency that employs them, owns the copyright in code until it is assigned to you in writing. Here is what to put in the contract.</description>
<content:encoded><![CDATA[<ul><li>Under the Copyright, Designs and Patents Act 1988, copyright in code belongs to the person who wrote it, or to their employer, unless it is assigned in writing.</li><li>An agency or freelancer you hire is not your employee, so paying for software does not by itself make you the owner of its copyright.</li><li>A good contract assigns the bespoke code to you, licenses any reusable components, lists open-source licences and puts every account in your company’s name.</li></ul><p>When you commission software in the UK, the copyright in the code belongs to whoever wrote it, or to their employer, until it is assigned to you in a signed written document. Copyright is the legal right to copy, change and distribute a work, and the Copyright, Designs and Patents Act 1988 treats computer programs as literary works, so the same rules apply to code as to a book. If your contract with an agency or freelancer says nothing about ownership, you have paid for the software without owning it.</p>
<h2>The default rule under UK copyright law</h2>
<p>The Copyright, Designs and Patents Act 1988 (CDPA) protects computer programs, and the preparatory design material for them, as literary works. The first owner of copyright is the author, meaning the person who created the work. There is one important exception. When an employee writes code in the course of their employment, their employer owns it, unless they have agreed otherwise.</p>
<p>That exception is why ownership catches people out. A developer on your payroll writes code that your company owns automatically. A developer at an agency writes code that the agency owns, because the developer is the agency’s employee and not yours. A freelancer working on their own account owns what they write personally. In both of the last two cases, the work only becomes yours if the copyright is assigned to you.</p>
<p>An assignment of copyright has no effect unless it is in writing and signed by or on behalf of the person transferring it. An invoice marked “paid” does not meet that test, and neither does a conversation in which the agency said “of course it’s yours”.</p>
<h2>What happens if the contract says nothing</h2>
<p>Without a written assignment you are not left with nothing. A court is likely to find that you have at least an implied licence to use the software for the purpose it was commissioned, and in some cases courts have gone further and found that the client owns the work in equity. Both outcomes depend on the facts, and finding out means a dispute with legal fees on both sides.</p>
<p>An implied licence is also narrower than ownership. It may let you run the software without letting you license it to others, sell it with your business or stop the developer reusing the same code for a competitor. Investors and buyers check intellectual property during due diligence, and a gap here can hold up a funding round or a sale.</p>
<h2>Assignment and licence compared</h2>
<p>An assignment transfers ownership of the copyright to you. A licence is permission to use something that someone else continues to own. Most commissioned software involves both, because an agency will assign the code it writes for you and license the parts it brought with it.</p>
<table><thead><tr><th>Question</th><th>Assignment</th><th>Licence</th></tr></thead><tbody><tr><td>Who owns the copyright</td><td>You</td><td>The supplier, or the original author</td></tr><tr><td>Can the supplier reuse the code for other clients</td><td>No, unless you license it back</td><td>Yes, unless the licence is exclusive</td></tr><tr><td>Can you change it and hire another developer</td><td>Yes</td><td>Only if the licence allows it</td></tr><tr><td>Can it pass to a buyer if you sell the business</td><td>Yes</td><td>Only if the licence allows transfer</td></tr><tr><td>Formal requirement</td><td>In writing, signed by the supplier</td><td>Best in writing; an exclusive licence must be in writing and signed</td></tr><tr><td>Typical use</td><td>Code written specifically for you</td><td>The supplier’s existing tools, libraries and templates</td></tr></tbody></table>
<h2>Background code, open source and third-party services</h2>
<p>Not everything in your software can be assigned to you. Most applications contain three kinds of code that someone else owns.</p>
<h3>The supplier’s background code</h3>
<p>Agencies reuse their own libraries, starter templates and internal tools across projects, and it is reasonable for them to keep ownership of these. What you need is a licence to use them as part of your software that is perpetual, irrevocable, royalty-free and transferable, and that lets another developer maintain them after the relationship ends. Ask for the background code to be identified, so you know where the line sits.</p>
<h3>Open-source components</h3>
<p>Almost every application depends on open-source packages. Their authors keep the copyright and grant everyone a licence. Permissive licences such as MIT and Apache 2.0 mainly ask for the copyright notice and licence text to be kept. Copyleft licences such as the GPL can require you to make your own source code available if you distribute software that includes the component, and the AGPL can extend that duty to modified software that people use over a network. A supplier should list every open-source component with its licence, and agree with you in advance which licences are acceptable.</p>
<h3>Third-party services</h3>
<p>Payments through Stripe, maps from Google, email sending and AI models are services you use under each provider’s terms. You never own them. What you should own is the account, so the contract with the provider is between your company and them, and nobody can switch it off by walking away.</p>
<h2>What to put in the contract</h2>
<p>These are the clauses we’d expect to see in a UK software development agreement. Treat them as a checklist for a conversation with your solicitor. This article explains the general position and is not legal advice, so take advice on your own contract, especially if the software is central to the value of your business.</p>
<ol><li><b>Assignment of the deliverables.</b> The supplier assigns to you all intellectual property in the code, designs and documentation created for the project. Wording that assigns future copyright as it is created avoids a gap between the work being written and a later signature. The clause should cover everything delivered, however it was produced, including code written with AI tools, where the law on ownership is less settled.</li><li><b>Timing.</b> Agencies often assign on payment. That is fair, as long as each payment transfers the work it covers and ownership does not wait for the final invoice.</li><li><b>Background licence.</b> A perpetual, irrevocable, royalty-free licence to any background code built into the deliverables, which you can pass on if you sell the business.</li><li><b>Subcontractors.</b> The supplier confirms it has written assignments from everyone else who worked on the project, including freelancers. An agency cannot assign rights it never acquired.</li><li><b>Open source.</b> A list of components and licences, and a commitment not to include copyleft code without your written approval.</li><li><b>Moral rights.</b> In the UK, the moral rights to be named as author and to object to derogatory treatment do not apply to computer programs. They do apply to designs, illustrations and written content, so the contract should include a waiver for those.</li><li><b>Non-infringement.</b> A warranty that the deliverables do not infringe anyone else’s rights, with an indemnity if they do.</li><li><b>Handover.</b> Delivery of the full source code with its history, and all credentials, at the end of the project or on termination.</li></ol>
<p>Raise ownership in your first conversation with any supplier. It is one of the questions in our guide on <a class="link" href="https://flavoureak.co.uk/blog/how-to-brief-an-agency/">how to brief an agency</a>.</p>
<h2>Repositories and accounts in your name</h2>
<p>Owning the copyright is of limited use if you cannot get at the code. Ownership on paper and control in practice are separate questions, and the second is settled by whose name is on each account.</p>
<ul><li>The source code repository, in an organisation on GitHub, GitLab or Bitbucket that your company controls, with the supplier invited as members.</li><li>The Apple Developer Program and Google Play Console accounts, enrolled as your organisation. Both ask an organisation for a D-U-N-S Number, which can take a while to obtain, so start early.</li><li>The domain name, DNS and email, registered to your company.</li><li>Cloud hosting, databases and backups, billed to your company.</li><li>Payment, email, SMS, maps and AI provider accounts.</li><li>Design files and brand assets.</li></ul>
<p>If an app is published under the supplier’s developer account, moving it later is possible on both stores, but it involves a formal transfer with conditions attached. Our <a class="link" href="https://flavoureak.co.uk/blog/you-get-the-repository/">handover checklist</a> covers the rest of what you should receive.</p>
<div class="callout"><b>Commissioning software?</b> Ask any supplier, including us, to show you the ownership clause before you sign. <a class="link" href="https://flavoureak.co.uk/contact/">Talk to us about your project</a>.</div>
<h2>Questions people ask</h2>
<h3>Do I own the code if I paid for it?</h3>
<p>Not automatically. Under UK law, payment does not transfer copyright, and an assignment has to be in writing and signed by the supplier. Without one you probably have a licence to use the software, while the supplier or the individual developer still owns it.</p>
<h3>Can a freelancer reuse code they wrote for me?</h3>
<p>If they still own the copyright, yes, unless your contract restricts it. Once the copyright is assigned to you, they need your permission to reuse that code. General skills and know-how stay with the developer either way.</p>
<h3>Who owns code my employee writes?</h3>
<p>Your company does, if the employee wrote it in the course of their employment and nothing in their contract says otherwise. Code written by a director who is not an employee, or by a co-founder before the company existed, is a different matter and should be assigned to the company in writing.</p>
<h3>Does an agency’s standard contract usually cover ownership?</h3>
<p>Some do and some don’t, and the wording varies widely. Read the intellectual property clause before signing, check that it assigns the bespoke work to you instead of only licensing it, and ask a solicitor to review it if the software matters to the value of your business.</p>]]></content:encoded>
</item>
<item>
<title>B2B SaaS onboarding: getting new accounts to a first useful outcome</title>
<link>https://flavoureak.co.uk/blog/saas-onboarding/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/saas-onboarding/</guid>
<pubDate>Sat, 27 Jun 2026 08:00:00 GMT</pubDate>
<category>SaaS</category>
<description>Good onboarding gets a new account to one useful result in its first session, then brings the rest of the team in. Sample data, short checklists and helpful empty states do most of the work.</description>
<content:encoded><![CDATA[<ul><li>SaaS onboarding is everything between sign-up and the moment a new account first gets something useful from the product, such as sending its first quote or booking its first lesson.</li><li>Define that first outcome as a measurable activation event, track the share of new accounts that reach it within a set number of days, and remove steps that don’t lead to it.</li><li>In B2B products the customer is a team, so onboarding also covers inviting colleagues, importing existing data and completing the few settings that block real use.</li></ul><p>SaaS onboarding is the path from sign-up to the first time a new account gets a useful result from your product. In B2B software that result is usually a piece of real work done faster, such as a quote sent, an order placed or a lesson booked. Good onboarding finds the shortest route to that result, uses sample data, checklists and empty states to guide people along it, and measures how many accounts get there.</p>
<h2>Define the first useful outcome and clear the path to it</h2>
<p>Start by naming the moment when a new account has had value from the product. It should be an action the customer takes, that you can record, and that matters to their business. Some examples from different kinds of product:</p>
<ul><li><b>Construction management software:</b> the first quote sent to a client.</li><li><b>A tutoring platform:</b> the first lesson booked.</li><li><b>A trade ordering portal:</b> the first order placed by a customer.</li><li><b>A reporting tool:</b> the first report shared with a colleague.</li></ul>
<p>For a product like CurioKids, where parents use an app to book, join and follow lessons, the natural first outcome is a booked lesson. Everything in onboarding should either lead to that moment or wait until after it.</p>
<h3>Remove steps before the first outcome</h3>
<p>List every screen between the sign-up form and the first outcome, then question each one. Products tend to ask for too much, too early.</p>
<p>Ask only for what the first outcome needs. A quote needs a company name and a client. It doesn’t need a VAT number, a logo, bank details or the size of the team. Ask for the VAT number when the user creates their first invoice, and the logo when they first preview a document, where the reason is obvious.</p>
<p>Offer sign-in with Google or Microsoft accounts if your buyers use them, so there is no password to create. Send the email verification message, let people keep working while it waits, and require verification only before they send anything to their own customers.</p>
<h2>Sample data and templates</h2>
<p>A blank product is hard to judge. Sample data lets a new user see what a finished job, schedule or report looks like before they have entered anything of their own.</p>
<p>Keep sample data clearly labelled, out of real numbering sequences and totals, and removable in one click. A sample invoice that takes invoice number 1, or a demo client that appears in a VAT report, causes real problems later.</p>
<p>Templates are often more useful than samples, because the user keeps them. In construction software, the price book and labour rates decide how quickly a first quote can be built, so a starter price book the user can edit gets them further than an empty one. Imports matter too. Most new B2B customers arrive with a spreadsheet, and a CSV import with column matching and a preview gets their real data in quickly.</p>
<h2>Checklists that point at real work</h2>
<p>A short checklist shows a new admin what to do next and how far they have got. Keep it to five items or fewer, tick each one only when the action has really happened, and let people dismiss it. A checklist for construction software might read as follows.</p>
<ol><li>Add the company details that appear on quotes.</li><li>Add your first client.</li><li>Build a quote from the price book.</li><li>Send the quote.</li><li>Invite a colleague.</li></ol>
<p>Each item links straight to the place where the work is done. Ticking an item because someone opened a page teaches them nothing and tells you nothing.</p>
<h2>Invitations and the rest of the team</h2>
<p>A B2B account succeeds when the team uses it, so invitations are part of onboarding. Ask the person inviting to choose a role at the same time, using your <a class="link" href="https://flavoureak.co.uk/blog/saas-roles-permissions/">roles and permissions</a> model, so new users arrive with the right access.</p>
<p>Design the invited user’s first screen separately from the account owner’s. An estimator invited to a job management tool should land on the jobs assigned to them, with the setup checklist left to the admin. Make invitation links expire, let admins resend them, and show pending invitations in the user list. Larger customers will want people added automatically through <a class="link" href="https://flavoureak.co.uk/blog/sso-for-saas/">single sign-on and SCIM</a>.</p>
<h2>Empty states that tell people what to do</h2>
<p>An empty state is what a screen shows before it has any data. Every new account sees all of them, so they are part of onboarding whether you design them or not.</p>
<table><thead><tr><th>Screen</th><th>Unhelpful empty state</th><th>Useful empty state</th></tr></thead><tbody><tr><td>Jobs</td><td>“No jobs found.”</td><td>What a job holds, a button to create one and a link to import a spreadsheet</td></tr><tr><td>Schedule</td><td>An empty calendar</td><td>A note that jobs appear here once they have dates, with a link to the jobs list</td></tr><tr><td>Reports</td><td>Charts with no data</td><td>One sentence on what each report shows and what data it needs</td></tr><tr><td>Team</td><td>Only the owner’s name</td><td>An invite button, with each role explained in a line</td></tr></tbody></table>
<p>A useful empty state says what belongs on the screen, why it matters and the one action that fills it.</p>
<h2>Measuring activation</h2>
<p>Activation is the share of new accounts that reach the first useful outcome within a set period. Record an event when an account is created and when it first completes the outcome. Then calculate the rate for each monthly group of sign-ups, called a cohort.</p>
<p>For example, if 80 accounts signed up in March and 30 of them sent a quote within 14 days, activation for March is 30 ÷ 80 = 37.5%. Track the median time to that first outcome as well, and the step where accounts that don’t activate stop. That step is where to work next. Activation sits alongside revenue and churn among the <a class="link" href="https://flavoureak.co.uk/blog/saas-metrics/">metrics an early SaaS should track</a>.</p>
<p>Numbers show where people stop. Conversations show why. Each month, call or email a few accounts that stalled and ask what got in the way.</p>
<h2>Emails and human help</h2>
<p>Emails triggered by what an account has or hasn’t done, such as a reminder to send a quote they drafted, are more useful than a fixed sequence sent to everyone. Keep them about using the service the customer signed up for. Messages that promote other products or offers count as direct marketing, which has its own rules under PECR and UK GDPR, so keep the two kinds separate. For larger accounts, a short call to help set up the first real job is often the quickest onboarding of all.</p>
<div class="callout"><b>Building a product people need to adopt quickly?</b> <a href="https://flavoureak.co.uk/founders/">See how we scope an MVP</a> around the first useful outcome.</div>
<h2>Questions people ask</h2>
<h3>What is activation in SaaS?</h3><p>Activation is the point at which a new account first gets real value from the product, defined as a specific recorded action such as sending a first invoice. The activation rate is the share of new accounts that reach it within a set number of days. It is one of the earliest signs of whether a product will keep its customers.</p>
<h3>Should onboarding include a product tour?</h3><p>We’d use short hints at the moment a feature is needed in preference to a long tour at sign-up, which people can click through without learning anything. If you use a tour, keep it to a few steps and end it on the action that leads to the first useful outcome.</p>
<h3>Should sample data sit in the customer’s real account?</h3><p>It can, if it is clearly labelled, excluded from totals, reports and numbering, and removable in one click. The alternative is a separate demo account, which keeps real data clean but means the user has to switch back to start their own work.</p>
<h3>How do you onboard the rest of a customer’s team?</h3><p>Make inviting colleagues part of the setup checklist, ask for each person’s role when inviting them, and give invited users a first screen built around their own work. For larger customers, provisioning through single sign-on and SCIM adds people automatically as their IT team grants access.</p>]]></content:encoded>
</item>
<item>
<title>Marketplace payments in the UK: taking money on behalf of sellers</title>
<link>https://flavoureak.co.uk/blog/marketplace-payments-uk/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/marketplace-payments-uk/</guid>
<pubDate>Sun, 21 Jun 2026 08:00:00 GMT</pubDate>
<category>E-commerce</category>
<description>A marketplace that takes payment for its sellers is handling other people’s money, which UK payment rules regulate. Most use a payment platform with connected accounts to verify and pay sellers.</description>
<content:encoded><![CDATA[<ul><li>A UK marketplace that collects buyers’ money and passes it to sellers may be providing a regulated payment service, which is why most use a regulated payment provider to hold and move the funds.</li><li>Payment platforms with connected accounts, such as Stripe Connect, verify sellers, split each payment, take the platform’s fee and pay sellers out to their bank accounts.</li><li>Since 1 January 2021, online marketplaces have been responsible for the VAT on many sales of goods by overseas sellers to UK consumers, so take VAT advice before launch.</li></ul><p>Marketplace payments are the systems that let a platform take money from a buyer, keep a fee and pass the rest to the seller who supplied the goods or service. In the UK, receiving money on behalf of others can be a regulated payment service under the Payment Services Regulations 2017, so most marketplaces leave the holding of funds to a payment platform with connected accounts, such as Stripe Connect. The marketplace then designs onboarding, fees, payouts, refunds and disputes around that provider, and takes advice on VAT and tax reporting.</p>
<h2>Why taking payment for others is regulated</h2>
<p>If your business receives a buyer’s money and later pays it to a seller, you are holding funds that belong to someone else. Under the Payment Services Regulations 2017 that can be a payment service needing authorisation or registration with the Financial Conduct Authority (FCA). There is a commercial agent exclusion, but it covers an agent acting for only the buyer or only the seller. Most marketplaces act for both sides and handle the money, so they cannot rely on it. Take legal advice on where your model sits.</p>
<p>The practical answer for most marketplaces is to let a regulated payment provider hold and move the money. Options include Stripe Connect, Adyen for Platforms, PayPal’s marketplace products and Mangopay. They structure funds flows differently, so read how each one handles the money and who it treats as the merchant of record.</p>
<table><thead><tr><th>Approach</th><th>Who holds the buyer’s money</th><th>Regulatory position</th><th>Usually fits</th></tr></thead><tbody>
<tr><td>Platform collects and pays sellers itself</td><td>The platform</td><td>Likely to need FCA authorisation or registration, unless an exclusion clearly applies</td><td>Rarely the right place to start</td></tr>
<tr><td>Payment platform with connected accounts</td><td>The regulated provider</td><td>The provider is regulated, and the platform follows its rules</td><td>Most marketplaces</td></tr>
<tr><td>Sellers take payment with their own provider</td><td>Each seller</td><td>The platform does not touch the funds</td><td>Directories and listings sites that charge a subscription or lead fee</td></tr>
</tbody></table>
<h2>How connected accounts work</h2>
<p>A connected account is an account the payment provider opens for each seller, linked to your platform’s account. The buyer pays once, and the provider splits the money between the seller’s connected account and your platform fee. The main design choices are these.</p>
<ul>
<li><b>Who is the merchant of record.</b> The charge can be made on the seller’s account, so the seller’s name appears on the buyer’s statement and the seller carries refunds and disputes. Or it can be made on the platform’s account, with funds then transferred to the seller.</li>
<li><b>Who pays the processing fee.</b> It can come out of the seller’s share or out of the platform fee, or be built into the price.</li>
<li><b>How much of the provider sellers see.</b> Hosted onboarding pages and dashboards save build time. Embedded or fully custom screens keep sellers inside your product.</li>
</ul>
<h2>Onboarding and KYC for sellers</h2>
<p>KYC, short for know your customer, is the identity checking a regulated provider must do before it pays anyone. Under the Money Laundering Regulations 2017, the payment provider has to verify each seller. What it asks for depends on who the seller is:</p>
<ul>
<li>for individuals, name, date of birth, address and sometimes an identity document;</li>
<li>for companies, registration details that can be checked against Companies House, plus directors and people with significant control;</li>
<li>for every seller, a bank account in their own name for payouts.</li>
</ul>
<p>Design onboarding around the points where sellers give up. Where the provider allows it, let sellers create listings before checks are complete and hold payouts until they pass. Show exactly what is missing. Expect the provider to ask for more information later, for example when a seller’s volume grows, and build a way to tell sellers when that happens.</p>
<h2>Platform fees and payouts, with a worked example</h2>
<p>Take a £100 order on a marketplace that keeps a 12% commission and absorbs an example processing fee of 2% plus 20p. These are illustrative rates and do not come from any provider’s price list.</p>
<table><thead><tr><th>Line</th><th>Amount</th></tr></thead><tbody>
<tr><td>Buyer pays</td><td>£100.00</td></tr>
<tr><td>Platform fee at 12%</td><td>£12.00</td></tr>
<tr><td>Seller receives</td><td>£88.00</td></tr>
<tr><td>Processing fee (2% of £100 plus 20p), paid by the platform</td><td>£2.20</td></tr>
<tr><td>Platform keeps</td><td>£9.80</td></tr>
</tbody></table>
<p>If the platform is VAT registered, VAT is usually due on its commission, so decide whether the 12% includes VAT before publishing your fees.</p>
<p>Payouts move the seller’s share from their connected account to their bank. Decide the schedule, daily, weekly or after delivery, and whether to hold funds until an order is fulfilled or a service delivered. Holding funds protects buyers, but sellers will want to know exactly when they will be paid, so show it in their dashboard.</p>
<h2>Refunds and disputes</h2>
<p>When a buyer is refunded, up to three amounts have to be reversed: the seller’s share, the platform fee and, depending on the provider, the processing fee. Settle the rules in your seller terms before launch.</p>
<ol>
<li>Who can issue a refund: the seller, the platform, or both.</li>
<li>Whether the platform returns its fee on a full refund, and on a partial one.</li>
<li>Who carries the processing fee if the provider does not return it.</li>
<li>How a chargeback is handled: who supplies evidence, who pays the dispute fee and what happens if the seller has already been paid.</li>
<li>What happens when a seller leaves with refunds still owed.</li>
</ol>
<p>In some connected account set-ups the platform is ultimately liable for negative balances its sellers cannot cover. A reserve, or a delay before new sellers are paid out, is a sensible protection.</p>
<h2>VAT and reporting</h2>
<p>Marketplaces have their own VAT rules, and this is an area to take specialist advice on. Since 1 January 2021, an online marketplace is treated as the supplier for VAT purposes, and must account for the VAT, when it facilitates sales of goods to UK consumers that are:</p>
<ul>
<li>imported in consignments worth £135 or less; or</li>
<li>already in the UK and sold by a seller not established in the UK.</li>
</ul>
<p>Separately, UK digital platforms must collect information about their sellers and report their income to HMRC each year, under rules based on an OECD model that took effect from 1 January 2024. Build seller onboarding to collect what that reporting needs, such as tax identification numbers, so you are not chasing sellers for it later.</p>
<h2>Strong Customer Authentication</h2>
<p>Strong Customer Authentication (SCA) is the UK requirement for many online payments to be confirmed with two of three factors: something the customer knows, has or is. For cards it usually works through 3D Secure, with the bank asking the customer to approve the payment, often in their banking app. The FCA required full enforcement for UK online card payments from 14 March 2022. Exemptions exist, for example for some low-value payments and for merchant-initiated transactions such as later instalments, and the card issuer makes the final decision.</p>
<p>Your payment provider handles the mechanics. Your job is to design checkout so that an authentication step does not lose the sale, and to authenticate the card up front whenever you will charge it later, for example when a deposit is followed by a final payment. The same applies to repeat charges, which we cover in <a class="link" href="https://flavoureak.co.uk/blog/subscription-ecommerce/">subscription e-commerce</a>, and to lesson marketplaces such as those in <a class="link" href="https://flavoureak.co.uk/blog/building-a-tutoring-platform/">building a tutoring platform</a>.</p>
<div class="callout"><b>Planning a marketplace?</b> <a href="https://flavoureak.co.uk/ecommerce/">See how we build e-commerce platforms</a>.</div>
<h2>Questions people ask</h2>
<h3>Do I need FCA authorisation to run a marketplace?</h3>
<p>If your marketplace receives buyers’ money and pays it to sellers itself, you may be providing a regulated payment service that needs FCA authorisation or registration. Most marketplaces avoid this by using a regulated payment provider that holds and moves the funds. Take legal advice on your specific model.</p>
<h3>What is Stripe Connect?</h3>
<p>Stripe Connect is Stripe’s product for platforms and marketplaces. It creates a connected account for each seller, verifies their identity, splits each payment between the seller and the platform and pays sellers out to their bank accounts. Other providers offer similar products for platforms.</p>
<h3>Who pays VAT on marketplace sales in the UK?</h3>
<p>A VAT-registered seller normally accounts for VAT on their own sales. Since 1 January 2021, the marketplace is responsible instead for goods sold to UK consumers in imported consignments worth £135 or less, and for goods already in the UK sold by sellers not established in the UK. Take VAT advice, because the details depend on your model.</p>
<h3>Does Strong Customer Authentication apply to marketplace payments?</h3>
<p>Yes. Online card payments by UK customers generally need SCA, usually through 3D Secure, unless an exemption applies. Payment providers handle the authentication, and marketplaces should design checkout around it and authenticate up front when cards will be charged later.</p>]]></content:encoded>
</item>
<item>
<title>CEFR levels explained: A1 to C2 for English learning products</title>
<link>https://flavoureak.co.uk/blog/cefr-levels-explained/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/cefr-levels-explained/</guid>
<pubDate>Tue, 16 Jun 2026 08:00:00 GMT</pubDate>
<category>Education</category>
<description>The CEFR describes language ability in six levels from A1 to C2, with Pre-A1 below them for young beginners. Here is what each level means and how learning products use them.</description>
<content:encoded><![CDATA[<ul><li>The CEFR, or Common European Framework of Reference for Languages, is a Council of Europe framework that describes language ability in six levels: A1, A2, B1, B2, C1 and C2.</li><li>Pre-A1 describes learners below A1 and is used mainly with young learners, for example in the Cambridge English Pre A1 Starters exam.</li><li>English learning products use CEFR levels to place learners, tag content and show progress, using can-do statements that describe what a learner can do in the language.</li></ul><p>CEFR levels are the six steps of the Common European Framework of Reference for Languages, a Council of Europe framework that describes what a learner can do in a language at each stage. They run from A1 for beginners to C2 for near-complete mastery, grouped as basic users (A), independent users (B) and proficient users (C), with Pre-A1 below A1 for the youngest beginners. Because each level is defined by what a learner can do, an English learning product can use the levels to place learners, choose content and show progress in terms that teachers and parents already recognise.</p>
<h2>Where the CEFR comes from</h2>
<p>The Council of Europe published the CEFR in 2001 to give teachers, examiners and learners a common way to describe language ability. It is used far beyond Europe and is not tied to any one language. In 2020 the Council of Europe published a Companion Volume that updated and extended the descriptors, adding detail for Pre-A1, mediation and online interaction among other areas.</p>
<p>The framework describes ability through can-do descriptors. A can-do descriptor is a short statement of what a learner can do at a level, such as introducing themselves or following the main points of a talk on a familiar subject. The descriptors describe performance in real tasks. They are deliberately not a grammar syllabus or a vocabulary count, which is why products have to do some work to turn them into lessons.</p>
<h2>The levels in plain language</h2>
<p>The table paraphrases the overall picture at each level. The official descriptors are longer and are split by skill, such as reading, listening, speaking and writing.</p>
<table><thead><tr><th>Level</th><th>Group</th><th>What a learner can do</th><th>Cambridge English exam at this level</th></tr></thead><tbody>
<tr><td>Pre-A1</td><td>Below basic</td><td>Recognise and use a small set of words and memorised phrases about themselves and the world around them, with a lot of support</td><td>Pre A1 Starters</td></tr>
<tr><td>A1</td><td>Basic user</td><td>Use familiar everyday expressions and very simple phrases, introduce themselves and answer simple questions if the other person speaks slowly and helps</td><td>A1 Movers</td></tr>
<tr><td>A2</td><td>Basic user</td><td>Handle short, routine exchanges about familiar things such as family, shopping and their local area, and describe their background simply</td><td>A2 Flyers, A2 Key</td></tr>
<tr><td>B1</td><td>Independent user</td><td>Understand the main points of clear speech on familiar topics, cope with most situations while travelling, and describe experiences, plans and opinions</td><td>B1 Preliminary</td></tr>
<tr><td>B2</td><td>Independent user</td><td>Understand the main ideas of complex text, including discussion in their own field, talk with fluent speakers without strain on either side and write clear, detailed text</td><td>B2 First</td></tr>
<tr><td>C1</td><td>Proficient user</td><td>Understand long, demanding texts and implied meaning, and express ideas fluently and flexibly for social, study and work purposes</td><td>C1 Advanced</td></tr>
<tr><td>C2</td><td>Proficient user</td><td>Understand almost everything they hear or read, summarise from different sources and express fine shades of meaning precisely</td><td>C2 Proficiency</td></tr>
</tbody></table>
<p>Pre A1 Starters, A1 Movers and A2 Flyers are Cambridge’s exams for young learners of primary school age. The framework also uses “plus” levels, such as A2+ and B1+, for learners who are well above one level but not yet at the next. Some exam boards and publishers use their own finer scales mapped to the CEFR, such as the Cambridge English Scale and Pearson’s Global Scale of English.</p>
<h2>Level is not the same as age</h2>
<p>A seven-year-old and a forty-year-old can both be at A1, and they need completely different lessons. The level says what the learner can do in English. Age decides the topics, the length of each activity, how much play is involved and how instructions are given. A product for children therefore needs both on every piece of content.</p>
<p>The steps between levels are not equal either. Each higher level covers more language and more demanding tasks, so moving from B2 to C1 usually takes much longer than moving from A1 to A2. Progress screens should not suggest that every level takes the same time.</p>
<h2>How products use CEFR levels</h2>
<h3>Placement</h3>
<p>A placement test estimates a learner’s level at the start, so that the first lessons are neither too easy nor too hard. Adaptive tests, which get harder or easier after each answer, reach an estimate in fewer questions. For young children, a short spoken conversation with a teacher often tells you more than a written test, because many of them cannot yet read well in any language. Treat the result as a starting point and adjust it after the first few lessons.</p>
<h3>Content</h3>
<p>Every lesson, text and activity should carry a level tag, together with the skills it practises and the can-do statement it works towards. Reference resources such as the English Vocabulary Profile and the English Grammar Profile, which map words and grammar points to CEFR levels, help writers check that a text fits its level.</p>
<h3>Progress</h3>
<p>Moving up a whole level takes months, so learners and parents need smaller markers along the way. Can-do checklists work well for this. “I can say what I like and don’t like” is something a parent can hear their child do. Show progress against the can-do statements of the current level as well as the level itself.</p>
<h2>Mapping lesson content to levels</h2>
<p>We’d recommend working through a content library in a fixed order.</p>
<ol>
<li>Choose the can-do statements you will teach at each level, written in language a learner or parent understands.</li>
<li>For each statement, list the vocabulary, grammar and functions a learner needs, checked against a reference such as the English Vocabulary Profile.</li>
<li>Tag each lesson with one main level, one or more can-do statements and the skills it practises.</li>
<li>Tag age suitability separately from level.</li>
<li>Have a teacher review a sample at each level, because automated text checks miss cultural references and how hard a task really is.</li>
<li>Re-tag after real use, when lessons that look like A2 on paper turn out to be harder in practice.</li>
</ol>
<p>In CurioKids, lesson content is generated for each child starting from their CEFR level, age and interests, and the teacher can change it. The level keeps the language right, the age keeps the activity right and the interests give the child something they want to talk about. We explain the approach in <a class="link" href="https://flavoureak.co.uk/blog/lesson-for-the-child/">a lesson for the child</a>, and the wider product in <a class="link" href="https://flavoureak.co.uk/blog/building-a-tutoring-platform/">building a tutoring platform</a>.</p>
<h2>Common mistakes with levels</h2>
<ul>
<li>Labelling content by grammar alone, so an A1 lesson asks for tasks only a B1 learner can do.</li>
<li>Treating levels as equal steps when the higher ones take far longer.</li>
<li>Showing young children a letter and a number, where a can-do statement or a picture would mean more to them.</li>
<li>Claiming a learner has “reached B2” on the strength of the product’s own quiz. Only a properly designed and validated assessment supports that kind of claim.</li>
<li>Forgetting that a learner’s skills can sit at different levels, for example B1 in reading and A2 in speaking.</li>
</ul>
<div class="callout"><b>Building an English learning product?</b> <a href="https://flavoureak.co.uk/curiokids/">See how CurioKids builds lessons from a child’s level</a>.</div>
<h2>Questions people ask</h2>
<h3>What are the CEFR levels?</h3>
<p>The CEFR has six main levels: A1 and A2 for basic users, B1 and B2 for independent users, and C1 and C2 for proficient users. Pre-A1 sits below A1 and is used mainly for young beginners. Each level is described by can-do statements about what a learner can do in the language.</p>
<h3>What is Pre-A1?</h3>
<p>Pre-A1 describes learners who rely on single words and memorised phrases and cannot yet build their own sentences. The Council of Europe added descriptors for it in the 2020 Companion Volume. Cambridge English Pre A1 Starters is an exam for young learners at this stage.</p>
<h3>Which CEFR level counts as fluent?</h3>
<p>The CEFR does not use “fluent” as a level. At B2 a learner can talk with fluent speakers without strain on either side, and at C1 they express themselves fluently and spontaneously for social, study and work purposes. Which counts as fluent depends on what the learner needs English for.</p>
<h3>How long does it take to move up a CEFR level?</h3>
<p>There is no fixed time. It depends on hours of study, use of the language outside lessons, age and the learner’s starting point. Higher levels take longer than lower ones, because each covers more language and harder tasks.</p>]]></content:encoded>
</item>
<item>
<title>What an event app needs, and what attendees rarely open</title>
<link>https://flavoureak.co.uk/blog/what-an-event-app-needs/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/what-an-event-app-needs/</guid>
<pubDate>Wed, 10 Jun 2026 08:00:00 GMT</pubDate>
<category>Events</category>
<description>An event app earns its place with the ticket, the schedule, the map and live updates. Networking and sponsor pages need a clear purpose, or they tend to sit unopened.</description>
<content:encoded><![CDATA[<ul><li>An event app needs four things attendees use at the venue: a ticket that opens without signal, a schedule with a personal agenda, a venue map and live updates.</li><li>Networking features only work when enough attendees opt in and there is a specific reason to connect, such as booking meetings at a trade show.</li><li>For a one-off event, a fast mobile website with Apple Wallet and Google Wallet passes often does the job, while a native app suits events the same people return to.</li></ul><p>An event app is software on an attendee’s phone that holds their ticket and tells them where to be and when. The features that earn their place are the ones people need at the venue: the ticket, the schedule with a personal agenda, a map and live updates. Networking, sponsor pages and gamification are the features most likely to go unopened unless each one is designed around a real reason to use it.</p>
<h2>The four features attendees use on the day</h2>
<p>Start from what an attendee does between arriving and leaving. They show a ticket at the door. They check what is on next and where. They find a room, a stage or the toilets. They find out when something changes. Every feature in an event app should support one of those four jobs, or have a clear reason to exist alongside them.</p>
<h3>Ticket wallet</h3>
<p>The ticket should open in two taps, with no login and no signal. The most reliable way to do that is a wallet pass. Apple Wallet and Google Wallet both support event tickets and store the barcode on the phone, so the pass opens even when the venue network is down. If the app has its own ticket screen as well, cache the barcode on the device the first time it loads. Our article on <a class="link" href="https://flavoureak.co.uk/blog/offline-event-check-in/">offline event check-in</a> covers the scanning side at the door.</p>
<h3>Schedule and personal agenda</h3>
<p>A schedule lists every session. A personal agenda is the attendee’s own shortlist, built by starring the sessions they want. The agenda is what people check during the day, so it should be one tap from the home screen and should flag a clash when two starred sessions overlap. Filters by day, track, stage and time matter more than clever recommendations. Each session needs a room, a start and end time, the speakers and a short description, all kept in step with the organiser’s master schedule.</p>
<h3>Venue map</h3>
<p>A map answers “where is Hall 3” faster than any signpost. It does not need to be an interactive 3D model. A clear floor plan for each level, with room names that match the schedule exactly, does the job. Link every session to its room on the map, and mark the things people look for most: entrances, toilets, food, the cloakroom and first aid. If the venue changes between years, the organiser must be able to replace the map without a new app release.</p>
<h3>Live updates</h3>
<p>Room changes, delays, cancelled sessions and safety messages need to reach people quickly. Push notifications do this best, with an announcements feed inside the app as the record of what was sent. Keep pushes rare and useful. An app that sends sponsor promotions every hour teaches people to mute it, and then the safety message goes unseen. We cover this in <a class="link" href="https://flavoureak.co.uk/blog/push-notifications-that-help/">push notifications that help</a>.</p>
<h2>Features that often go unopened</h2>
<p>Some features appear on every event app feature list and get little use in practice. They are not useless, but each one needs a plan.</p>
<ul>
<li><b>Attendee networking.</b> A directory of attendees only works if enough people opt in and complete a profile. At a small event most people would rather walk over and talk. Networking earns its place when it solves a specific problem, such as booking 15-minute meetings between buyers and suppliers at a trade show. Appearing in a directory shares personal data, so make it opt-in and explain who will see each profile.</li>
<li><b>Sponsor pages.</b> A list of sponsor logos with a paragraph each is rarely read. Sponsors usually get more from a session they host appearing in the agenda, or a stand marked on the map.</li>
<li><b>Gamification.</b> Points for visiting stands can work at a large exhibition with a prize worth chasing. At a one-day conference it is mostly noise.</li>
<li><b>Session chat and Q&amp;A.</b> Q&amp;A with upvoting works well when a moderator reads the questions out on stage. Open chat channels for each session often stay empty.</li>
<li><b>Social feeds.</b> An in-app feed competes with the social networks attendees already use, and it needs moderating.</li>
</ul>
<p>A useful test for any feature is to name the moment at the event when someone will open it, and what they will do in the next 30 seconds. If nobody on the team can name that moment, leave the feature out of the first version and add it when there is evidence people want it.</p>
<h2>Working offline at a busy venue</h2>
<p>Venue Wi-Fi and mobile networks struggle when thousands of phones share one building. An event app should assume a weak connection at exactly the moment it matters most. In practice that means five rules.</p>
<ol>
<li>Download the full schedule, speaker list and maps when the app is first opened, and refresh them in the background whenever there is a connection.</li>
<li>Store the ticket barcode on the device.</li>
<li>Save agenda changes on the phone first and sync them later.</li>
<li>Show when the schedule was last updated, so an attendee knows whether they are looking at the latest version.</li>
<li>Keep images small, so the first download finishes on a weak signal.</li>
</ol>
<p>This is the same approach we use in Estimark, which saves to the phone first and syncs later, because a building site has the same problem as a packed exhibition hall.</p>
<h2>Native app, mobile website or white-label platform</h2>
<p>There are three ways to put an event guide on an attendee’s phone. The right one depends mainly on how often your audience will use it.</p>
<table><thead><tr><th></th><th>Mobile website or PWA</th><th>White-label event platform</th><th>Custom native app</th></tr></thead><tbody>
<tr><td>Attendee has to install</td><td>No</td><td>Usually</td><td>Yes</td></tr>
<tr><td>Push notifications</td><td>Limited on iPhone</td><td>Yes</td><td>Yes</td></tr>
<tr><td>Works offline</td><td>Yes, if built for it</td><td>Depends on the platform</td><td>Yes, if built for it</td></tr>
<tr><td>Wallet tickets</td><td>Yes</td><td>Usually</td><td>Yes</td></tr>
<tr><td>Design and features</td><td>Full control</td><td>Within the platform’s options</td><td>Full control</td></tr>
<tr><td>Time to launch</td><td>Short</td><td>Shortest</td><td>Longest, including app store review</td></tr>
<tr><td>Best fit</td><td>Single events and one-off conferences</td><td>Organisers who want standard features now</td><td>Event series, membership bodies, products where the app is the business</td></tr>
</tbody></table>
<p>Requiring an install asks a lot of someone attending for one day. Many will not bother, and the organiser ends up running a printed programme alongside the app. A fast mobile website, linked from the confirmation email and a QR code at the door, reaches everyone. The main gap is notifications. On iPhone, a website can only send push notifications once the person has added it to their Home Screen. Our guide to <a class="link" href="https://flavoureak.co.uk/blog/pwa-or-native-app/">PWAs and native apps</a> goes into the detail.</p>
<p>A native app makes sense when the same people come back, for example to a festival series, an annual conference run by a membership body, or a venue with a season of events. The install is then paid for once, and the app becomes a channel the organiser owns all year.</p>
<h2>When to build and when to use a white-label platform</h2>
<p>White-label event platforms sell a ready-made app that carries your branding. They are quick to set up and include the standard features above. The trade-offs are that you work within their data model, their design options and their pricing. Attendee data also sits on their systems under their terms, which you need to check as part of your obligations as controller under UK GDPR.</p>
<p>A custom build makes sense when the app needs to do something the platforms do not, or when it must connect closely to your own systems. Signs that you are in that position:</p>
<ul>
<li>Your ticketing, membership or CRM system is your own, and the app must read and write to it in real time.</li>
<li>You run many events a year and per-event or per-attendee platform charges keep adding up.</li>
<li>The app is part of what you sell, for example meeting slots at a trade show or cashless top-ups at a festival.</li>
<li>You need offline behaviour or accessibility that the platforms you have tried do not provide.</li>
</ul>
<p>Whichever route you take, test on a crowded network before the event, check the app with a screen reader and large text settings, and agree who on the organiser’s team can send a push notification on the day.</p>
<div class="callout"><b>Planning an event app or attendee portal?</b> <a href="https://flavoureak.co.uk/events/">See how we approach event software</a>.</div>
<h2>Questions people ask</h2>
<h3>What features does an event app need?</h3>
<p>An event app needs a ticket that opens without signal, a schedule with a personal agenda, a venue map and live updates through push notifications or an announcements feed. These cover what attendees actually do at the venue. Networking, sponsor pages and gamification are optional and should be added only with a specific use in mind.</p>
<h3>Do attendees download event apps?</h3>
<p>Installing an app is a bigger ask for a single day than for a multi-day event or one people attend every year. For a one-off event, a mobile website reached by a link or QR code removes the install step entirely. Whichever you choose, also offer the ticket as an Apple Wallet or Google Wallet pass.</p>
<h3>Should an event app work offline?</h3>
<p>Yes. Venue networks are often overloaded when an event is busiest, which is when attendees need the app. The ticket, schedule and maps should be stored on the phone, with any changes synced when a connection returns.</p>
<h3>Is a white-label event app good enough?</h3>
<p>For many organisers it is, especially for standard conferences that need a schedule, a map and a ticket. It becomes limiting when the app must connect closely to your own ticketing or membership systems, or when you need features the platform does not offer. Check where attendee data is stored, and on what terms, before signing.</p>]]></content:encoded>
</item>
<item>
<title>Job costing for UK builders: tracking costs against the estimate</title>
<link>https://flavoureak.co.uk/blog/job-costing-for-builders/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/job-costing-for-builders/</guid>
<pubDate>Thu, 04 Jun 2026 08:00:00 GMT</pubDate>
<category>Construction</category>
<description>Job costing compares what a job is costing with what you estimated, while it is still running. Track labour, materials, plant, subcontractors and overheads weekly, and forecast the final margin.</description>
<content:encoded><![CDATA[<ul><li>Job costing is tracking the actual costs of a construction job, by cost type, against the estimate while the work is still in progress.</li><li>The figure that matters is the forecast final cost, meaning cost to date plus the estimated cost to complete, compared with the contract sum including agreed variations.</li><li>Variations should be priced and agreed in writing before the work is done, or the cost lands on the job and the margin shrinks.</li></ul><p>Job costing for builders means recording every cost on a job, sorted into labour, materials, plant, subcontractors and overheads, and comparing it with the estimate while the job is still running. The aim is to forecast the final margin early enough to act on it. A builder who only works out the profit when the final account is settled finds out about an overrun when it is too late to recover it.</p>
<h2>What job costing is</h2>
<p>A job cost is any cost that can be traced to a specific job. Job costing groups those costs under the same headings as the estimate, so each line can be compared four ways: estimated, spent to date, forecast to complete and forecast final cost. The difference between the contract sum and the forecast final cost is the forecast margin.</p>
<p>The habit matters more than the software. A weekly review of each live job, with the site manager’s view of what is left to do, catches most problems while there is still time to change the plan, talk to the client or price a variation.</p>
<h2>The cost headings</h2>
<table>
<thead><tr><th>Heading</th><th>What goes in it</th><th>Where the figure comes from</th></tr></thead>
<tbody>
<tr><td>Labour</td><td>Your own operatives’ time at their full cost rate</td><td>Timesheets or clocking in, by job</td></tr>
<tr><td>Materials</td><td>Everything bought for the job, less returns</td><td>Purchase orders, delivery notes, supplier invoices</td></tr>
<tr><td>Plant</td><td>Hired and owned plant, skips, scaffold, welfare</td><td>Hire invoices and internal plant rates</td></tr>
<tr><td>Subcontractors</td><td>Each package let to another firm</td><td>Subcontract orders, applications and payments</td></tr>
<tr><td>Overheads</td><td>A share of office, vehicles, insurance and management</td><td>An agreed allocation, often a percentage of value</td></tr>
</tbody>
</table>
<h3>Labour at full cost</h3>
<p>The most common error is to cost labour at the hourly wage. An operative’s real cost includes employer’s National Insurance, pension contributions, holiday pay, sick pay, travel and non-productive time. As an example, if an operative is paid £18.00 an hour and those extras add £5.40, the cost rate is £23.40 an hour. Estimating and job costing should both use the full rate, or every job will look more profitable than it is. A list of labour rates by trade and grade keeps this consistent.</p>
<h3>Materials when committed</h3>
<p>Record materials when they are ordered, without waiting for the supplier’s invoice a month later. A purchase order against the job shows the commitment at once, so the forecast reflects it. Match delivery notes to orders, and record returns and credits against the same job.</p>
<h2>A worked example</h2>
<p>A contractor is refurbishing a ground-floor office for a commercial client for a contract sum of £60,000, excluding VAT. The estimate breaks down as labour £18,000, materials £16,500, plant and skips £2,500, subcontractors £10,000 and overheads £4,800 (8% of the contract sum). The estimated cost is £51,800, leaving a margin of £8,200, or 13.7% of the contract sum.</p>
<p>Midway through the job, the site manager reviews each line, recording what has been spent and estimating what is left to do.</p>
<table>
<thead><tr><th>Cost line</th><th>Estimate</th><th>Spent to date</th><th>Forecast to complete</th><th>Forecast final</th><th>Variance</th></tr></thead>
<tbody>
<tr><td>Labour</td><td>£18,000</td><td>£11,400</td><td>£8,100</td><td>£19,500</td><td>−£1,500</td></tr>
<tr><td>Materials</td><td>£16,500</td><td>£9,800</td><td>£6,200</td><td>£16,000</td><td>+£500</td></tr>
<tr><td>Plant and skips</td><td>£2,500</td><td>£1,700</td><td>£900</td><td>£2,600</td><td>−£100</td></tr>
<tr><td>Subcontractors</td><td>£10,000</td><td>£4,200</td><td>£6,300</td><td>£10,500</td><td>−£500</td></tr>
<tr><td>Overheads</td><td>£4,800</td><td>£2,640</td><td>£2,160</td><td>£4,800</td><td>£0</td></tr>
<tr><td><b>Total</b></td><td><b>£51,800</b></td><td><b>£29,740</b></td><td><b>£23,660</b></td><td><b>£53,400</b></td><td><b>−£1,600</b></td></tr>
</tbody>
</table>
<p>The forecast final cost is £53,400, so the forecast margin is £60,000 − £53,400 = £6,600, or 11.0%. The job is still profitable, and £1,600 of margin has gone. The table shows where. Labour is £1,500 over because stripping out took longer than planned, and the electrical subcontractor has asked for £500 more for extra containment.</p>
<p>That is the moment to act. The site manager checks whether the extra strip-out came from conditions the contract puts at the client’s risk, and whether the extra containment was a design change. If either was, it should be priced and claimed as a variation.</p>
<h2>Variations and how they affect margin</h2>
<p>A variation is a change to the work in the contract, such as an addition, an omission or a change of specification, usually instructed by the client or their representative. Variations are where margin is most often lost, because work gets done on a verbal request and priced later, if at all.</p>
<p>Continuing the example, the client asks for an extra glazed partition. The contractor prices it at £2,400, made up of £1,800 cost and £600 margin, and the client confirms in writing before the work starts. The contract sum becomes £62,400 and the forecast final cost becomes £55,200. The forecast margin is £7,200, or 11.5%. If the same work had been done without agreement and never paid for, the margin would have fallen to £4,800, or 8.0% of the original sum.</p>
<ul>
<li>Get every instruction in writing, even if it is a text or email confirming a conversation.</li>
<li>Price the variation and get agreement before doing the work, where the contract allows.</li>
<li>Give each variation its own reference and cost code, so its cost is tracked separately.</li>
<li>Include agreed variations in the next valuation instead of saving them for the final account.</li>
</ul>
<h2>Valuations and applications for payment</h2>
<p>On most commercial jobs the contractor is paid in stages. An application for payment, or valuation, sets out the value of work done to date and claims the amount due. Job costing measures cost and valuations measure value, and the two should be read together.</p>
<p>In the example, at the midway review the contractor values the work done at £33,000. The contract has a retention of 5%, which the client holds back, usually releasing part at practical completion and the rest after the defects period. Previous applications have been paid at £22,000.</p>
<table>
<thead><tr><th>Line</th><th>Amount</th></tr></thead>
<tbody>
<tr><td>Gross value of work done to date</td><td>£33,000</td></tr>
<tr><td>Less retention at 5%</td><td>−£1,650</td></tr>
<tr><td>Net valuation</td><td>£31,350</td></tr>
<tr><td>Less previously paid</td><td>−£22,000</td></tr>
<tr><td><b>Due this application, excluding VAT</b></td><td><b>£9,350</b></td></tr>
</tbody>
</table>
<p>Comparing value with cost shows whether the job is earning its keep as it goes. Here the value to date is £33,000 against costs of £29,740, a margin to date of £3,260, or 9.9%. That is below the forecast 11.0% for the whole job, which suggests either the valuation is light or costs have come early, and it deserves a second look. The rules on payment notices, CIS deductions and the VAT reverse charge are covered in <a class="link" href="https://flavoureak.co.uk/blog/construction-rules/">CIS, reverse charge and payment notices</a>.</p>
<h2>Reviewing jobs weekly</h2>
<ol>
<li>Post all timesheets, purchase orders and subcontract costs to the job by the end of each week.</li>
<li>Ask the site manager for the cost to complete on each line, based on what is left to do.</li>
<li>Compare the forecast final cost with the contract sum plus agreed variations.</li>
<li>Investigate any line that has moved by more than an amount you set in advance.</li>
<li>Decide what to do, whether that is changing the method, pricing a variation, talking to the client or accepting the loss knowingly.</li>
<li>Record the forecast margin so you can see how it moves from week to week.</li>
</ol>
<h2>Spreadsheet or software</h2>
<p>A spreadsheet works for one or two jobs if someone keeps it up to date. It breaks down when costs come from several people, when the estimate lives in a different file, and when variations, valuations and invoices are tracked somewhere else again. Software that keeps the estimate, purchase orders, timesheets, variations and invoices in one record per job removes the re-keying that makes figures go stale.</p>
<p>Estimark is built around one record per job, from the estimate through scheduling to invoicing, with a price book, labour rates and reports. For choosing a system, see <a class="link" href="https://flavoureak.co.uk/blog/choosing-estimating-software/">how to choose estimating and quoting software</a>.</p>
<div class="callout"><b>Need job costing that fits how you work?</b> See <a class="link" href="https://flavoureak.co.uk/estimark/">Estimark</a>, or talk to us about <a class="link" href="https://flavoureak.co.uk/construction/">construction software built around your process</a>.</div>
<h2>Questions people ask</h2>
<h3>What is a good margin for a building job?</h3>
<p>There is no single right margin for a building job, because it depends on the type of work, the risk, the client and how overheads are recovered. What matters is knowing the margin in the estimate and tracking whether the job is achieving it. Compare the forecast final margin with the estimated margin on every job.</p>
<h3>Should overheads be included in job costs?</h3>
<p>Including a share of overheads in job costs shows whether each job contributes enough to run the business. A common method is to allocate overheads as a percentage of each job’s value, based on annual overheads divided by annual turnover. Some builders track direct costs and overheads separately and report both margins.</p>
<h3>What is the difference between a valuation and an invoice?</h3>
<p>A valuation, or application for payment, sets out the value of work done to date and the amount claimed. The invoice is the document used for VAT and accounting. On many construction contracts the invoice is issued once the amount due has been agreed or notified.</p>
<h3>How often should I review job costs?</h3>
<p>Review job costs weekly on any live job of meaningful size. Weekly reviews catch overruns while there is still time to act. Monthly reviews often find problems after the money has already been spent.</p>]]></content:encoded>
</item>
<item>
<title>How to measure traffic from ChatGPT, Perplexity and AI Overviews</title>
<link>https://flavoureak.co.uk/blog/measuring-ai-search-traffic/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/measuring-ai-search-traffic/</guid>
<pubDate>Fri, 29 May 2026 08:00:00 GMT</pubDate>
<category>GEO</category>
<description>Some AI search visits show up as referrals, many arrive as direct traffic, and Google counts AI Overviews inside normal search data. Combine referrers, UTMs, server logs and a prompt list.</description>
<content:encoded><![CDATA[<ul><li>Visits from ChatGPT, Perplexity, Copilot, Gemini and Claude can be identified in analytics by their referring domains, but some arrive with no referrer and are counted as direct.</li><li>Google counts clicks and impressions from AI Overviews in Search Console’s ordinary web search figures, so AI Overview traffic is hard to isolate there.</li><li>Server logs show which AI crawlers and user-triggered fetchers read your pages, and a monthly prompt list shows whether answers mention and cite you.</li></ul><p>To measure traffic from AI search, combine referral data in your analytics, UTM parameters that some engines add to links, server logs that show AI crawlers and fetchers, and a regular check of whether answers mention and cite you. AI search traffic means visits that arrive after someone reads an answer in a tool such as ChatGPT, Perplexity, Microsoft Copilot, Gemini, Claude or Google’s AI Overviews and clicks a cited link. No single report captures all of it, so aim for a consistent estimate you can track over time.</p>
<h2>What you can and cannot see</h2>
<p>An AI answer can affect a business in three ways. Someone clicks a citation and visits the site. Someone reads your name in the answer and searches for you later, or types your address. Or the engine reads your page and uses it without anyone visiting. Analytics can only see the first, and only partly. The second shows up as branded search or direct traffic. The third leaves traces in server logs and nowhere else.</p>
<table>
<thead><tr><th>Source</th><th>What it shows</th><th>What it misses</th></tr></thead>
<tbody>
<tr><td>Analytics referrers</td><td>Clicks from AI tools that pass a referrer</td><td>Clicks from apps or settings that strip it</td></tr>
<tr><td>UTM parameters</td><td>Clicks from engines that tag their links</td><td>Engines that don’t tag, and copied links</td></tr>
<tr><td>Server logs</td><td>Crawlers and fetchers reading your pages</td><td>Whether anyone saw or clicked the answer</td></tr>
<tr><td>Search Console</td><td>Google search clicks and impressions, AI features included</td><td>A separate figure for AI Overviews</td></tr>
<tr><td>Prompt tracking</td><td>Whether answers mention or cite you</td><td>How many people asked those questions</td></tr>
</tbody>
</table>
<h2>Referrers in analytics</h2>
<p>When someone clicks a link in a browser, the browser usually tells the destination site which page they came from. That information is the referrer. Clicks from the web versions of AI tools often arrive with a referrer that names the tool, so they can be grouped in analytics.</p>
<p>The domains to look for include chatgpt.com (and the older chat.openai.com), perplexity.ai, copilot.microsoft.com, gemini.google.com and claude.ai. In Google Analytics 4 you can create a custom channel group with a rule that matches these sources, so AI referrals appear as their own channel instead of being mixed into general referrals. A regular expression such as <code>chatgpt|openai|perplexity|copilot|gemini|claude</code> on the session source is a reasonable start. Review it every few months, because new tools and domains appear.</p>
<h3>The limits of referrers</h3>
<p>Referrer data undercounts. Desktop and mobile apps may open links without passing a referrer. Browser privacy settings and some sites’ referrer policies remove or shorten it. People copy a link from an answer and paste it into a new tab. In each case the visit lands in direct traffic.</p>
<p>Google AI Overviews are a separate problem. A click from an overview arrives with a Google referrer and looks like any other organic visit from Google. Treat the AI channel in analytics as a floor. A rise in it is a real signal, and a flat line doesn’t prove that nothing is happening.</p>
<h2>UTM parameters some engines add</h2>
<p>A UTM parameter is a tag on the end of a link, such as <code>?utm_source=chatgpt.com</code>, that analytics tools read to attribute a visit. ChatGPT adds <code>utm_source=chatgpt.com</code> to many of the links it shows, which makes those visits identifiable even when the referrer is missing. Other engines may or may not tag their links, and the behaviour can change without notice.</p>
<p>Two practical points follow. Make sure your analytics keeps UTM parameters and nothing strips query strings before they are recorded. And make sure your site treats a tagged address as the same page as the untagged one. A canonical tag on each page tells search engines which address is the real one, and keeps one page’s data from splitting across many rows.</p>
<h2>Server logs and AI crawlers</h2>
<p>Your server or CDN logs record every request, including requests from automated agents. The main AI companies identify their agents by a user-agent name. Several run more than one agent, each with a different job, such as collecting training data, building a search index, or fetching a page when a user’s question needs it.</p>
<table>
<thead><tr><th>Company</th><th>Agents to look for</th></tr></thead>
<tbody>
<tr><td>OpenAI</td><td>GPTBot, OAI-SearchBot, ChatGPT-User</td></tr>
<tr><td>Anthropic</td><td>ClaudeBot, Claude-SearchBot, Claude-User</td></tr>
<tr><td>Perplexity</td><td>PerplexityBot, Perplexity-User</td></tr>
<tr><td>Microsoft</td><td>Bingbot, which feeds Bing and Copilot</td></tr>
<tr><td>Google</td><td>Googlebot, which feeds Search, including AI Overviews</td></tr>
</tbody>
</table>
<p>The user-triggered fetchers are the most interesting for measurement. A request from ChatGPT-User or Claude-User usually means a person asked a question and the assistant went to read your page to answer it. Counting those requests by page shows which content is being used in answers, even when nobody clicks through.</p>
<p>User-agent names can be faked, so verify important findings. Several operators publish the IP ranges their agents use, and a reverse DNS lookup confirms genuine Googlebot and Bingbot requests. For how to allow or block these agents, see <a class="link" href="https://flavoureak.co.uk/blog/ai-crawlers-robots/">AI crawlers and robots.txt</a>.</p>
<h2>What Google Search Console separates and what it doesn’t</h2>
<p>Google’s documentation says that traffic from its AI features, including AI Overviews and AI Mode, is counted in the Search Console Performance report under the Web search type. Your Search Console totals therefore include it. There is no standard filter that shows AI Overview clicks on their own, so you cannot read off how many visits came from an overview. Google changes Search Console regularly, so check the current report before relying on this.</p>
<p>What you can do is watch for patterns. A query whose impressions rise while clicks fall may now show an AI Overview that answers it on the results page. Searching those queries yourself shows whether an overview appears and whether your page is among its links. The guide to <a class="link" href="https://flavoureak.co.uk/blog/google-ai-overviews/">Google AI Overviews</a> covers what they change for business sites.</p>
<h2>Tracking mentions in answers</h2>
<p>Many people read an answer and never click. To see whether you are part of those answers, ask the engines directly and record what they say. This can be done by hand or with one of the commercial tools that run prompts at scale, and the method is the same either way.</p>
<ol>
<li>Write a list of 20 to 50 questions buyers ask before they choose a supplier like you, in their own words.</li>
<li>Run each question in the engines your customers use, signed out where possible, on a fixed day each month.</li>
<li>Record whether your business is mentioned, whether your site is cited, and which other sources are cited.</li>
<li>Save the answer text and the date, because answers change from one run to the next.</li>
<li>Compare month on month, noting which of your pages are cited and which competitors appear in your place.</li>
</ol>
<p>Answers vary between runs, accounts and locations, so treat one result as a sample. Trends over several months are more reliable than any single answer.</p>
<h2>Putting it into one monthly report</h2>
<p>A short monthly report keeps the numbers honest. We would include AI referral sessions and their conversions from analytics, branded search impressions from Search Console, requests from user-triggered AI fetchers by page from the logs, and the mention and citation counts from the prompt list. None of these is complete alone. Together they show direction, which is what a business needs to decide whether its content work is paying off.</p>
<p>Also ask new enquiries how they found you, and add “an AI assistant such as ChatGPT” as an option on contact forms. Self-reported sources are imperfect, but they catch visits that analytics files as direct.</p>
<div class="callout"><b>Want AI search measured properly?</b> Our <a class="link" href="https://flavoureak.co.uk/geo/">GEO guide</a> explains how we set up tracking and content for answer engines.</div>
<h2>Questions people ask</h2>
<h3>Why does ChatGPT traffic show as direct in my analytics?</h3>
<p>ChatGPT traffic shows as direct when the visit arrives without a referrer or a UTM tag. This happens when links open from an app, when browser privacy settings strip the referrer, or when someone copies the link into a new tab. Analytics then has no record of where the visitor came from.</p>
<h3>Can I see AI Overview clicks in Google Search Console?</h3>
<p>Google counts clicks and impressions from AI Overviews within the Web search type in the Search Console Performance report. They are included in the totals and are not shown as a separate figure. Checking key queries by hand is the practical way to see whether an overview cites you.</p>
<h3>Which AI crawlers should I look for in my logs?</h3>
<p>The main AI agents are OpenAI’s GPTBot, OAI-SearchBot and ChatGPT-User, Anthropic’s ClaudeBot, Claude-SearchBot and Claude-User, and Perplexity’s PerplexityBot and Perplexity-User. Googlebot and Bingbot also matter, because Google’s AI features and Microsoft Copilot draw on their search indexes. Verify important findings, because user-agent names can be faked.</p>
<h3>Is it worth paying for an AI visibility tool?</h3>
<p>An AI visibility tool saves time when you track many questions across several engines. For a small business, a monthly manual check of 20 to 50 questions in a spreadsheet gives much of the same insight. Whichever you choose, keep the question list stable so the results can be compared over time.</p>]]></content:encoded>
</item>
<item>
<title>How to redesign or migrate a website without losing search traffic</title>
<link>https://flavoureak.co.uk/blog/redesign-without-losing-traffic/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/redesign-without-losing-traffic/</guid>
<pubDate>Sun, 24 May 2026 08:00:00 GMT</pubDate>
<category>SEO</category>
<description>Search traffic is usually lost in a redesign through changed URLs, missing redirects and removed content. Map every old URL, redirect with 301s, keep what ranks and watch Search Console.</description>
<content:encoded><![CDATA[<ul><li>Before a redesign, list every URL that has traffic, links or rankings, and map each one to its closest equivalent on the new site.</li><li>Redirect every changed URL with a permanent 301 redirect to its most relevant new page, and keep the redirects in place for at least a year.</li><li>For a move to a new domain, use the Change of Address tool in Google Search Console once the redirects are live.</li></ul><p>You keep search traffic through a redesign by making sure every page that earns traffic today either keeps its URL or permanently redirects to an equivalent page, and by keeping the content that made it rank. A site migration is any change to a site’s URLs, domain, structure or platform that search engines have to relearn. The work is mostly an inventory, a redirect map and careful checks on launch day and in the weeks after.</p>
<h2>Why redesigns lose traffic</h2>
<p>Search engines rank individual URLs. When a URL disappears, its rankings and the links pointing to it have nowhere to go unless a redirect says where the page now lives. The usual causes of a drop after a relaunch are a short list.</p>
<ul>
<li>URLs changed without redirects, so old links and rankings lead to 404 pages</li>
<li>Pages that ranked were removed, merged badly or cut down to a few lines</li>
<li>Titles, headings and internal links all changed at once</li>
<li>The staging site’s noindex tag or robots.txt block went live with the new site</li>
<li>The new site draws its content with JavaScript that crawlers can’t see</li>
</ul>
<p>Every one of these is preventable. The safest migration changes as few things at once as possible. If you can keep the URLs the same and change only the design, do that.</p>
<h2>Build a full URL inventory</h2>
<p>Start with a list of every URL that matters on the current site. No single source has them all, so combine several:</p>
<ul>
<li>a crawl of the live site with a site crawler tool</li>
<li>your XML sitemaps</li>
<li>pages with clicks or impressions in Search Console’s Performance report, which keeps 16 months of data</li>
<li>landing pages from your analytics</li>
<li>pages with links from other sites, from Search Console’s Links report</li>
<li>PDFs, images and other files that receive traffic or links</li>
</ul>
<p>Put them in one spreadsheet, with columns for traffic, external links and the new URL. That spreadsheet becomes the redirect map. Before launch, also record benchmarks for the pages that matter most, such as their clicks, impressions and average position, so you have something to compare against afterwards.</p>
<h2>Write the 301 redirect map</h2>
<p>A 301 redirect is a permanent redirect. It tells browsers and search engines that a page has moved for good, and Google treats it as a strong signal that the new URL should replace the old one in its index. Map each old URL to the single most relevant new page.</p>
<table><thead><tr><th>Old page</th><th>Situation on the new site</th><th>What to do</th></tr></thead><tbody>
<tr><td>Same content, new URL</td><td>A direct equivalent exists</td><td>301 to the new URL</td></tr>
<tr><td>Two pages merged into one</td><td>A combined page covers both</td><td>301 both old URLs to the combined page</td></tr>
<tr><td>Product or service discontinued</td><td>A close replacement or category exists</td><td>301 to the closest relevant page</td></tr>
<tr><td>Content removed with no equivalent</td><td>Nothing relevant on the new site</td><td>Return 404 or 410; don’t redirect to the home page</td></tr>
<tr><td>URL unchanged</td><td>Same page, new design</td><td>No redirect; check the content survived</td></tr>
</tbody></table>
<p>Avoid sending everything to the home page. Google may treat a redirect to an unrelated page as a soft 404, so the old page’s value is lost anyway. Redirect straight to the final URL with no chains of one redirect leading to another, and update redirects left over from earlier redesigns so they point at the new destinations. Google advises keeping redirects for as long as possible, and generally for at least a year. Keeping them permanently is safer, because other sites may link to the old addresses for years.</p>
<h2>Keep the content, titles and canonical tags</h2>
<p>A redesign is a common moment to cut copy, and cutting copy from pages that rank is one of the quickest ways to lose traffic. Keep the substance of each ranking page, including its main headings, its answers to the questions people search for and its links to related pages. Improve it if you like, but check that nothing a searcher relied on has gone.</p>
<p>Then check the technical details page by page. Each new page should have a canonical tag pointing to its own new URL, never to the old one. Titles and meta descriptions should carry over unless there is a reason to improve them. Internal links should point straight at the new URLs instead of relying on redirects, which is part of good <a class="link" href="https://flavoureak.co.uk/blog/internal-linking/">internal linking</a>. Structured data, hreflang tags, image URLs and analytics tracking need the same check.</p>
<h2>Sitemaps and Change of Address for domain moves</h2>
<p>Publish an XML sitemap listing only the new URLs, and submit it in Search Console on launch day. It helps Google discover the new pages quickly. Google will recrawl the old URLs over time and follow the redirects it finds there.</p>
<p>If you are moving to a new domain, verify both domains in Search Console, put the 301 redirects live, and then use the Change of Address tool in the old domain’s property. It tells Google the whole site has moved. The tool is meant for moves between domains or subdomains. It isn’t needed for a switch from HTTP to HTTPS, and it doesn’t apply to URL changes within the same domain.</p>
<h2>A step-by-step migration checklist</h2>
<ol>
<li>Record benchmarks for traffic, key rankings and the number of indexed pages.</li>
<li>Build the URL inventory and the redirect map, and have someone check it.</li>
<li>Protect the staging site with a password so it can’t be indexed.</li>
<li>Crawl the staging site to check titles, canonical tags, internal links and content.</li>
<li>Test every redirect in the map on staging, looking for chains, loops and 404s.</li>
<li>Launch at a quiet time for the business, with the team available that day.</li>
<li>Check straight away that no noindex tag or robots.txt block has come across from staging.</li>
<li>Submit the new sitemap and, for a domain move, use Change of Address.</li>
<li>Crawl the full list of old URLs against the live site and confirm each one redirects correctly.</li>
<li>Inspect the most important pages with the URL Inspection tool.</li>
<li>Monitor Search Console and analytics daily for the first two weeks, then weekly.</li>
</ol>
<h2>Monitoring after launch</h2>
<p>Expect some movement. Rankings can shift for a few weeks while search engines recrawl the site and process the redirects, and larger sites take longer. What matters is anything that doesn’t recover.</p>
<p>In Search Console, watch the Page indexing report for unexpected rises in “Not found (404)” and “Soft 404”, and use the Performance report to compare clicks for your key pages before and after. Check server logs or a fresh crawl for redirect errors, and fix them the same day. If a page that used to rank has dropped and stayed down, compare the old and new versions side by side. The cause is often something that was removed. New sites built on a JavaScript framework need extra checks, set out in <a class="link" href="https://flavoureak.co.uk/blog/seo-for-javascript-sites/">SEO for JavaScript sites</a>.</p>
<div class="callout"><b>Planning a redesign or a move to a new platform?</b> We treat the URL inventory and redirect map as part of the build, with the same care as the design. <a class="link" href="https://flavoureak.co.uk/seo/">See our SEO guide</a>.</div>
<h2>Questions people ask</h2>
<h3>Will a website redesign affect SEO?</h3>
<p>It can, in either direction. If URLs, content and internal links are kept or properly redirected, rankings usually hold and can improve with a faster, clearer site. If pages disappear without redirects or lose their content, traffic to them falls.</p>
<h3>How long should 301 redirects stay in place?</h3>
<p>Google recommends keeping redirects for as long as possible, and generally for at least a year. Keeping them permanently is safer, because other websites, emails and bookmarks may point at the old URLs for many years.</p>
<h3>How long does it take for traffic to settle after a migration?</h3>
<p>A well-planned migration often settles within a few weeks, as search engines recrawl the old URLs and follow the redirects. Larger sites take longer. A drop that lasts well beyond that usually points to missing redirects or removed content.</p>
<h3>Should we change domain and redesign at the same time?</h3>
<p>Separating them is safer, because if traffic falls you can tell which change caused it. If you must do both together, the redirect map, content checks and monitoring matter even more.</p>]]></content:encoded>
</item>
<item>
<title>PWA or native app: what a progressive web app can and can’t do</title>
<link>https://flavoureak.co.uk/blog/pwa-or-native-app/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/pwa-or-native-app/</guid>
<pubDate>Mon, 18 May 2026 08:00:00 GMT</pubDate>
<category>Engineering</category>
<description>A progressive web app installs from the browser, works offline and can send push notifications, including on iPhone. Native apps still lead on device access, background work and store presence.</description>
<content:encoded><![CDATA[<ul><li>A progressive web app is a website that can be installed to the home screen, work offline through a service worker and send push notifications.</li><li>On iPhone, a PWA is installed through the browser’s Add to Home Screen option, and it can send push notifications only after it has been installed.</li><li>Choose native when you need Bluetooth, background location or background tasks, dependable offline storage for large amounts of data, or a listing in both app stores.</li></ul><p>A progressive web app (PWA) is a website built so it can be installed to a phone’s home screen, open in its own window, work offline and send push notifications, all from a single web codebase. On Android, a PWA comes close to a native app for most business uses. On iPhone it can install, work offline and receive push notifications, but it has less access to the device and almost no ability to run in the background, so native apps remain the better choice for anything that depends on hardware, background work or store presence.</p>
<h2>What makes a website a progressive web app</h2>
<p>Three pieces turn a website into a PWA. A web app manifest is a small file that gives the app its name, icon, colours and how it should open. A service worker is a script the browser runs separately from the page, which can serve pages and data from a cache when there is no connection and receive push messages. HTTPS is required for both. Everything else is ordinary web development, which is why one team can build a PWA for every platform at once.</p>
<p>A PWA is still a website. People can use it in a browser tab without installing anything, share a link to any screen and get every update the next time they open it. On a desktop computer, Chrome and Edge can install it too, and Safari on a Mac can add it to the Dock.</p>
<h2>What a PWA can do on Android and iPhone</h2>
<p>The gap between the two platforms is wider than the gap between a PWA and a native app on Android. This table reflects current versions of Chrome on Android and Safari on iOS. Browser support changes with each release, so check the specific features you need before you commit.</p>
<table><thead><tr><th>Capability</th><th>PWA on Android</th><th>PWA on iPhone</th><th>Native app</th></tr></thead><tbody><tr><td>Install to home screen</td><td>Yes, and the browser can prompt</td><td>Yes, manually through the browser’s Add to Home Screen option</td><td>Yes, from the App Store or Google Play</td></tr><tr><td>Push notifications</td><td>Yes</td><td>Yes, once added to the home screen and permission is granted</td><td>Yes</td></tr><tr><td>Works offline</td><td>Yes, through a service worker</td><td>Yes, through a service worker, though stored data can be cleared</td><td>Yes, with full control of local storage</td></tr><tr><td>Background sync</td><td>Partly, in Chrome</td><td>No</td><td>Yes, within the operating system’s limits</td></tr><tr><td>Location while closed</td><td>No</td><td>No</td><td>Yes, with permission</td></tr><tr><td>Bluetooth and NFC</td><td>Yes, through Web Bluetooth and Web NFC in Chrome</td><td>No</td><td>Yes</td></tr><tr><td>Camera and photos</td><td>Yes</td><td>Yes</td><td>Yes</td></tr><tr><td>Passkeys with Face ID or fingerprint</td><td>Yes, through WebAuthn</td><td>Yes, through WebAuthn</td><td>Yes, plus wider use of biometrics</td></tr><tr><td>Home screen widgets</td><td>No</td><td>No</td><td>Yes</td></tr><tr><td>Store listing</td><td>Google Play, packaged as a Trusted Web Activity</td><td>Only inside a native wrapper</td><td>Yes</td></tr><tr><td>Updates</td><td>Instant, on the next visit</td><td>Instant, on the next visit</td><td>Through store review, then installed by the user</td></tr></tbody></table>
<h2>The iPhone limits in detail</h2>
<p>Apple has added many PWA features to Safari in recent years, and web push, which arrived in iOS 16.4, closed the largest gap. Several limits remain that matter in practice.</p>
<ul><li><b>No install prompt.</b> People have to find the Add to Home Screen option themselves. Many need to be shown how, so your site has to explain it.</li><li><b>Push only after install.</b> A PWA cannot ask for notification permission until it has been added to the home screen, and the request has to follow a tap or similar action from the person.</li><li><b>Storage can be cleared.</b> Safari may delete data stored by websites that have not been used for a while. Web apps added to the home screen are treated more generously, but storage on iPhone is still less dependable than in a native app, which matters for anything offline-first.</li><li><b>Little background work.</b> A PWA on iPhone cannot sync in the background or follow the person’s location while it is closed.</li></ul>
<p>What a PWA can do on iPhone depends on what Apple supports in WebKit, the engine behind Safari, so the list can grow with each iOS release. Plan around what works today.</p>
<h2>Store presence and discovery</h2>
<p>Being in the App Store and Google Play matters for some products and not at all for others. A consumer app that people search for by name benefits from a store listing, ratings and the trust that comes with them. Internal tools, customer portals and booking systems are usually reached through a link in an email or on a website, so a store listing adds little.</p>
<p>On Android, a PWA can be published to Google Play by packaging it as a Trusted Web Activity, which shows the web app full screen inside a thin native shell. On iPhone there is no equivalent. Wrapping a website in a web view and submitting it risks rejection under Apple’s minimum functionality guideline, as covered in our guide to <a class="link" href="https://flavoureak.co.uk/blog/app-store-review-first-time/">passing app review first time</a>.</p>
<h2>Cost and maintenance</h2>
<p>A PWA is usually the cheaper option to build and to run. There is one codebase, built with web skills, and every user gets an update the next time they open it, with no store review and no old versions left in use. Payments taken on the web follow your payment provider’s terms instead of the stores’ in-app purchase rules.</p>
<p>Native apps cost more because there are two platforms to build for, or one cross-platform codebase with platform-specific work, plus store accounts, review and a yearly cycle of operating system changes. The trade-offs between separate native apps and a shared codebase are covered in <a class="link" href="https://flavoureak.co.uk/blog/native-or-cross-platform/">native or cross-platform</a>.</p>
<h2>When each one fits</h2>
<h3>Choose a PWA when</h3>
<ul><li>People reach the product through links, email or a website, rather than by searching a store.</li><li>The main tasks are forms, lists, bookings, dashboards and messages.</li><li>You want to test demand before paying for native apps.</li><li>Offline use is occasional and the data involved is small.</li></ul>
<h3>Choose native when</h3>
<ul><li>The app depends on Bluetooth, NFC, background location or other device features, especially on iPhone.</li><li>People must work offline for long periods with a lot of data, such as records and photos on a building site.</li><li>Notifications and background updates are central to the product on iPhone.</li><li>Your customers expect to find you in the app stores.</li></ul>
<p>Estimark sits in the second group. It has web, iPhone and Android apps, and it works offline on site by saving to the phone first and syncing later, the pattern described in <a class="link" href="https://flavoureak.co.uk/blog/offline-first-mobile/">offline first</a>. On iPhone, a PWA could not make the same promise about work stored on the device.</p>
<p>The choice is also reversible. A product can start as a PWA and add native apps later, reusing the same API, data and design, and the web version can stay for people who work at a desk. Building the API first, with the web and mobile apps as clients of it, makes that later step much cheaper.</p>
<div class="callout"><b>Not sure which to build?</b> The answer usually comes from where your users find you and what the app must do without a connection. <a class="link" href="https://flavoureak.co.uk/mobile/">See how we approach mobile apps</a>.</div>
<h2>Questions people ask</h2>
<h3>Can a PWA be listed in the App Store?</h3>
<p>Not directly. To appear in the App Store, a PWA has to be wrapped in a native app, and Apple may reject a wrapper that adds nothing beyond the website. Google Play does accept PWAs packaged as a Trusted Web Activity.</p>
<h3>Do PWAs work offline on iPhone?</h3>
<p>Yes. A service worker can cache pages and data so the app opens and works without a connection. Safari may clear stored data from sites that have not been used for some time, so a PWA on iPhone should not hold unsynced work for long periods.</p>
<h3>Is a PWA cheaper than a native app?</h3>
<p>Usually, because one web codebase serves every platform and updates need no store review. The saving shrinks if you later need native apps anyway, so decide based on what the product must do in its first year.</p>
<h3>Can a PWA use Face ID?</h3>
<p>Indirectly. A PWA can use passkeys through the WebAuthn standard, and the phone may confirm the passkey with Face ID or a fingerprint. It cannot use biometrics for other purposes, such as protecting data stored on the device, in the way a native app can.</p>]]></content:encoded>
</item>
<item>
<title>How to test an AI feature before your customers see it</title>
<link>https://flavoureak.co.uk/blog/testing-ai-features/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/testing-ai-features/</guid>
<pubDate>Tue, 12 May 2026 08:00:00 GMT</pubDate>
<category>AI</category>
<description>Build a test set from real cases, agree what a pass looks like, score every change against it and keep watching once it is live. Most of the work is the test set.</description>
<content:encoded><![CDATA[<ul><li>An evaluation set is a fixed collection of real inputs, each with the result you expect, that you run an AI feature against before every release.</li><li>Use automated checks for anything a rule can verify, such as format, required fields and numbers, and keep human review for tone, judgement and subtle errors.</li><li>Rerun the whole evaluation set whenever the prompt, the model or the retrieved data changes, because a fix in one case often breaks another.</li></ul><p>You test an AI feature by running it against a fixed set of real cases, each with an agreed definition of a correct result, and counting how many it gets right before anything reaches a customer. That fixed set is called an evaluation set, or eval set. It is a collection of example inputs, each paired with an expected output or a written rule for judging the output, and it is rerun every time the prompt or the model changes.</p>
<h2>Why AI features need a different kind of testing</h2>
<p>Ordinary software is deterministic. Give a VAT calculation the same invoice twice and it returns the same figure twice, so one test proves the rule works. A language model can return a different answer to the same input, and a change that improves one answer can quietly make another worse. A good demo on five hand-picked examples proves very little.</p>
<p>The answer is to test with volume and to test the same way every time. The useful question is how many of 150 real cases the feature gets right, and which ones it gets wrong. That number, and the list of failures behind it, becomes the thing the team protects from one release to the next.</p>
<h2>How to build an evaluation set from real cases</h2>
<p>The best test cases are the ones the feature will actually meet. Examples written by the team tend to be tidy, and real inputs are not. Emails arrive with signatures and forwarded chains. Documents arrive as phone photos taken at an angle. Customers ask two questions in one message and leave out the order number.</p>
<ol>
<li><b>Collect real inputs.</b> Pull past support tickets, documents, enquiries or records that the feature would have handled. Remove or replace any personal data the test doesn’t need.</li>
<li><b>Write down the expected result for each one.</b> For extraction, that is the correct value of every field. For drafting, it is a short list of what a good answer must include and must not include.</li>
<li><b>Cover the awkward cases on purpose.</b> Include blank fields, contradictory information, very long inputs, other languages and requests the feature should decline.</li>
<li><b>Tag each case.</b> Label it by type, such as “handwritten”, “two questions” or “out of scope”, so you can see where failures cluster.</li>
<li><b>Version it.</b> Store the set alongside the code, and add every real failure found later as a new case.</li>
</ol>
<p>Fifty well-chosen cases are enough to start. A few hundred give a clearer picture of rare failures. The set should keep growing for as long as the feature is live, because production keeps finding inputs nobody thought of.</p>
<h2>Deciding what counts as a pass</h2>
<p>Pass criteria have to be written before anyone looks at the results. Otherwise the team will talk itself into accepting whatever came out. Write them per case and per feature, and make each one something two reviewers would agree on.</p>
<p>Take a feature like the lesson content in CurioKids, which is generated for each child from their CEFR level, age and interests, and which the teacher can change. A pass for a case like that can be broken into checks. The vocabulary suits the stated level. The topic reflects at least one of the child’s interests. Nothing in it is unsuitable for a child. The activities fit the length of the lesson. Each check is met or not met, which is far easier to agree on than a score out of ten.</p>
<p>Then set a release threshold for the whole set. Say the set has 150 cases and the current version passes 138 of them, which is 92%. A new version ships only if it passes at least as many, and only if every case that has gone from pass to fail has been read and accepted by a named person.</p>
<h2>Automated checks and human review</h2>
<p>Some checks a computer can make instantly and without disagreement. Others need a person. The aim is to automate everything a rule can decide, so that human time goes on the judgements that really need it.</p>
<table><thead><tr><th>Check</th><th>Who or what does it</th><th>Example</th></tr></thead><tbody>
<tr><td>Format and structure</td><td>Code</td><td>The output is valid JSON with every required field present</td></tr>
<tr><td>Exact values</td><td>Code</td><td>The extracted invoice total matches the expected £1,240.00</td></tr>
<tr><td>Numbers within tolerance</td><td>Code</td><td>A measured quantity is within 1% of the expected figure</td></tr>
<tr><td>Must include or must avoid</td><td>Code</td><td>A reply mentions the refund window and never promises a delivery date</td></tr>
<tr><td>Meaning and coverage</td><td>A second model, spot-checked by people</td><td>A summary covers the three key points of the source document</td></tr>
<tr><td>Tone, judgement and safety</td><td>People</td><td>A reply to an upset customer is accurate, calm and appropriate</td></tr>
</tbody></table>
<p>Using a second model to grade outputs, often called model-graded evaluation, is useful for checking meaning at scale. It makes mistakes of its own, so check a sample of its grades by hand. Don’t let it be the only judge of anything that affects money, safety or a customer’s rights.</p>
<p>For human review, show reviewers the input, the output and the written criteria side by side, and ask for pass or fail against each criterion. Having two people review the same sample now and then shows quickly whether the criteria are clear enough.</p>
<h2>Regression testing when the prompt or model changes</h2>
<p>A regression is something that used to work and now doesn’t. AI features regress more easily than most software, because a single change can affect every output at once. Rerun the full evaluation set whenever any of these change:</p>
<ul>
<li>the prompt or system instructions, including small wording edits</li>
<li>the model, or the model version, including updates made by the provider</li>
<li>settings such as temperature or the maximum output length</li>
<li>the retrieval step, such as how documents are split, searched or ranked</li>
<li>the data the feature reads, such as a new price book or product catalogue</li>
<li>the code that validates or reshapes the output afterwards</li>
</ul>
<p>Compare the new run with the last accepted run, case by case. A headline pass rate hides movement, because a version can fix six cases and break five and still look better on paper. Read every case that changed from pass to fail before you ship.</p>
<p>Where your provider allows it, call a specific model version instead of an alias that moves automatically. The model then changes only when you decide, and you can run the evaluation set first.</p>
<h2>Monitoring and logging in production</h2>
<p>An evaluation set tells you how the feature behaves on the cases you thought of. Production tells you about the rest. Log enough to reconstruct what happened on any single call:</p>
<ul>
<li>the input and the output, or a reference to them where they hold personal data</li>
<li>the prompt version and the model version used</li>
<li>any documents or records retrieved and passed to the model</li>
<li>time taken, tokens used and cost</li>
<li>what the person did next, whether they accepted, edited, rejected or retried</li>
</ul>
<p>The last item is the most valuable. A draft accepted unchanged is a strong signal of quality. A draft that is rewritten every time is a failure, even if nobody reports it. The screen design that makes this signal easy to collect is covered in <a class="link" href="https://flavoureak.co.uk/blog/designing-for-wrong-answers/">designing AI features for when the answer is wrong</a>.</p>
<p>Review a sample of real calls every week and turn the interesting failures into new evaluation cases. Set alerts for anything measurable, such as a rise in validation failures, slower responses or a higher cost per call, which we break down in <a class="link" href="https://flavoureak.co.uk/blog/ai-cost-per-call/">what an AI feature costs per call</a>.</p>
<p>Logs of AI calls often contain personal data, so treat them like any other personal data under UK GDPR. Keep only what you need, restrict who can read the logs, and delete them on a set schedule. The wider rules are in <a class="link" href="https://flavoureak.co.uk/blog/ai-and-uk-gdpr/">using AI with customer data under UK GDPR</a>.</p>
<div class="callout"><b>Adding AI to your product?</b> We write the evaluation set and design the review screen alongside the prompt, before anything goes live. <a class="link" href="https://flavoureak.co.uk/ai/">See how we build AI features</a>.</div>
<h2>Questions people ask</h2>
<h3>How many test cases does an AI feature need?</h3>
<p>Start with around 50 real cases that cover both the common inputs and the awkward ones. Grow the set to a few hundred as the feature matures, adding every real failure you find. The spread of the cases matters more than the count.</p>
<h3>Can we test an AI feature without real customer data?</h3>
<p>You can start with realistic examples written by the team, but they tend to be cleaner than real inputs. Where you use real data, remove personal details the test doesn’t need, and check that your lawful basis and privacy notice cover the use.</p>
<h3>What is an AI eval?</h3>
<p>An eval is a repeatable test of an AI feature. It runs the feature against a fixed set of inputs and scores each output against agreed criteria, so the team can compare versions with numbers instead of impressions.</p>
<h3>How often should we rerun the tests?</h3>
<p>Rerun the full evaluation set before releasing any change to the prompt, model, settings, retrieval or output handling. Running it on a regular schedule as well helps catch changes on the provider’s side that nobody announced.</p>]]></content:encoded>
</item>
<item>
<title>Agency, freelancers or an in-house team: choosing who builds your product</title>
<link>https://flavoureak.co.uk/blog/agency-freelancer-or-in-house/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/agency-freelancer-or-in-house/</guid>
<pubDate>Wed, 06 May 2026 08:00:00 GMT</pubDate>
<category>Working with us</category>
<description>Agencies give you a complete team quickly, freelancers give flexibility with less overhead, and an in-house team gives long-term control. Many products use more than one over their life.</description>
<content:encoded><![CDATA[<ul><li>An agency or studio supplies a complete team, including design, development, testing and project management, under one contract, which is usually the quickest way to start but costs more per day than individual freelancers.</li><li>Freelancers usually cost less per day and are flexible, but you take on the coordination yourself, and continuity depends on particular people staying available.</li><li>An in-house team is the slowest and most expensive option to set up and builds knowledge the company keeps; a common path is to start with an agency and move work in-house as the product proves itself.</li></ul><p>The three ways to get a software product built are an agency or studio, freelancers, or your own in-house team, and each trades cost, speed, risk and control differently. An agency provides a ready-made team under one contract. Freelancers are individual contractors you coordinate yourself. An in-house team is employed by you. The right choice depends mostly on how central the software is to your business and how long it will need active development.</p>
<h2>The three options compared</h2>
<table><thead><tr><th></th><th>Agency or studio</th><th>Freelancers</th><th>In-house team</th></tr></thead><tbody><tr><td>Cost structure</td><td>Day or project rate covering the team, management and overheads</td><td>Day or hourly rate per person</td><td>Salaries plus employment costs, paid whether or not there is work</td></tr><tr><td>Speed to start</td><td>Quick, because the team already exists</td><td>Quick for one person, slower to assemble several</td><td>Slow, because hiring takes time</td></tr><tr><td>Range of skills</td><td>Design, web, mobile, testing and operations in one place</td><td>Whatever you hire</td><td>Whatever you hire and can keep busy</td></tr><tr><td>Management load</td><td>Low, the agency coordinates</td><td>High, you coordinate</td><td>High, you manage and develop people</td></tr><tr><td>Continuity</td><td>Depends on the agency and on handover</td><td>Depends on individuals</td><td>Strong, while people stay</td></tr><tr><td>Ownership of code</td><td>Must be assigned to you in the contract</td><td>Must be assigned to you in the contract</td><td>Yours, for work done in the course of employment</td></tr></tbody></table>
<h2>Cost structure</h2>
<p>Comparing day rates alone is misleading. An agency’s rate covers project management, design, testing, holidays, equipment and the cost of finding and training people. A freelancer’s rate covers their own time, and the coordination falls to you. An employee’s salary is only part of what they cost. The full cost of an in-house developer includes the items below.</p>
<ul><li>Salary.</li><li>Employer’s National Insurance contributions.</li><li>Workplace pension contributions under auto-enrolment.</li><li>Recruitment fees, or the time spent hiring.</li><li>Equipment, software licences and training.</li><li>Management time, and the cost of the team during quieter periods.</li></ul>
<p>In-house becomes cheaper per day of work when there is a steady, long-term stream of development. For a burst of work followed by quieter maintenance, an agency or freelancers usually cost less overall.</p>
<h2>Speed and risk</h2>
<p>An agency can usually start within weeks, because the team exists. Hiring an experienced developer can take months once you count advertising, interviews, notice periods and getting up to speed. A single freelancer can start quickly, while assembling and coordinating several takes time and skill.</p>
<p>The risks differ too. With an agency, you depend on one supplier, so the contract, the repository and the handover terms matter. With freelancers, the main risk is key-person dependency, where one person holds knowledge nobody else has. With an in-house team, the risks are a poor hire and gaps in skills, because a small team rarely covers design, backend, mobile, testing and operations equally well.</p>
<h2>Continuity, knowledge and ownership</h2>
<h3>Keeping the knowledge</h3>
<p>Whoever builds the product, the knowledge needs to outlive them. Insist on the code being in a repository your company owns from day one (<a class="link" href="https://flavoureak.co.uk/blog/you-get-the-repository/">you get the repository</a>), with documentation, automated tests and a deployment process someone else can run. These make it possible to change supplier, bring work in-house or rescue a project that has stalled (<a class="link" href="https://flavoureak.co.uk/blog/taking-over-a-stalled-project/">taking over a stalled project</a>).</p>
<p>Ask any supplier how they would hand the product over if you parted ways, and what you would have on the last day. A good answer is specific. It covers the repository, hosting accounts, credentials, documentation and a few days of overlap with whoever comes next.</p>
<h3>Who owns the code</h3>
<p>Under the Copyright, Designs and Patents Act 1988, work an employee creates in the course of their employment belongs to the employer, unless agreed otherwise. Work by a freelancer or an agency belongs to its author unless it is assigned, and an assignment of copyright has to be in writing and signed by the person assigning it. So contracts with agencies and freelancers need a clear assignment of intellectual property to you. <a class="link" href="https://flavoureak.co.uk/blog/who-owns-the-code-uk/">Who owns the code in the UK</a> covers the detail, The contract wording matters, so ask a solicitor to check it.</p>
<h2>Freelancers and IR35</h2>
<p>If a freelancer works through their own limited company, the off-payroll working rules, often called IR35, may apply. Where the client is a medium or large business, the client must decide whether the freelancer would be an employee for tax purposes if engaged directly, and give them a status determination statement. Small businesses don’t have that duty, and the freelancer’s company decides instead. Public sector clients always make the decision themselves. HMRC publishes guidance and its Check Employment Status for Tax tool, and an accountant can advise on your case.</p>
<h2>A hybrid path</h2>
<p>Many products use more than one option over their life. A path we’d often suggest runs in five stages.</p>
<ol><li><b>Start with an agency or studio</b> for discovery and the first release, so you get a full team without hiring.</li><li><b>Hire a technical lead</b> once the product has paying customers or a clear future, to own technical decisions and review the agency’s work.</li><li><b>Share the work.</b> The in-house lead and the agency work in the same repository, and the agency hands over areas as the internal team grows.</li><li><b>Move the core in-house.</b> Your own team runs the product day to day.</li><li><b>Use specialists for peaks</b>, such as an agency for a mobile app or an AI feature, or a freelancer for a short, well-defined piece of work.</li></ol>
<p>The handover in the third and fourth stages goes best when everyone expects it from the start. Tell the agency at the outset that you plan to build an in-house team, so the code, documentation and ways of working are set up for it.</p>
<h2>How to decide</h2>
<p>Answer these questions honestly before you choose.</p>
<ul><li>Is the software the core of your business, or a tool that supports it?</li><li>Will it need active development for years, or a build followed by maintenance?</li><li>Do you have someone who can manage developers and judge their work?</li><li>How soon do you need something live?</li><li>Would losing one person stop the project?</li></ul>
<p>If the software is your product and you can wait for hiring, plan to build an in-house team, with outside help to get started. If it supports your business, an agency or studio on a retainer is usually simpler. If you already have a strong technical lead, freelancers can fill specific gaps well.</p>
<div class="callout"><b>Weighing up who should build your product?</b> <a href="https://flavoureak.co.uk/businesses/">See how we work with businesses</a> on building and running their software.</div>
<h2>Questions people ask</h2>
<h3>Is an agency more expensive than hiring developers?</h3><p>Per day, an agency usually costs more than an employee’s salary, because its rate covers management, testing, design and overheads. Over a whole project the answer depends on how long you need the work. Employees cost money every month whether or not there is work, while an agency can be scaled down when the build is done.</p>
<h3>Who owns the code a freelancer writes?</h3><p>Under UK copyright law, a freelancer owns the copyright in the code they write unless a written, signed agreement assigns it to you. Make sure the contract includes an assignment of intellectual property, and that the code sits in a repository your company controls.</p>
<h3>Can an agency hand a product over to an in-house team?</h3><p>Yes, if the handover is planned. It goes best when the code has been in your repository from the start, the documentation and deployment process are up to date, and your new developers work alongside the agency for a period before taking over.</p>
<h3>Do I need a technical co-founder or CTO before hiring an agency?</h3><p>It isn’t required. You do need someone on your side who can make product decisions quickly and check that the work is going well. Some founders use an independent technical adviser for a few days a month to review code, estimates and plans, then hire a permanent technical lead as the product grows.</p>]]></content:encoded>
</item>
<item>
<title>Roles and permissions in SaaS: RBAC, ABAC and per-record sharing</title>
<link>https://flavoureak.co.uk/blog/saas-roles-permissions/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/saas-roles-permissions/</guid>
<pubDate>Fri, 01 May 2026 08:00:00 GMT</pubDate>
<category>SaaS</category>
<description>Start with a few fixed roles built from named permissions, check them on the server, and scope every check to the tenant. Add attribute rules, custom roles and sharing as customers need them.</description>
<content:encoded><![CDATA[<ul><li>Role-based access control (RBAC) gives each user a role, such as Admin or Member, and each role a set of permissions; it is the right starting point for most B2B SaaS products.</li><li>Attribute-based access control (ABAC) decides access from facts about the user, the record and the context, such as a site manager editing only jobs on their own sites, while per-record sharing grants access to one item at a time.</li><li>Check named permissions in code instead of role names, enforce them on the server for every request, and store a user’s role on their membership of a tenant, because these choices are expensive to change later.</li></ul><p>Roles and permissions decide who can see and do what inside each customer’s account. A permission is a single allowed action, such as approving an invoice, and a role is a named bundle of permissions given to a user. Most B2B SaaS products should start with role-based access control and a handful of default roles, then add attribute rules, custom roles and per-record sharing as customers ask for them.</p>
<h2>Four models of access control</h2>
<p>The terms are used loosely, so the table sets them side by side with an example of each.</p>
<table><thead><tr><th>Model</th><th>How access is decided</th><th>Example</th><th>Good for</th></tr></thead><tbody><tr><td>Role-based (RBAC)</td><td>The user’s role grants a fixed list of permissions</td><td>Members can create quotes, and only Admins can invite users</td><td>Almost every product, from day one</td></tr><tr><td>Attribute-based (ABAC)</td><td>Rules combine facts about the user, the record and the context</td><td>Managers can approve invoices under £5,000 for their own region</td><td>Limits by site, region, amount or status</td></tr><tr><td>Per-record sharing</td><td>Access to one record is granted to one person or group</td><td>A client can view their own project and nothing else</td><td>Guests, clients and subcontractors</td></tr><tr><td>Relationship-based (ReBAC)</td><td>Access follows links between records, such as team, owner or parent folder</td><td>Anyone on a project team can see that project’s documents</td><td>Products with nested workspaces and folders</td></tr></tbody></table>
<p>Mature products usually combine them. Roles set the broad shape of access, attribute rules narrow it, and sharing handles the exceptions.</p>
<h2>How role-based access control works</h2>
<p>Write permissions as named actions on a resource, such as <code>quotes.create</code>, <code>invoices.approve</code>, <code>users.invite</code> and <code>reports.export</code>. Roles are lists of these permissions. Before every action, the code checks whether this user, in this tenant, has the permission the action needs.</p>
<p>The important rule is to check permissions, never role names. Code that asks whether the user is an Admin has to be found and changed every time roles change. Code that checks <code>invoices.approve</code> keeps working when you add a Finance role or let customers build their own roles.</p>
<h2>Default roles to start with</h2>
<p>Four or five default roles cover most early customers.</p>
<ul><li><b>Owner.</b> Billing, plan changes, deleting the account and transferring ownership. Every account always has at least one.</li><li><b>Admin.</b> Users, roles, settings and integrations, without billing.</li><li><b>Member.</b> Day-to-day work, such as creating and editing records.</li><li><b>Viewer.</b> Read-only access, often for directors or auditors.</li><li><b>Guest.</b> Access only to specific records shared with them, such as a client viewing their own job.</li></ul>
<p>Products used in the field often need a role shaped around one kind of work. Estimark’s menu includes a field worker mode as well as office screens such as Invoices and Reports. People on site and people who handle money need different things in front of them, and roles are how a product decides who sees which.</p>
<h2>When you need attribute rules or per-record sharing</h2>
<p>Roles alone can’t express conditions. Requests such as “site managers can only edit jobs on their own sites”, “a submitted timesheet can’t be edited” and “managers can approve spend up to £5,000” all depend on facts about the record or the person. That is attribute-based access control. Implement it as a small set of policy functions that take the user, the record and the action, so the rules live in one place and can be tested.</p>
<p>Per-record sharing is for exceptions to the role. A subcontractor needs to see one job’s drawings. A client needs to approve one quote. Store each share as a row naming the record, the person or group and the access level, and include shares in every query that lists records.</p>
<h2>Custom roles</h2>
<p>Larger customers eventually ask to define their own roles. If permissions are already named and checked in code, custom roles are mostly an admin screen. Let admins copy a default role and adjust it, describe each permission in plain language, and stop anyone removing the last Owner or their own ability to manage roles. Customers using <a class="link" href="https://flavoureak.co.uk/blog/sso-for-saas/">single sign-on</a> will also want groups from their identity provider mapped to these roles.</p>
<p>Decide what happens when you release a new permission. Default roles get it as you intend. Custom roles should not get it automatically, because a customer who built a careful read-only finance role does not expect it to gain new powers after an update.</p>
<h2>How to model roles and permissions in the database</h2>
<p>A simple, durable model uses the tables below. The key choice is that the role sits on the membership, the link between a user and a tenant. One person can then be an Admin in their own company’s account and a Guest in a client’s.</p>
<table><thead><tr><th>Table</th><th>Holds</th><th>Notes</th></tr></thead><tbody><tr><td>tenants</td><td>One row per customer account</td><td>Every other business table carries a tenant ID</td></tr><tr><td>users</td><td>People, identified by email or SSO identity</td><td>Global, because one person may join several tenants</td></tr><tr><td>memberships</td><td>User ID, tenant ID, role ID and status</td><td>The role lives here</td></tr><tr><td>roles</td><td>Name, description and tenant ID</td><td>Tenant ID is empty for default roles shared by every tenant</td></tr><tr><td>role_permissions</td><td>Role ID and a permission key</td><td>Permission keys are defined in code</td></tr><tr><td>record_shares</td><td>Record type and ID, grantee, access level and expiry</td><td>For guests and exceptions</td></tr></tbody></table>
<p>Keep the list of permission keys in code, under version control, and the assignment of permissions to roles in the database. Cache a user’s permissions for the session, and clear the cache when their role changes.</p>
<h2>Mistakes that are expensive to fix later</h2>
<ol><li><b>Putting the role on the user.</b> It works until the first person needs access to two tenants, and then many queries have to change.</li><li><b>Checking role names in code.</b> Custom roles become impossible without a rewrite.</li><li><b>Enforcing permissions only in the interface.</b> Hiding a button is a convenience. The server must check every request, including those to the API.</li><li><b>Forgetting lists, search, exports and reports.</b> A user blocked from opening a record can often still see it in a CSV export or a dashboard total.</li><li><b>Missing tenant scoping.</b> Every permission check must also confirm the record belongs to the user’s tenant (<a class="link" href="https://flavoureak.co.uk/blog/multi-tenancy-explained/">multi-tenancy explained</a>).</li><li><b>Keeping no record of changes.</b> Role changes, new shares and support staff access should all be written to an <a class="link" href="https://flavoureak.co.uk/blog/saas-audit-logs/">audit log</a>.</li></ol>
<p>Each of these is cheap to get right in the first version and slow to change once hundreds of screens, queries and reports depend on it.</p>
<div class="callout"><b>Designing a multi-user product?</b> <a href="https://flavoureak.co.uk/saas/">See how we build SaaS</a> with roles, tenancy and audit trails planned before the first screen.</div>
<h2>Questions people ask</h2>
<h3>What is the difference between a role and a permission?</h3><p>A permission is one specific allowed action, such as creating a quote or exporting a report. A role is a named group of permissions, such as Admin or Viewer, that you give to a user. Users get their permissions through their role, which keeps access manageable as a team grows.</p>
<h3>How many roles should a new SaaS have?</h3><p>Start with four or five, usually Owner, Admin, Member, Viewer and a guest role for people outside the customer’s company. Fewer roles are easier to explain and to test. Add roles when several customers ask for the same shape of access, and offer custom roles when larger customers need their own.</p>
<h3>Do I need a policy engine for permissions?</h3><p>Most early products don’t. Permission checks in application code, kept in one module and covered by tests, are enough for roles and simple attribute rules. Policy engines and authorisation services such as Open Policy Agent, Cedar or OpenFGA help when rules are complex, shared across several services or need to change without a release.</p>
<h3>Should permissions be checked in the front end?</h3><p>The front end should use permissions to hide or disable actions a user can’t take, which keeps the interface clear. It must never be the only check. Every request to the server has to be checked again, because anyone can call an API directly.</p>]]></content:encoded>
</item>
<item>
<title>Shopify or a custom e-commerce build: how to decide what fits</title>
<link>https://flavoureak.co.uk/blog/shopify-or-custom/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/shopify-or-custom/</guid>
<pubDate>Sat, 25 Apr 2026 08:00:00 GMT</pubDate>
<category>E-commerce</category>
<description>Shopify handles a standard online shop very well. Businesses outgrow it with complex B2B pricing, unusual catalogues or checkout rules, and custom apps on Shopify are often the middle path.</description>
<content:encoded><![CDATA[<ul><li>Shopify is usually the right choice for a shop selling a standard catalogue to consumers, because hosting, checkout, payments and card security come built in.</li><li>Businesses tend to outgrow an off-the-shelf set-up when they need customer-specific B2B pricing, complex product rules, custom checkout logic or deep integration with their own systems.</li><li>Custom apps or a headless storefront on top of Shopify can meet those needs without rebuilding the parts Shopify already does well.</li></ul><p>Choosing between Shopify and a custom e-commerce build comes down to how standard your selling is. Shopify is a hosted commerce platform that provides the shop, checkout, payments, hosting and security as a subscription service, and for most consumer shops it is the sensible default. A custom build makes sense when your pricing, catalogue, checkout or integrations are unusual enough that you would spend more effort working around a platform than building your own, and custom apps or storefronts built on Shopify sit between the two.</p>
<h2>What Shopify does well</h2>
<ul>
<li><b>Checkout and payments.</b> A hosted, well-tested checkout, with Shopify Payments and a range of other payment providers. Shopify is certified as a PCI DSS Level 1 service provider, so most of the work of handling card data safely is done for you.</li>
<li><b>Hosting and security.</b> There are no servers to run, patch or scale up for a big sale.</li>
<li><b>Themes and the editor.</b> Merchandisers can change pages, menus and collections without a developer.</li>
<li><b>Apps.</b> A large app store covers reviews, subscriptions, loyalty, shipping and much more.</li>
<li><b>Channels.</b> Point of sale, social and marketplace channels run from one product catalogue.</li>
<li><b>International selling.</b> Multiple currencies, languages and domains through Shopify Markets.</li>
</ul>
<p>For a business selling a few hundred or a few thousand products to consumers at one price each, that combination is hard to beat on cost and time to launch. Building the same from scratch would mean paying to recreate features that already work.</p>
<h2>Where businesses outgrow off-the-shelf</h2>
<p>Off-the-shelf here means using the platform as designed, plus apps from its store. These are the needs that commonly push beyond it.</p>
<h3>B2B pricing and ordering</h3>
<p>Trade customers expect their own price lists, contract prices, volume breaks, payment on account, credit limits, order approval and quick reordering. Shopify has added B2B features, such as company accounts, catalogues, price lists and payment terms, so check what your plan includes before assuming you need a custom build. Where prices come from an ERP and change per customer, or ordering needs quotes and approval steps, a dedicated portal is often cleaner. See <a class="link" href="https://flavoureak.co.uk/blog/b2b-ordering-portals/">B2B ordering portals</a>.</p>
<h3>Complex catalogues</h3>
<p>Made-to-measure products, configurable products with rules between options, bundles built from parts and products priced by size all strain a model built on fixed variants. Shopify’s limits on product options and variants have changed over time, so check the current limits against your most complicated product before deciding.</p>
<h3>Custom checkout logic</h3>
<p>Shopify’s checkout can be extended with checkout UI extensions and Shopify Functions, which let you add fields and content, validate carts and customise discounts, delivery and payment options. Some of these capabilities depend on your plan. The core flow of the checkout stays Shopify’s. If you need, for example, a checkout that books an installation slot in your own engineers’ calendar and takes a deposit against it, that is custom work on any platform.</p>
<h3>Integrations</h3>
<p>Shopify connects to ERPs, warehouses and accounting systems through its Admin API and webhooks, and simple integrations come ready-made as apps. When stock, prices and orders must sync both ways with an ERP in near real time, you need integration code that you own and monitor, whichever platform you choose.</p>
<h2>The middle path of custom apps on Shopify</h2>
<p>A custom app is software built for one store, using Shopify’s APIs, that adds a feature the platform lacks without leaving the platform. It keeps Shopify doing what it does well and puts your own code only where your business is different. Typical examples:</p>
<ol>
<li>An app that syncs customer-specific prices from your ERP into Shopify’s B2B price lists.</li>
<li>A product configurator built as a theme app extension, writing the chosen options to the order.</li>
<li>Shopify Functions that apply your own discount or delivery rules at checkout.</li>
<li>A headless storefront using the Storefront API, with Shopify still running checkout and orders. See <a class="link" href="https://flavoureak.co.uk/blog/headless-commerce-explained/">headless commerce explained</a>.</li>
<li>An internal tool that reads orders through the Admin API to plan production or deliveries.</li>
</ol>
<p>Agree who owns the code for anything built this way. The app should live in your own repository, even though it runs against Shopify. Our article on <a class="link" href="https://flavoureak.co.uk/blog/who-owns-the-code-uk/">who owns the code in the UK</a> covers the contract terms to look for.</p>
<h2>What a fully custom build involves</h2>
<p>A fully custom build means you own the catalogue, cart, checkout, order management and admin screens, usually built on a web framework or an open-source commerce engine. You also take on the responsibilities Shopify carried for you:</p>
<ul>
<li>card security, where using a payment provider’s hosted payment fields or hosted checkout keeps card data off your servers and reduces your PCI DSS scope, though some obligations remain;</li>
<li>hosting, scaling for peak sales, backups and security patches;</li>
<li>admin screens for staff, which take longer to build than people expect;</li>
<li>maintenance, which is a permanent cost and grows with every feature you add.</li>
</ul>
<p>In return, the buying process can work exactly as your business does, and nothing about it depends on a platform’s roadmap or plan structure.</p>
<h2>Shopify, custom apps and a custom build compared</h2>
<table><thead><tr><th></th><th>Shopify with theme and apps</th><th>Shopify with custom apps or headless</th><th>Fully custom build</th></tr></thead><tbody>
<tr><td>Time to launch</td><td>Shortest</td><td>Medium</td><td>Longest</td></tr>
<tr><td>Upfront cost</td><td>Lowest</td><td>Medium</td><td>Highest</td></tr>
<tr><td>Running costs</td><td>Subscription, app fees and payment fees</td><td>Platform costs, plus hosting and maintenance of your code</td><td>Hosting, payment fees, maintenance and support</td></tr>
<tr><td>Checkout control</td><td>Platform checkout with extensions</td><td>Platform checkout with extensions</td><td>Full</td></tr>
<tr><td>B2B and complex pricing</td><td>Within platform features and apps</td><td>Your logic, synced into the platform</td><td>Any model you can specify</td></tr>
<tr><td>Card security burden</td><td>Mostly carried by the platform</td><td>Mostly carried by the platform</td><td>Reduced by hosted payment fields, still yours to manage</td></tr>
<tr><td>Code you own</td><td>Theme customisations</td><td>Your apps and storefront</td><td>Everything</td></tr>
<tr><td>Best fit</td><td>Standard consumer shops</td><td>Shops with one or two unusual needs</td><td>Businesses where the buying process is the product</td></tr>
</tbody></table>
<h2>How to decide</h2>
<p>A short exercise settles most cases.</p>
<ol>
<li>Write down the five things your buying process must do that a basic online shop does not.</li>
<li>For each one, check whether Shopify does it natively on the plan you would use.</li>
<li>If not, check whether an established, well-reviewed app does it.</li>
<li>Mark anything left as needing custom code, with a rough size.</li>
<li>If nothing needs code, use Shopify. If one or two items do, build custom apps on Shopify. If most do, or the checkout itself is unusual, a custom build is likely to be the better long-term choice.</li>
</ol>
<p>The first choice is not permanent. Shopify lets you export products, customers and orders, so a business can start on the platform, learn what its customers actually need and move later if the list of custom work keeps growing. Plan URLs and redirects carefully if you do move, so the shop keeps its search traffic.</p>
<div class="callout"><b>Not sure which route fits your shop?</b> <a href="https://flavoureak.co.uk/ecommerce/">See how we scope e-commerce projects</a>.</div>
<h2>Questions people ask</h2>
<h3>Is Shopify good for B2B?</h3>
<p>Shopify has B2B features such as company accounts, catalogues, price lists and payment terms, so check what your plan includes. It suits wholesale with straightforward price lists. Where prices come from an ERP per customer, or ordering needs quotes and approvals, a custom app or dedicated portal is often a better fit.</p>
<h3>Can you customise the Shopify checkout?</h3>
<p>Yes, within limits. Checkout UI extensions and Shopify Functions let you add fields and content, validate carts and customise discounts, delivery and payment options, and some of these depend on your plan. You cannot replace the checkout’s core flow.</p>
<h3>Is a custom e-commerce site more secure than Shopify?</h3>
<p>Not by default. Shopify is certified as a PCI DSS Level 1 service provider and runs its own hosting and security. A custom site can be secure, but your team carries that responsibility, including patches, monitoring and card data handling.</p>
<h3>Who owns the code if an agency builds a Shopify app for us?</h3>
<p>That depends on the contract. Under UK law, code written by an agency’s employees belongs to the agency unless the contract assigns it to you, so ask for a written assignment of intellectual property and access to the repository. The Shopify platform itself always remains Shopify’s.</p>
<h3>When should a business move off Shopify?</h3>
<p>Consider moving when most of what makes your buying process different needs custom code, or when the checkout itself has to work in a way Shopify does not allow. If only one or two needs fall outside the platform, custom apps on Shopify usually cost less than leaving it.</p>]]></content:encoded>
</item>
<item>
<title>The Children’s code for edtech: what the ICO expects from learning apps</title>
<link>https://flavoureak.co.uk/blog/childrens-code-for-edtech/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/childrens-code-for-edtech/</guid>
<pubDate>Sun, 19 Apr 2026 08:00:00 GMT</pubDate>
<category>Education</category>
<description>The ICO’s Age Appropriate Design Code sets 15 standards for online services likely to be used by children in the UK. Here is what each one means for an edtech product.</description>
<content:encoded><![CDATA[<ul><li>The Age Appropriate Design Code, or Children’s code, is a statutory code of practice from the ICO for online services likely to be accessed by children under 18 in the UK.</li><li>It sets 15 standards, including high privacy by default, geolocation off by default, no nudge techniques that weaken privacy, and a data protection impact assessment.</li><li>Whether it applies to an edtech product sold to schools depends on who decides how pupils’ data is used, so take legal advice on your position.</li></ul><p>The Age Appropriate Design Code, usually called the Children’s code, is a statutory code of practice from the Information Commissioner’s Office (ICO) that sets 15 standards for online services likely to be accessed by children in the UK. For an edtech product it means high privacy settings by default, collecting only what each feature needs, keeping location off and avoiding design that nudges children to share more. It is a practical guide and does not replace legal advice.</p>
<h2>What the Children’s code is</h2>
<p>The code was made under section 123 of the Data Protection Act 2018. It came into force in September 2020 with a 12-month transition period, so services have been expected to conform since September 2021. It is not a separate law. It explains how UK GDPR and the Data Protection Act 2018 apply to children’s data in online services, and the ICO and the courts must take it into account. A service that ignores it will find it hard to show it meets UK GDPR, and fines under UK GDPR can reach £17.5 million or 4% of annual worldwide turnover, whichever is higher.</p>
<h2>Who the code applies to</h2>
<p>The code applies to information society services likely to be accessed by children. An information society service is, broadly, an online service normally provided for payment, including one funded by advertising, that a user requests individually, such as an app, website, online game or connected device. “Children” means anyone under 18. “Likely to be accessed” means it is more probable than not that children will use the service, whether or not they are the intended audience.</p>
<p>For edtech, the position depends on how the product reaches children.</p>
<ul>
<li><b>Sold to families or directly to learners</b>, such as a home tutoring app or a language app for children: in scope.</li>
<li><b>Sold to schools, where the school decides how pupils’ data is used</b> and you act only as the school’s processor: the code may not apply to you directly, though the school still has to meet UK GDPR and will expect your help.</li>
<li><b>Sold to schools, but you also use pupil data for your own purposes</b>, such as product development, analytics or marketing: for that processing you are a controller, and the code is likely to apply.</li>
</ul>
<p>Where the line falls in your contracts and data flows is a legal question. Take legal advice, and write down your reasoning so you can show it later. The Online Safety Act 2023, enforced by Ofcom, is a separate regime that may also apply if children can share content with each other in your app.</p>
<h2>The 15 standards in plain English</h2>
<table><thead><tr><th>Standard</th><th>What it asks</th><th>What it means in a learning app</th></tr></thead><tbody>
<tr><td>1. Best interests of the child</td><td>Make the child’s best interests a primary consideration in design</td><td>Weigh the learning benefit of each feature against its use of data</td></tr>
<tr><td>2. Data protection impact assessments</td><td>Assess and reduce the risks to children before processing their data</td><td>A DPIA before launch, updated for each feature that uses children’s data</td></tr>
<tr><td>3. Age appropriate application</td><td>Establish users’ ages with certainty that matches the risk, or apply the code to everyone</td><td>Use age bands from sign-up, or treat every user as a child</td></tr>
<tr><td>4. Transparency</td><td>Privacy information that is concise, prominent and in language children understand</td><td>Short explanations at the moment a feature asks for data, written for the age group</td></tr>
<tr><td>5. Detrimental use of data</td><td>Do not use children’s data in ways shown to harm their wellbeing</td><td>Take care with streaks, rewards and notifications designed to keep children in the app</td></tr>
<tr><td>6. Policies and community standards</td><td>Uphold your own published terms, policies and standards</td><td>If you say chat is moderated, moderate it</td></tr>
<tr><td>7. Default settings</td><td>High privacy by default, unless there is a compelling reason</td><td>Profiles private, sharing off, analytics minimal</td></tr>
<tr><td>8. Data minimisation</td><td>Collect and keep only the minimum data for the parts of the service a child actively uses</td><td>Separate choices for optional features, with no data collected for features left off</td></tr>
<tr><td>9. Data sharing</td><td>Do not disclose children’s data without a compelling reason</td><td>No sharing with advertisers or data brokers, and a list of every processor</td></tr>
<tr><td>10. Geolocation</td><td>Location off by default, with an obvious sign when it is on</td><td>Most learning apps do not need location at all</td></tr>
<tr><td>11. Parental controls</td><td>Explain them to the child, and show an obvious sign when a parent is monitoring</td><td>If parents can see activity, the child can see that they can</td></tr>
<tr><td>12. Profiling</td><td>Profiling off by default, and only with measures to protect the child</td><td>Adaptive learning can count as profiling, so record why it is needed in the DPIA</td></tr>
<tr><td>13. Nudge techniques</td><td>No nudges that lead children to share unnecessary data or weaken their privacy</td><td>Equal weight for “yes” and “no” on every privacy choice</td></tr>
<tr><td>14. Connected toys and devices</td><td>Connected toys and devices must include effective tools to meet the code</td><td>Applies to learning hardware such as tablets or smart speakers for children</td></tr>
<tr><td>15. Online tools</td><td>Prominent, accessible tools for children to use their data rights and report concerns</td><td>A visible “ask about my data” or “report a problem” route in the app</td></tr>
</tbody></table>
<h2>High privacy by default</h2>
<p>Default settings matter because many users, and children in particular, keep whatever a service starts with. Under the code, every setting that affects privacy should begin at its most protective option, and the child or parent can relax it when there is a reason. In a learning app that typically means:</p>
<ul>
<li>profiles visible only to the child, their parent and their teacher;</li>
<li>leaderboards that use nicknames, or are off;</li>
<li>chat with other learners off, or limited to a moderated class group;</li>
<li>analytics that collect only what is needed to run and improve the service;</li>
<li>geolocation off, and any setting that makes a child’s location visible to others switching back to off at the end of each session.</li>
</ul>
<h2>Nudge techniques and engagement features</h2>
<p>A nudge technique is a design feature that steers a choice, such as a large bright “Accept” button next to a small grey “Not now” link. The code rules out nudges that lead children to share unnecessary data or to weaken or turn off their privacy protections. Nudges towards the more private option are acceptable.</p>
<p>Edtech has its own temptations here, such as streaks, badges and notifications that reward time spent in the app. They can support learning, but the detrimental use standard asks you to avoid features that keep children engaged beyond what is good for them. For each mechanic, ask whether it rewards learning or simply time in the app, and record the answer in your DPIA.</p>
<h2>Parental controls in practice</h2>
<p>Parents have a legitimate role, and parental controls are common in edtech. The code accepts that, but asks you to explain controls to the child in language suited to their age, and to show the child an obvious sign whenever a parent is monitoring their activity or location.</p>
<p>In CurioKids, the parent’s app is where lessons are booked, joined and followed, and homework happens in a child mode opened with the parent’s PIN. That split, with the parent as account holder and a separate, limited space for the child, is a useful pattern for products used by young children. We cover the wider product in <a class="link" href="https://flavoureak.co.uk/blog/building-a-tutoring-platform/">building a tutoring platform</a>, and how lessons are personalised in <a class="link" href="https://flavoureak.co.uk/blog/lesson-for-the-child/">a lesson for the child</a>.</p>
<h2>Doing the DPIA</h2>
<p>A data protection impact assessment, or DPIA, is a written assessment of how a service uses personal data, what could harm the people involved and how you will reduce those risks. The code expects one for any service in scope, done before launch and kept up to date. The steps:</p>
<ol>
<li>Map every piece of data you collect, where it goes and why.</li>
<li>Decide how you will establish users’ ages, and record why that level of certainty is enough for the risks.</li>
<li>For each feature, assess the risks to children of different ages and stages of development, including physical, emotional and developmental harm as well as data risks.</li>
<li>Seek the views of children and parents where you can, and of your data protection officer if you have one.</li>
<li>Record the measures you take for each standard, and the reason for any departure from a high privacy default.</li>
<li>Review the DPIA with each significant release.</li>
</ol>
<div class="callout"><b>Designing a product for children?</b> We plan the Children’s code standards into the first design. <a href="https://flavoureak.co.uk/education/">See our education work</a>.</div>
<h2>Questions people ask</h2>
<h3>Does the Children’s code apply to my edtech app?</h3>
<p>It applies if your app is an online service likely to be accessed by under-18s in the UK. Apps sold to families or directly to learners are in scope. If you sell only to schools and act purely as the school’s processor, the code may not apply to you directly, so take legal advice on your position.</p>
<h3>Is the Children’s code a law?</h3>
<p>It is a statutory code of practice made under the Data Protection Act 2018, and it does not create new legal duties of its own. The ICO and the courts must take it into account when deciding whether a service meets UK GDPR. Ignoring it makes compliance very hard to show.</p>
<h3>Does an edtech app need a DPIA?</h3>
<p>Yes, if it is in scope of the Children’s code, which expects a DPIA before launch. Offering online services directly to children is also on the ICO’s list of processing likely to need a DPIA under UK GDPR. Update it whenever a new feature uses children’s data.</p>]]></content:encoded>
</item>
<item>
<title>Event registration forms and UK GDPR: what you can collect and why</title>
<link>https://flavoureak.co.uk/blog/event-registration-uk-gdpr/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/event-registration-uk-gdpr/</guid>
<pubDate>Mon, 13 Apr 2026 08:00:00 GMT</pubDate>
<category>Events</category>
<description>An event registration form should collect only what you need to run the event and explain why. Dietary and access needs, marketing consent and sponsor sharing need extra care.</description>
<content:encoded><![CDATA[<ul><li>An event registration form should collect only the personal data needed to run the event, and tell people at the point of collection why and how it will be used.</li><li>Dietary and access requirements can reveal health or religious belief, which are special category data under UK GDPR and need an extra condition, usually explicit consent.</li><li>Marketing emails to individuals need consent under PECR unless the soft opt-in applies, and consent to share details with sponsors must name the sponsors.</li></ul><p>Under UK GDPR, an event registration form should collect only the personal data you need to run the event, explain why at the point of collection, and rely on a clear lawful basis for each use. UK GDPR is the UK’s version of the General Data Protection Regulation, which works alongside the Data Protection Act 2018 to govern how organisations use personal data. Dietary and access needs, marketing consent and sharing attendee lists with sponsors need the most care.</p>
<h2>Collect what you need and say why</h2>
<p>Data minimisation, one of the principles of UK GDPR, requires personal data to be adequate, relevant and limited to what is necessary for the purpose. For each field on a registration form, ask what you will actually do with it. If the answer is nothing, remove the field.</p>
<table>
<thead><tr><th>Field</th><th>Typical purpose</th><th>Likely lawful basis</th></tr></thead>
<tbody>
<tr><td>Name and email</td><td>Send the ticket and event updates</td><td>Contract, or legitimate interests for free events</td></tr>
<tr><td>Job title and organisation</td><td>Badges and planning a business event</td><td>Legitimate interests</td></tr>
<tr><td>Dietary requirements</td><td>Catering</td><td>Contract or legitimate interests, plus an Article 9 condition if it reveals health or religion</td></tr>
<tr><td>Access requirements</td><td>Making the venue and sessions accessible</td><td>As above, because health data needs an Article 9 condition</td></tr>
<tr><td>Marketing preferences</td><td>Emails about future events</td><td>Consent, or the soft opt-in where it applies</td></tr>
<tr><td>Sponsor sharing</td><td>Passing details to named sponsors</td><td>Consent</td></tr>
</tbody>
</table>
<p>Tell people at the point of collection who you are, what you will use the data for, the lawful basis, who you will share it with, how long you will keep it and what their rights are. A short notice beside the form with a link to the full privacy notice is the usual approach.</p>
<h2>Choosing a lawful basis</h2>
<p>Every use of personal data needs one of the six lawful bases in Article 6 of UK GDPR. Three come up most often for events.</p>
<ul>
<li><b>Contract.</b> Processing that is necessary to provide what someone has signed up for, such as issuing a ticket, admitting them and telling them about changes.</li>
<li><b>Legitimate interests.</b> Processing that is necessary for a genuine interest of yours, balanced against the attendee’s interests and reasonable expectations. Badges, attendance records and sensible follow-up after the event often fit here. Record the balancing test.</li>
<li><b>Consent.</b> A freely given, specific, informed and unambiguous indication, given by a clear positive action and as easy to withdraw as to give. Pre-ticked boxes are not consent.</li>
</ul>
<p>Pick the basis before you collect the data. Consent is a poor fit for things you must do anyway, such as sending joining instructions, because people expect to be able to withdraw it.</p>
<h2>Dietary and access needs are often special category data</h2>
<p>Special category data is personal data that needs extra protection under Article 9 of UK GDPR, including data about health, religious or philosophical beliefs, racial or ethnic origin and sexual orientation. Event forms collect it more often than organisers realise.</p>
<ul>
<li>An allergy, coeliac disease or a need for step-free access is health data.</li>
<li>A request for halal or kosher food may reveal religious belief.</li>
<li>A vegetarian meal choice on its own usually reveals neither, though it can in some contexts.</li>
</ul>
<p>To process special category data you need a lawful basis under Article 6 and a separate condition under Article 9. For events, explicit consent is the condition most organisers rely on. Explicit consent means a clear statement that the person confirms, such as a tick box beside the words “I agree to my dietary and access requirements being shared with the caterer and venue to arrange this event.”</p>
<p>Make these fields optional, ask only for what you will act on, share them only with the catering or venue staff who need them, and delete them soon after the event. A short list of options plus “other, we will contact you” collects less than a free-text box.</p>
<h2>Marketing consent under PECR</h2>
<p>The Privacy and Electronic Communications Regulations 2003, known as PECR, sit alongside UK GDPR and govern electronic marketing. Sending marketing emails or texts to individuals generally needs their prior consent. The exception is the soft opt-in. It can apply when you obtained the person’s details in the course of a sale or negotiations for a sale, the marketing is about your own similar products or services, and you gave them a simple way to refuse when you collected the details and in every message since.</p>
<p>For paid events, the soft opt-in may cover emails about your future similar events, provided the opt-out was offered on the form. For free events it is less clear-cut, so many organisers ask for consent. Under PECR, emails to people at corporate addresses are treated differently, although sole traders and some partnerships count as individuals and UK GDPR still applies. The Data (Use and Access) Act 2025 amends parts of UK GDPR and PECR, with changes taking effect in stages, so check the ICO’s current guidance before relying on any of these rules.</p>
<ul>
<li>Use an unticked box for marketing consent, separate from the terms and conditions.</li>
<li>Keep a record of when and how each person consented and what they were told.</li>
<li>Include an unsubscribe link in every message and act on it promptly.</li>
</ul>
<h2>Sharing attendee lists with sponsors</h2>
<p>Passing the delegate list to a sponsor discloses personal data to another organisation, which becomes a controller in its own right. Attendees must be told, and where the sponsor will use the data for its own marketing you will normally need their consent.</p>
<p>The ICO’s guidance on consent says that you must name any third-party controllers who will rely on the consent. Wording such as “our carefully selected partners” is not enough. List the sponsors by name on the form, give a separate tick box, and pass on only the details of people who ticked it.</p>
<p>A common alternative is badge scanning at the event, where a sponsor scans the badge of anyone who visits their stand and agrees to be contacted. Tell attendees in advance how this works and what a scan shares.</p>
<h2>Photography and filming</h2>
<p>Photos and video in which people can be identified are personal data. Organisers usually rely on legitimate interests for general event photography, which means telling people in advance and giving them a practical way to object. Put a photography notice in the registration confirmation and at the venue entrance, explain how images will be used, and offer a simple opt-out such as a different coloured lanyard. Close-up photos of individuals for promotional use, and any photography involving children, need more care, and consent is often the right basis.</p>
<h2>Retention, suppliers and security</h2>
<p>The storage limitation principle in UK GDPR means personal data must not be kept for longer than the purpose needs. There is no fixed period for event data, so set your own, write it down and apply it.</p>
<table>
<thead><tr><th>Data</th><th>Example retention</th><th>Reason</th></tr></thead>
<tbody>
<tr><td>Dietary and access requirements</td><td>Delete within 30 days of the event</td><td>Only needed to run the event</td></tr>
<tr><td>Registration details</td><td>Until the next edition of the event, then review</td><td>Reasonable follow-up and repeat events</td></tr>
<tr><td>Payment and invoice records</td><td>As required for tax, generally six years for a company</td><td>HMRC record-keeping</td></tr>
<tr><td>Marketing consent records</td><td>For as long as you rely on the consent, and a period after</td><td>Proof that consent was given</td></tr>
</tbody>
</table>
<p>The periods here are examples. Choose ones that fit your events, and make sure your event platform can actually delete or anonymise data when each period ends.</p>
<p>Your ticketing platform, email service and event app process attendees’ data on your behalf, which makes them processors. You need a written contract with each one that meets Article 28 of UK GDPR, and you need to know where they store the data. Limit access to the attendee list to the people who need it. Check-in devices hold a copy of the list too, as covered in <a class="link" href="https://flavoureak.co.uk/blog/offline-event-check-in/">offline event check-in</a>, so encrypt it and wipe it afterwards. For the rest of the event stack, see <a class="link" href="https://flavoureak.co.uk/blog/what-an-event-app-needs/">what an event app needs</a>.</p>
<p>Attendees can ask for a copy of their data, and you must respond to a subject access request within one month. A personal data breach that is likely to pose a risk to people must be reported to the ICO without undue delay and, where feasible, within 72 hours of becoming aware of it.</p>
<div class="callout"><b>Building registration into your own platform?</b> We design forms, consent records and retention rules in from the start. See our <a class="link" href="https://flavoureak.co.uk/events/">events sector page</a>.</div>
<h2>Questions people ask</h2>
<h3>Do I need consent to collect dietary requirements?</h3>
<p>Dietary requirements that reveal health, such as allergies, or religious belief, such as halal or kosher meals, are special category data under UK GDPR. Processing them needs an Article 9 condition as well as a lawful basis, and for events that is usually explicit consent. Keep the field optional and delete the data soon after the event.</p>
<h3>Can I share my attendee list with sponsors?</h3>
<p>Sharing an attendee list with sponsors usually needs attendees’ consent, given separately from registration, with each sponsor named. Generic wording such as “selected partners” is not specific enough. Badge scanning at stands, where attendees choose to share their details, is a common alternative.</p>
<h3>Can I email attendees about future events?</h3>
<p>Under PECR you can email individuals about future events if they consented, or if the soft opt-in applies because they bought a ticket and were offered a way to opt out at the time and in every message. The emails must be about your own similar products or services.</p>
]]></content:encoded>
</item>
<item>
<title>Digital RAMS: risk assessments and method statements on a phone</title>
<link>https://flavoureak.co.uk/blog/digital-rams/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/digital-rams/</guid>
<pubDate>Wed, 08 Apr 2026 08:00:00 GMT</pubDate>
<category>Construction</category>
<description>RAMS pair a risk assessment, which the law requires, with a method statement showing how the work is done safely. On a phone they can be briefed, signed and version-controlled on site.</description>
<content:encoded><![CDATA[<ul><li>RAMS stands for risk assessment and method statement, a pair of documents in which the risk assessment identifies hazards and controls and the method statement sets out how a task will be done safely, step by step.</li><li>Employers must assess risks under the Management of Health and Safety at Work Regulations 1999 and, with five or more employees, record the significant findings; method statements are common practice and often required by contract.</li><li>Digital RAMS make it easy to brief the current version on site, capture each worker’s sign-off with a name and time, and show an auditor who saw which version.</li></ul><p>Digital RAMS are risk assessments and method statements written, briefed and signed on a phone or tablet instead of on paper. RAMS stands for risk assessment and method statement, where the risk assessment identifies hazards and how they are controlled, and the method statement sets out how a particular task will be done safely, step by step. Done well on a phone, RAMS reach the people doing the work in their current version, and the sign-off records who was briefed, on what and when.</p>
<h2>What RAMS are</h2>
<p>RAMS is an industry term for two documents that are usually prepared together for a task or a package of work.</p>
<ul>
<li><b>A risk assessment</b> identifies the hazards in the work, who could be harmed and how, how serious the risk is, and the precautions that control it.</li>
<li><b>A method statement</b> describes how the task will be carried out, including the sequence of work, the people and competences needed, the plant and equipment, the access arrangements, and how the controls from the risk assessment apply at each step.</li>
</ul>
<p>Principal contractors commonly ask each subcontractor to submit RAMS for review before starting a package, such as roofing, groundworks or steel erection, and then to brief them to the operatives who will do the work.</p>
<h2>What the law requires</h2>
<p>The legal duty is to assess risk. Regulation 3 of the Management of Health and Safety at Work Regulations 1999 requires every employer to make a suitable and sufficient assessment of the risks to its employees and to others affected by its work. Employers with five or more employees must record the significant findings. HSE’s guidance describes the process as identifying hazards, assessing the risks, controlling them, recording the findings and reviewing the controls.</p>
<p>No law names a method statement as a required document. It is a well-established way of showing how risks will be controlled for a complex or high-risk task, HSE guidance refers to method statements for some types of work, and principal contractors and clients often require them in their contracts. In practice, a builder working for a main contractor will rarely start work without accepted RAMS.</p>
<h2>Where RAMS sit alongside CDM 2015</h2>
<p>The Construction (Design and Management) Regulations 2015 set out duties for clients, designers, principal designers, principal contractors and contractors on construction projects. Contractors must plan, manage and monitor their work so that, so far as is reasonably practicable, it is carried out without risks to health and safety. Every project needs a construction phase plan, drawn up by the principal contractor or, where there is only one contractor, by that contractor. Contractors must give workers under their control appropriate supervision, instructions and information, including a suitable site induction.</p>
<p>RAMS are one of the main ways contractors meet those duties for specific tasks. The construction phase plan sets out the arrangements for the project as a whole. The RAMS for each package show how that package will be done safely, and the briefing record shows that the information reached the people doing it. For how this and other rules shape construction software, see <a class="link" href="https://flavoureak.co.uk/blog/construction-rules/">the unglamorous half of construction software</a>.</p>
<h2>Briefing and sign-off on site</h2>
<p>Paper RAMS tend to fail in the same ways. A long generic document is printed, signed by a supervisor and filed in the site cabin. Operatives sign a sheet they haven’t read. The method changes after a design revision, and the old copy is still in the folder. When something goes wrong, nobody can show which version was briefed or who was there.</p>
<p>Digital RAMS can’t fix a poor assessment. They make the good habits easier to repeat every time, with the right version briefed to the right people and a record that can be found later.</p>
<h3>How to run a briefing on a phone</h3>
<p>A RAMS briefing is the conversation in which a supervisor explains the method and the controls to the operatives before they start. A signature is evidence that the briefing happened, and the briefing itself is what keeps people safe. A phone helps with both.</p>
<ol>
<li><b>Open the current version on site.</b> The supervisor opens the RAMS for the task on a phone or tablet, and the app shows the revision number and approval status.</li>
<li><b>Brief it in person.</b> Walk through the sequence, the hazards and the controls, at the place where the work will happen if possible.</li>
<li><b>Check understanding.</b> Ask operatives to explain the key controls back, and use pictures or a translated summary where English isn’t someone’s first language.</li>
<li><b>Capture each sign-off.</b> Each operative confirms on the device, and the record stores their name, the time, the RAMS revision and, where useful, the location.</li>
<li><b>Brief late starters.</b> Anyone who joins later is briefed before starting, and their sign-off is added to the same record.</li>
<li><b>Stop and re-brief when things change.</b> If the method has to change on site, the work stops, the RAMS are revised, and the new version is briefed.</li>
</ol>
<p>An electronic sign-off with a name, date, time and document version is generally accepted as a record of a briefing. Check what your clients and principal contractors require, as some specify their own forms or systems.</p>
<h2>Version control</h2>
<p>Every RAMS document should have a revision number, a date, an author and an approver. When it changes, the old version is kept, marked as superseded, and never shown as current. A digital system makes this automatic, and it links each sign-off to the exact revision that was briefed.</p>
<table>
<thead><tr><th>Event</th><th>What should happen</th><th>What the record shows</th></tr></thead>
<tbody>
<tr><td>First issue</td><td>Written, reviewed and approved</td><td>Revision A, author, approver, date</td></tr>
<tr><td>Submitted to principal contractor</td><td>Accepted, or returned with comments</td><td>Status and comments with dates</td></tr>
<tr><td>Briefed on site</td><td>Operatives briefed, then sign</td><td>Names, times and the revision briefed</td></tr>
<tr><td>Design or method change</td><td>Revised, re-approved, old version superseded</td><td>Revision B and the reason for the change</td></tr>
<tr><td>Re-briefed</td><td>Affected operatives briefed on the change</td><td>New sign-offs against revision B</td></tr>
<tr><td>Incident or near miss</td><td>Controls reviewed</td><td>Review date and outcome</td></tr>
</tbody>
</table>
<h2>Audits and inspections</h2>
<p>When a client, a principal contractor or an HSE inspector asks about a task, the questions are predictable. Was the work assessed? Were the controls appropriate? Did the people doing it know about them? A digital record answers the last question directly, by showing the revision, the briefing and every sign-off with times. It also makes internal audits quicker, because a manager can see across all jobs which RAMS are out of date, waiting for approval or missing sign-offs.</p>
<p>Keep the RAMS with the job they belong to. When they live in the same record as the schedule, the subcontractor details and the <a class="link" href="https://flavoureak.co.uk/blog/site-diary-what-to-record/">site diary</a>, it is easy to see which packages were on site on a given day and which RAMS covered them.</p>
<h2>What to look for in a digital RAMS tool</h2>
<ul>
<li>It works offline, because briefings happen in places without signal.</li>
<li>It shows revision and approval status clearly on every screen.</li>
<li>It captures sign-offs against a specific revision, with name and time.</li>
<li>It builds RAMS from your own templates, which you then edit for the task and site.</li>
<li>It exports a clean PDF for principal contractors who ask for one.</li>
<li>It keeps superseded versions and the history of changes.</li>
</ul>
<p>In Estimark, health and safety paperwork sits in the same job record as estimates, scheduling and site diaries, the app works offline on site, and it includes a field worker mode.</p>
<div class="callout"><b>Building health and safety into your own software?</b> We design for offline use and audit trails from the start. See our <a class="link" href="https://flavoureak.co.uk/construction/">construction sector page</a> or <a class="link" href="https://flavoureak.co.uk/estimark/">Estimark</a>.</div>
<h2>Questions people ask</h2>
<h3>Are RAMS a legal requirement?</h3>
<p>Risk assessment is a legal requirement under the Management of Health and Safety at Work Regulations 1999, and employers with five or more employees must record the significant findings. No law names a method statement as a required document, though principal contractors and clients commonly require RAMS before work starts. CDM 2015 also requires work to be planned and workers to receive appropriate supervision, instructions and information.</p>
<h3>What is the difference between a risk assessment and a method statement?</h3>
<p>A risk assessment identifies hazards, who could be harmed and the controls needed. A method statement describes how a specific task will be carried out safely, step by step, applying those controls. Together they are known as RAMS.</p>
<h3>Can operatives sign RAMS on their phones?</h3>
<p>Operatives can confirm a RAMS briefing on a phone or tablet, and the record should store their name, the time and the revision briefed. Many contractors use the supervisor’s device so the sign-off happens at the briefing itself. Check whether your principal contractor has its own requirements.</p>
<h3>How often should RAMS be reviewed?</h3>
<p>RAMS should be reviewed whenever the work, the site conditions, the design or the people change, and after any incident or near miss. A risk assessment must be reviewed if there is reason to suspect it is no longer valid or if there has been a significant change. Each review should produce a new revision with a date and a reason.</p>]]></content:encoded>
</item>
<item>
<title>How to write web pages that AI answers quote, with a before and after</title>
<link>https://flavoureak.co.uk/blog/writing-for-ai-answers/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/writing-for-ai-answers/</guid>
<pubDate>Thu, 02 Apr 2026 08:00:00 GMT</pubDate>
<category>GEO</category>
<description>Answer engines lift short passages from pages. Pages that answer first, define terms, give facts with units and keep each paragraph self-contained are the ones that get quoted.</description>
<content:encoded><![CDATA[<ul><li>AI answer engines quote passages, so each paragraph should answer one question on its own without relying on the paragraph before it.</li><li>Put the direct answer in the first one or two sentences of a page or section, then explain, qualify and give examples.</li><li>Specific facts with units, dates, consistent names and a visible author are easier to quote safely than general marketing claims.</li></ul><p>To write a page that AI answers quote, put the direct answer first, define the main term in one sentence, and write paragraphs that each make sense when lifted out alone. Writing for AI answers means structuring a page so that a search-based assistant such as ChatGPT search, Perplexity, Microsoft Copilot or Google’s AI Overviews can find a passage that answers a question and repeat it accurately with a citation. Most of the work is plain, specific writing.</p>
<h2>How answer engines use a page</h2>
<p>An answer engine turns a question into one or more searches, reads some of the pages it finds and writes a reply from the passages that fit best. It then cites the pages those passages came from. The full process is set out in <a class="link" href="https://flavoureak.co.uk/blog/geo-become-the-source/">how to become the source an AI answer cites</a>.</p>
<p>The unit being judged is the passage, usually a paragraph or a short section. The engine needs to find it, understand it without the rest of the page, and be confident it is correct. Every rule below follows from those three needs.</p>
<h2>Put the answer and the definition first</h2>
<h3>Answer in the first two sentences</h3>
<p>Open the page, and each section, with the direct answer to the question it is about. Explanation, history and caveats come after. A reader skimming the top of the page gets what they came for, and an engine looking for a passage that answers the query finds it at once.</p>
<p>A useful test is to cover everything below the first paragraph. If the question in the heading is still answered, the page is answer-first. If the first paragraph only says the topic is important or complicated, the answer is buried and needs moving up.</p>
<h3>Define terms in one sentence</h3>
<p>Many questions are definitions in disguise, such as what a retention is, what a pay less notice is or what multi-tenancy means. A one-sentence definition in the form “X is a Y that does Z” is among the easiest things on the web to quote. Write it near the top, use the term the way your customers say it, and give the formal name alongside if it differs.</p>
<p>Keep the definition literal. “A retention is a percentage of each payment that the client holds back until the work is complete and defects are put right” can be quoted as it stands. “Retentions are a thorny issue for the industry” cannot.</p>
<h2>Write passages that stand alone</h2>
<p>A paragraph that starts “This means…” or “As mentioned above…” depends on something the engine may leave out. Repeat the subject instead of using a pronoun that points backwards. Name the product, the rule or the place again. It reads slightly more formally to a person and much more clearly out of context.</p>
<ul>
<li><b>One idea per paragraph.</b> If a paragraph answers two questions, split it in two.</li>
<li><b>Name the subject.</b> Write “Estimark saves each change to the phone first” rather than “It saves each change first”.</li>
<li><b>Keep conditions with the claim.</b> If a rule only applies to VAT-registered businesses, say so in the same sentence, so the condition travels with the quote.</li>
<li><b>Avoid “click here” and “see below”.</b> Link text and references should make sense on their own.</li>
</ul>
<h2>Use specific facts, units and tables</h2>
<p>Numbers, units, dates and names survive being summarised. Vague claims don’t, and an engine has no reason to repeat them. “Most installs take two to four days” is more useful than “fast turnaround”. “Deductions are 20% for registered subcontractors and 30% for unregistered ones” is more useful than “we handle CIS”.</p>
<p>Give the unit every time, whether that is days or working days, pounds including or excluding VAT, or a price per user per month. A number without a unit is likely to be misquoted or ignored. When a fact can change, say when it was true, for example “as of October 2026”.</p>
<h3>Tables for comparisons</h3>
<p>When a page compares options, a table with a header row is clearer than paragraphs. Each row is a self-contained fact, and the column headings supply the context. Use real HTML tables rather than images of tables, and keep one fact per cell.</p>
<table>
<thead><tr><th>Page element</th><th>Hard to quote</th><th>Easy to quote</th></tr></thead>
<tbody>
<tr><td>Opening</td><td>A story or a welcome</td><td>A two-sentence answer with a definition</td></tr>
<tr><td>Claims</td><td>“Industry-leading”, “competitive”</td><td>A number with a unit and a date</td></tr>
<tr><td>Comparisons</td><td>Long paragraphs</td><td>An HTML table with headings</td></tr>
<tr><td>Names</td><td>Three versions of the product name</td><td>One name used every time</td></tr>
<tr><td>Authorship</td><td>No author or date</td><td>A named author, a published date and an updated date</td></tr>
<tr><td>Questions</td><td>Buried in body text</td><td>A heading with a short answer below it</td></tr>
</tbody>
</table>
<h2>Keep names, dates and authorship consistent</h2>
<p>Use one name for your company and one for each product, and use them the same way on every page and profile. If the site says “Acme Build”, the footer says “Acme Construction Ltd” and a directory says “Acme Builders”, an engine has to guess whether they are the same business. The registered name can appear once alongside the trading name to connect them. <a class="link" href="https://flavoureak.co.uk/blog/entities-and-brand-mentions/">Entities and brand mentions</a> covers this in more depth.</p>
<p>Show who wrote the page, when it was published and when it was last updated. A dated page lets an engine, and a reader, judge whether a fact is current. Structured data can carry the same information in machine-readable form, as described in <a class="link" href="https://flavoureak.co.uk/blog/structured-data-for-business/">structured data for business sites</a>, but the visible text comes first.</p>
<h2>A before and after rewrite</h2>
<p>Here is a typical paragraph from a builder’s services page, followed by a rewrite. Acme Build is a fictional business used for the example.</p>
<p><b>Before.</b> “At Acme, we pride ourselves on delivering exceptional extensions that transform the way you live. Our experienced team works closely with you every step of the way, and our competitive pricing means you get outstanding value. Get in touch today to find out more.”</p>
<p><b>After.</b> “Acme Build designs and builds single-storey and two-storey house extensions in Leeds and within 20 miles of the city. A single-storey rear extension typically takes 10 to 14 weeks on site once planning permission or a lawful development certificate is in place. Every quote is fixed for 30 days, itemised by trade and shows the price with and without VAT.”</p>
<p>The first version contains no fact an engine could repeat. The second answers four questions people actually ask, covering what the business does, where it works, how long a job takes and how it prices. Each sentence works if quoted alone. Nothing in it is a superlative, so nothing in it needs evidence that isn’t there. The figures belong to a fictional firm; yours should come from your own jobs.</p>
<p>Apply the same test to every phrase like “passionate about quality” or “tailored solutions”. Each one gives an answer engine nothing to use and pushes useful sentences down the page. Replace it with the fact it was standing in for, such as what you check and how often, or what changes from one customer to the next.</p>
<h2>Add the questions customers ask</h2>
<p>Collect the questions that arrive by phone and email before someone buys, and answer the common ones on the page as headings with short answers beneath. Phrase each heading the way the customer phrases the question. Keep each answer to two to four sentences that include the subject, so the answer still makes sense without the heading.</p>
<ol>
<li>List twenty questions from real enquiries and sales calls.</li>
<li>Group them by page, such as pricing, process, timing and guarantees.</li>
<li>Write a two- to four-sentence answer for each, starting with the answer itself.</li>
<li>Add the facts that make it checkable, such as numbers, units, dates and conditions.</li>
<li>Read each answer alone and fix any that depend on another paragraph.</li>
</ol>
<div class="callout"><b>Building a site or product where content matters?</b> Our <a class="link" href="https://flavoureak.co.uk/geo/">GEO guide</a> sets out how we structure pages so people and answer engines can use them.</div>
<h2>Questions people ask</h2>
<h3>Does writing for AI answers hurt normal SEO?</h3>
<p>Writing for AI answers does not conflict with SEO. Answer-first paragraphs, clear headings, specific facts and consistent names are what search engines have rewarded for years. A page still has to be crawlable, indexed and useful to rank, and the same page is then easier for an answer engine to quote.</p>
<h3>How long should a passage be to get quoted?</h3>
<p>No passage length guarantees a quotation in an AI answer. In practice, a paragraph of two to four sentences that answers one question completely is easy to lift. Very long paragraphs tend to be summarised, and one-line fragments often lack the context needed to be quoted accurately.</p>
<h3>Should every page have an FAQ section?</h3>
<p>An FAQ section helps when it answers questions customers really ask about that page’s subject. A generic list of questions copied across every page adds repetition without adding facts. Write the questions from real enquiries and keep the answers specific to the page.</p>
<h3>Can I guarantee that ChatGPT or Google will quote my page?</h3>
<p>No one can guarantee a citation in an AI answer, because each engine chooses its sources for each query. What a business can control is whether its pages are found, understood and safe to quote. Clear, specific and consistent pages give an engine the best material to work with.</p>]]></content:encoded>
</item>
<item>
<title>Structured data for business and SaaS websites: which schema to use</title>
<link>https://flavoureak.co.uk/blog/structured-data-for-business/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/structured-data-for-business/</guid>
<pubDate>Fri, 27 Mar 2026 08:00:00 GMT</pubDate>
<category>SEO</category>
<description>Structured data is code that tells search engines what a page describes. A few schema.org types cover most business and SaaS sites. It can earn rich results, but it won’t lift rankings alone.</description>
<content:encoded><![CDATA[<ul><li>Structured data is information added to a page in a standard format, usually schema.org vocabulary written as JSON-LD, that describes what the page is about.</li><li>Most business and SaaS sites need Organization and WebSite on the home page, SoftwareApplication or Product on product pages, BlogPosting on articles and BreadcrumbList across the site.</li><li>Structured data can make a page eligible for rich results without guaranteeing them, and since 2023 Google shows FAQ rich results only for authoritative government and health sites.</li></ul><p>Structured data is a block of code on a web page that describes the page’s content in a standard vocabulary, so search engines and other software don’t have to guess what it is about. For business and SaaS websites, the useful set is small, usually Organization, WebSite, SoftwareApplication or Product, Article or BlogPosting, and BreadcrumbList. It can make pages eligible for rich results in Google, but it doesn’t raise rankings on its own.</p>
<h2>What structured data is</h2>
<p>The vocabulary comes from schema.org, a shared project founded by Google, Microsoft, Yahoo and Yandex. It defines types, such as Organization or SoftwareApplication, and properties for each type, such as <code>name</code>, <code>logo</code> or <code>offers</code>.</p>
<p>The format Google recommends is JSON-LD. It sits in a script element of type <code>application/ld+json</code>, separate from the visible HTML. That makes it easy to generate from your own data and hard to break with a design change. A minimal Organization block declares <code>"@type": "Organization"</code>, then gives the company’s name, website address and logo, and lists the profiles that confirm it exists, such as its Companies House record and its official social accounts.</p>
<p>Two older formats, Microdata and RDFa, add the same information as attributes woven into the page’s HTML. They still work, but they are harder to maintain, and there is rarely a reason to choose them for a new site.</p>
<h2>What structured data does and doesn’t do</h2>
<p>Structured data helps search engines understand a page and the entities on it, such as who the business is, what the product is, who wrote an article and when. That understanding can make a page eligible for a rich result, such as review stars, a product price or a breadcrumb trail beside the listing.</p>
<p>Eligibility isn’t a promise. Google decides whether to show a rich result, and its guidelines say so. Adding markup doesn’t, by itself, move a page up the rankings. Markup that describes content readers can’t see on the page breaks Google’s guidelines and can lead to a manual action, which removes the page’s eligibility for rich results.</p>
<p>For AI answer engines the picture is less clear. Some read the full HTML of pages they retrieve, and JSON-LD is part of that HTML, but the major providers haven’t published how much weight they give it. Treat it as a clear, machine-readable statement of facts about your business, with an uncertain direct benefit. Describing yourself the same way everywhere matters more, as covered in <a class="link" href="https://flavoureak.co.uk/blog/entities-and-brand-mentions/">entities and brand mentions</a>.</p>
<h2>The schema types most business sites need</h2>
<table><thead><tr><th>Type</th><th>Where it goes</th><th>What it describes</th><th>Possible effect in Google</th></tr></thead><tbody>
<tr><td>Organization</td><td>Home page or about page</td><td>The company, logo, contact details and official profiles</td><td>Helps Google show the right logo and business details</td></tr>
<tr><td>WebSite</td><td>Home page</td><td>The site and its preferred name</td><td>Helps Google choose the site name shown in results</td></tr>
<tr><td>SoftwareApplication</td><td>Product and pricing pages for software</td><td>The app, its category, operating system, price and ratings</td><td>Software rich results, where genuine ratings or reviews are included</td></tr>
<tr><td>Product</td><td>Product pages for physical or packaged goods</td><td>The item, price, availability and reviews</td><td>Product snippets and merchant listings</td></tr>
<tr><td>Article or BlogPosting</td><td>Articles and blog posts</td><td>Headline, author, dates and images</td><td>Clearer authorship and dates</td></tr>
<tr><td>BreadcrumbList</td><td>Every page below the home page</td><td>The page’s position in the site’s hierarchy</td><td>A breadcrumb trail in results</td></tr>
<tr><td>LocalBusiness</td><td>Sites for businesses with premises or a service area</td><td>Address, opening hours and area served</td><td>Supports local business details</td></tr>
</tbody></table>
<h3>Organization and WebSite</h3>
<p>Put both on the home page. Organization should use the same name, address and logo you use everywhere else, with <code>sameAs</code> links to your official profiles. WebSite tells Google the name you want shown for the site. Google retired the sitelinks search box in 2024, so there is no longer a search feature to gain from adding a search action to WebSite markup.</p>
<h3>SoftwareApplication for SaaS</h3>
<p>SaaS product pages fit SoftwareApplication, or its subtype WebApplication, better than Product. Include the name, application category, operating system (for a browser-based product, say so), and an offer with the price. Google’s software app rich result also needs ratings or reviews. Use them only if they come from real users and appear on the page. Invented ratings break Google’s policies, and fake reviews are banned under the Digital Markets, Competition and Consumers Act 2024.</p>
<h3>Product for online shops</h3>
<p>Online shops should mark up each product page with Product, including an offer with the price, currency and availability, and reviews only where they are real and shown on the page. Keep the markup consistent with any product feed you send to Google Merchant Center, where a price that doesn’t match the landing page can get a product disapproved.</p>
<h3>LocalBusiness for firms with premises</h3>
<p>LocalBusiness is a more specific kind of Organization for businesses with premises or a service area, such as a builder or a shop. Use the most specific subtype that fits, such as HomeAndConstructionBusiness or Store, and keep the address and opening hours identical to your Google Business Profile.</p>
<h3>Articles and breadcrumbs</h3>
<p>BlogPosting, a subtype of Article, should carry the headline, author, publication date, modified date and an image. Change <code>dateModified</code> only when the content really changes. BreadcrumbList should match the breadcrumb trail visitors can see on the page.</p>
<h2>FAQPage and the 2023 change</h2>
<p>FAQPage markup used to produce expandable questions under a search result, and many sites added it to every page they could. In August 2023 Google limited FAQ rich results to well-known, authoritative government and health websites. For every other site, the markup still validates, but the expandable result no longer appears in normal results. Google phased out HowTo rich results around the same time.</p>
<p>If a page genuinely contains a set of questions and answers, the markup does no harm and describes the content accurately. Just don’t add it expecting extra space in Google’s results.</p>
<h2>How to add structured data and test it</h2>
<ol>
<li><b>Generate it from your data.</b> Build the JSON-LD from the same database or content files as the page, so the price in the markup can’t drift from the price on the screen.</li>
<li><b>Render it in the HTML.</b> Include it in the server-rendered or static HTML. Google can read JSON-LD added by JavaScript, but other crawlers may not run scripts, as explained in <a class="link" href="https://flavoureak.co.uk/blog/seo-for-javascript-sites/">SEO for JavaScript sites</a>.</li>
<li><b>Mark up only what is visible.</b> Every fact in the markup should appear on the page for readers.</li>
<li><b>Use the Rich Results Test.</b> Google’s Rich Results Test shows which rich results a page is eligible for, with any errors and warnings.</li>
<li><b>Use the Schema Markup Validator.</b> The validator at validator.schema.org checks markup against the whole schema.org vocabulary, including types Google doesn’t use for rich results.</li>
<li><b>Watch Search Console.</b> Search Console reports problems with supported rich result types across the whole site once Google has crawled it.</li>
</ol>
<p>Keep the markup current as the site changes. Out-of-date prices, removed products and old logos in structured data cause the same confusion as out-of-date text on the page.</p>
<div class="callout"><b>Launching a SaaS or business website?</b> We build structured data from the same content as each page, so the two can’t disagree. <a class="link" href="https://flavoureak.co.uk/seo/">See our SEO guide</a>.</div>
<h2>Questions people ask</h2>
<h3>Does structured data improve SEO rankings?</h3>
<p>Not directly. Structured data helps search engines understand a page and can make it eligible for rich results, which can make a listing more visible. Rankings depend on the content and many other signals.</p>
<h3>What is the best format for structured data?</h3>
<p>JSON-LD is the format Google recommends. It sits in its own script element, separate from the visible HTML, so it is easier to generate and maintain than Microdata or RDFa, which are woven into the page markup.</p>
<h3>Should a SaaS website use Product or SoftwareApplication schema?</h3>
<p>SoftwareApplication, or its WebApplication subtype, describes software more accurately, with properties for application category and operating system. Product suits physical goods and packaged items sold through an online shop.</p>
<h3>Is FAQ schema still worth adding?</h3>
<p>Only where a page genuinely contains questions and answers. Since August 2023, Google shows FAQ rich results only for well-known government and health websites, so most sites gain no extra display in Google from it.</p>]]></content:encoded>
</item>
<item>
<title>How to send push notifications people don’t turn off</title>
<link>https://flavoureak.co.uk/blog/push-notifications-that-help/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/push-notifications-that-help/</guid>
<pubDate>Sun, 22 Mar 2026 08:00:00 GMT</pubDate>
<category>Engineering</category>
<description>Send fewer, more useful messages, ask for permission when the value is obvious, give people control over each type, and get separate consent for marketing.</description>
<content:encoded><![CDATA[<ul><li>Push notifications are sent from your server through Apple Push Notification service on iPhone and Firebase Cloud Messaging on Android, and current versions of both ask the user’s permission.</li><li>Ask for permission at a moment when the benefit is obvious, because on iPhone a refusal can only be reversed in the Settings app.</li><li>Treat marketing notifications like marketing email, with clear opt-in consent that is separate from the system permission and easy to withdraw in the app.</li></ul><p>Push notifications people keep switched on are timely, specific to them and expected, and they are split by type so each person can choose what they receive. A push notification is a short message sent from your server to an app on someone’s phone, which appears even when the app is closed. Ask for permission when the value is obvious, send transactional messages that save people effort, and treat marketing as a separate, opt-in category.</p>
<h2>How push notifications work</h2>
<p>When an app is set up for notifications, the phone gives it a device token, a long identifier that addresses that app on that device. The app sends the token to your server. To send a notification, your server passes the message and the token to the platform’s delivery service, which delivers it to the phone.</p>
<ul><li><b>APNs</b>, the Apple Push Notification service, delivers notifications to iPhone and iPad apps.</li><li><b>FCM</b>, Firebase Cloud Messaging, is Google’s service for Android. It can also send to iOS apps, by passing messages on to APNs, and to web browsers.</li></ul>
<p>Tokens change when an app is reinstalled or a phone is restored, and both services tell your server when a token is no longer valid, so remove those promptly. A notification payload is small, around 4 KB, so it carries a short message and an identifier for the app to fetch the rest. Delivery is best effort. A notification should never be the only place an important message lives, and a booking confirmation, for example, should also be visible in the app and sent by email.</p>
<h2>Transactional and marketing notifications</h2>
<p>Every notification you plan should fall into one of two groups, because the rules and people’s tolerance differ for each.</p>
<table><thead><tr><th></th><th>Transactional</th><th>Marketing</th></tr></thead><tbody><tr><td>Purpose</td><td>Tells someone about something they did, asked for or are directly involved in</td><td>Promotes products, offers or more use of the app</td></tr><tr><td>Examples</td><td>Booking confirmed, lesson starting soon, job assigned to you, payment failed</td><td>A sale, a new feature to try, a message to people who have not opened the app recently</td></tr><tr><td>Permission needed</td><td>The system permission</td><td>The system permission plus separate marketing consent</td></tr><tr><td>Default setting</td><td>On, with a switch per category</td><td>Off until the person opts in</td></tr><tr><td>What goes wrong</td><td>Late, duplicated or sent at night</td><td>Too frequent, irrelevant, or sent without consent</td></tr></tbody></table>
<p>Apple’s App Review Guidelines say push notifications must not be required for an app to work, should not carry sensitive personal or confidential information, and should not be used for promotions or direct marketing unless people have explicitly opted in through consent wording in the app, with a way to opt out in the app. Notifications also appear on the lock screen, where anyone nearby can read them. “Your lesson report is ready” is fine. The contents of the report are not.</p>
<h2>When to ask for permission</h2>
<p>On iPhone, the system permission prompt appears only once. If the person taps Don’t Allow, the app cannot ask again, and the only way back is through the Settings app. Since Android 13, Android also asks for permission before an app can show notifications, and after repeated refusals the prompt stops appearing.</p>
<p>That makes timing the most important decision. Asking on first launch, before the person has seen what the app does, invites a refusal. Ask at the moment the benefit is clear, such as straight after someone books a lesson or is assigned their first job. A short screen in your own design can explain what you will send before the system prompt appears, so the person knows what they are agreeing to.</p>
<p>iPhone also offers provisional authorisation, which delivers notifications quietly to Notification Centre without asking first. The person sees them there and decides whether to keep them. It suits apps whose value is easier to show than to explain.</p>
<h2>Giving people control over notifications</h2>
<p>People are more likely to switch everything off when the only choice is all or nothing. Give them categories and sensible limits.</p>
<ul><li><b>Per-type settings in the app.</b> Booking reminders, messages, reports and marketing as separate switches.</li><li><b>Android notification channels.</b> Since Android 8, each notification belongs to a channel that the person can silence or adjust in system settings. Match the channels to the categories in your app.</li><li><b>Interruption levels on iPhone.</b> Mark only genuinely urgent messages as time-sensitive, so they can break through Focus. Critical alerts that ignore the mute switch need a special entitlement from Apple.</li><li><b>Quiet hours.</b> Hold non-urgent messages overnight in the person’s own time zone, and batch low-priority updates into a daily digest.</li><li><b>Frequency limits.</b> Cap how many messages one person can receive in a day, across every feature.</li></ul>
<p>Time zones need care. In CurioKids, bookings are shown in the family’s own time zone, and the same principle applies to any notification about a booking. A reminder calculated in the server’s time zone arrives at the wrong hour for a family abroad. The same care applies to any app built around a schedule, such as <a class="link" href="https://flavoureak.co.uk/blog/what-an-event-app-needs/">an event app</a>.</p>
<h2>Marketing notifications and UK law</h2>
<p>Under the Privacy and Electronic Communications Regulations (PECR), unsolicited direct marketing sent by electronic mail, which covers email and text messages, generally needs the recipient’s prior consent, with a narrow exception for existing customers known as the soft opt-in. Whether a particular push notification counts as electronic mail under PECR depends on how it is sent and stored, and the position is less clear-cut than it is for email. UK GDPR applies either way, because the token and the data used to choose who gets a message relate to identifiable people.</p>
<p>The safe approach, and the one we’d recommend, is to treat marketing pushes as you would marketing email. Ask for marketing consent separately from the system permission, with plain wording about what you will send. Record when and how consent was given. Make it as easy to withdraw as it was to give, inside the app. Tapping Allow on a system prompt that says nothing about marketing is unlikely to count as consent to it. Take legal advice if marketing notifications are central to your plans.</p>
<h2>Measuring whether notifications help</h2>
<p>Measure each category separately, because a useful reminder and an unwelcome promotion can hide each other in an average.</p>
<ol><li><b>Permission rate.</b> The share of people who allow notifications when asked, broken down by where in the app you asked.</li><li><b>Opt-outs.</b> How many people switch a category off, or disable notifications in system settings. The app can check its permission status each time it opens.</li><li><b>Opens and actions.</b> Taps on the notification, and whether the person then did the thing it was about.</li><li><b>Outcomes.</b> Fewer missed lessons, faster responses to assigned jobs, fewer failed payments left unresolved.</li><li><b>Complaints and uninstalls</b> in the days after a campaign.</li></ol>
<p>Outcomes matter most. A notification that raises taps but also raises opt-outs is spending trust you will need for the messages that matter.</p>
<div class="callout"><b>Building an app with notifications?</b> Decide the categories, consent wording and quiet hours before writing the first message. <a class="link" href="https://flavoureak.co.uk/mobile/">See how we approach mobile apps</a>.</div>
<h2>Questions people ask</h2>
<h3>Can a web app send push notifications on iPhone?</h3>
<p>Yes, since iOS 16.4, but only once the web app has been added to the home screen and the person grants permission. A website open in a browser tab cannot send them. Our comparison of <a class="link" href="https://flavoureak.co.uk/blog/pwa-or-native-app/">progressive web apps and native apps</a> covers the other differences.</p>
<h3>Do I need consent to send push notifications in the UK?</h3>
<p>The phone’s own permission is needed for any visible notification. For marketing, get separate, clear opt-in consent, because PECR and UK GDPR may apply and Apple’s guidelines expect it. Transactional messages about something the person is doing do not need marketing consent.</p>
<h3>How many push notifications are too many?</h3>
<p>There is no fixed number. The right level depends on how useful each message is to the person receiving it. Watch opt-outs by category, and if one type of message is switched off more often than the rest, send it less often or stop.</p>
<h3>Why are some push notifications not delivered?</h3>
<p>APNs and FCM deliver on a best-effort basis. A phone that is off or without signal may receive a notification late, and when several are waiting, APNs keeps only the most recent one for each app. Invalid tokens, disabled permissions and Focus settings also stop messages appearing.</p>]]></content:encoded>
</item>
<item>
<title>How to extract structured data from invoices and forms with AI</title>
<link>https://flavoureak.co.uk/blog/ai-document-extraction/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/ai-document-extraction/</guid>
<pubDate>Mon, 16 Mar 2026 08:00:00 GMT</pubDate>
<category>AI</category>
<description>Define a schema, read the document with OCR, a layout model or a vision model, validate every value against rules, and send uncertain fields to a person with the source highlighted.</description>
<content:encoded><![CDATA[<ul><li>AI document extraction turns documents such as invoices and delivery notes into structured fields, using OCR, layout-aware models or vision-language models.</li><li>Validate every extracted value with rules, such as checking that line totals add up and that VAT is at a valid UK rate, before trusting any confidence score.</li><li>Link each value to the place on the page it came from, so a reviewer can check it at a glance and you can measure accuracy over time.</li></ul><p>AI document extraction is the process of reading documents such as invoices, delivery notes, tenders and forms and turning their contents into structured data, such as a supplier name, a date, line items and totals, that software can use. A dependable system combines a reading method, which can be OCR, a layout model or a vision-language model, with a fixed schema, validation rules and human review for anything uncertain. The model reads the document, and the rules and the review decide what is trusted.</p>
<h2>Three ways to read a document</h2>
<h3>OCR</h3>
<p>Optical character recognition (OCR) converts an image of a page into text, with the position of each word. It is mature, cheap and fast, and it understands nothing. It can tell you that “1,240.00” appears near the bottom right of the page, but not that it is the total. Older extraction systems paired OCR with a template for each supplier’s layout, which works for a few fixed formats and breaks when a supplier changes theirs.</p>
<h3>Layout models</h3>
<p>Layout models, sometimes sold as document AI, use both the text and its position on the page to find fields, tables and key-value pairs. Cloud services from Google, Amazon and Microsoft offer pre-built models for common documents such as invoices and receipts, and most can be customised for your own. They return the position of each field and a confidence score, which makes review easier.</p>
<h3>Vision-language models</h3>
<p>General-purpose multimodal language models can look at a page image and return the fields you ask for in the format you specify. They handle unfamiliar layouts, handwriting and unusual documents without training, and can follow instructions such as “the delivery address, not the invoice address”. They can also return a confident value that is not on the page, and they do not reliably say where on the page a value came from.</p>
<table><thead><tr><th>Method</th><th>Strengths</th><th>Weaknesses</th><th>Suits</th></tr></thead><tbody><tr><td>OCR with templates</td><td>Cheap, fast, gives word positions</td><td>No understanding; a template per layout</td><td>A few fixed layouts at high volume</td></tr><tr><td>Layout model</td><td>Fields, tables and positions, with confidence scores</td><td>Pre-built models cover common document types, and others need training</td><td>Invoices, receipts and forms from many suppliers</td></tr><tr><td>Vision-language model</td><td>Handles new layouts and follows instructions without training</td><td>Can invent values; positions unreliable; slower per page</td><td>Varied documents, tenders and one-off forms</td></tr><tr><td>Combined</td><td>A model extracts, and OCR locates each value on the page</td><td>More to build</td><td>Anything where values must be checked and traced</td></tr></tbody></table>
<h2>Defining a schema</h2>
<p>A schema is the exact list of fields you want, with the type and format of each. Defining it first turns a vague request into something you can check. For a supplier invoice it might include these fields.</p>
<ul><li>Supplier name, and the supplier’s VAT registration number if shown.</li><li>Invoice number, invoice date and due date, as ISO 8601 dates such as 2026-03-14.</li><li>Purchase order number, if quoted.</li><li>Line items, each with a description, quantity, unit price, VAT rate and line total.</li><li>Net total, VAT total and gross total, held in pence to avoid rounding errors.</li><li>Whether the invoice states that the VAT domestic reverse charge applies.</li></ul>
<p>Many model providers can force output to match a JSON schema, which removes a whole class of formatting errors. It does not make the values correct.</p>
<h2>Validation rules for extracted data</h2>
<p>Every extracted value should pass rules before it is trusted. Rules are cheap, explainable and consistent.</p>
<ul><li><b>Arithmetic.</b> Each line total equals quantity times unit price. The lines add up to the net total. Net plus VAT equals gross, within a penny for rounding.</li><li><b>VAT rates.</b> Each line is at one of the UK rates of 20%, 5% or 0%, or is marked as exempt or outside the scope of VAT.</li><li><b>VAT numbers.</b> A UK VAT registration number usually has nine digits after the GB prefix, and HMRC offers a service to check whether one is valid.</li><li><b>Dates.</b> The invoice date is not in the future, and the due date falls after it.</li><li><b>Duplicates.</b> The same supplier and invoice number have not been processed before.</li><li><b>Matching.</b> The purchase order exists and is open, the supplier matches it, and the quantities agree with what the delivery note says arrived. Checking the order, the delivery note and the invoice against each other is known as three-way matching.</li></ul>
<p>Construction adds its own rules. A subcontractor’s invoice under CIS has to separate labour from materials, because the deduction applies only to labour, and an invoice for many construction services between VAT-registered businesses may need to show the reverse charge. These rules are described in <a class="link" href="https://flavoureak.co.uk/blog/construction-rules/">CIS, reverse charge and payment notices</a>.</p>
<h2>Confidence and human review</h2>
<p>Treat confidence as one signal among several. Layout models return a confidence score for each field. Language models can be asked for one, but a model’s own estimate of its confidence is not a dependable measure of accuracy. A better approach combines four questions for each field.</p>
<ol><li>Did the value pass every validation rule?</li><li>Was the value found on the page, in the OCR text?</li><li>Do two methods agree, such as a layout model and a vision model?</li><li>Is an error in this field costly, as it would be for the total or the bank details?</li></ol>
<p>Documents where every field passes go straight through. Documents with a failed rule or a missing value go to a review queue, with the problem fields highlighted. Bank details deserve their own rule. A change to a supplier’s bank details on an invoice is a known sign of invoice fraud, and it should always be confirmed with the supplier by phone, using a number you already hold.</p>
<h2>Linking every value to its source</h2>
<p>Each extracted value should carry the page number and the area of the page it came from. When a vision-language model does the extraction, ask it to return the exact text it read for each field, then find that text in the OCR output to get its position. If the text cannot be found on the page, flag the field, because the model may have produced the value without reading it.</p>
<p>This brings three benefits. A reviewer sees the document beside the fields, with each value highlighted where it appears, and can confirm it at a glance. An auditor can trace any number in your accounts back to the document. And every correction a reviewer makes becomes a labelled example, which gives you a running measure of accuracy by field, supplier and document type.</p>
<h2>Working out the cost per document</h2>
<p>The cost of extraction per document is the sum of three parts. Work each out with your own volumes and your provider’s current rates.</p>
<table><thead><tr><th>Part</th><th>How to calculate it</th></tr></thead><tbody><tr><td>Reading</td><td>Pages × the per-page rate of an OCR or layout service</td></tr><tr><td>Model calls</td><td>(Input tokens × input rate) + (output tokens × output rate), where input includes the page images, instructions and schema</td></tr><tr><td>Review</td><td>Share of documents reviewed × minutes per review × staff cost per minute</td></tr></tbody></table>
<p>Here is an example, with figures you should replace with your own. Suppose 3,000 invoices a month, one in five needs review, a review takes two minutes, and staff time costs £30 an hour, or £0.50 a minute. Review then averages 0.2 × 2 × £0.50 = £0.20 per invoice, or £600 a month. The reading and model costs come from your provider’s price list. Providers count page images as tokens in different ways, so measure a few real documents before you estimate.</p>
<p>The review line is often the largest of the three. That is why time spent on validation rules and source highlighting usually saves more than time spent finding a cheaper model. The same sum for any AI feature is explained in <a class="link" href="https://flavoureak.co.uk/blog/ai-cost-per-call/">what an AI feature costs per call</a>.</p>
<div class="callout"><b>Keying in invoices by hand?</b> Start with a sample of real documents and a schema, then measure accuracy field by field. <a class="link" href="https://flavoureak.co.uk/ai/">See how we build AI features</a>.</div>
<h2>Questions people ask</h2>
<h3>How accurate is AI invoice extraction?</h3>
<p>It depends on the documents, the method and the field, so measure it on your own. Take a few hundred real documents, record the correct values and compare field by field. Accuracy on clean, typed invoices is usually much higher than on photographed or handwritten delivery notes.</p>
<h3>Can AI read handwritten delivery notes?</h3>
<p>Vision-language models and modern OCR can often read clear handwriting, but accuracy drops with poor photos, carbon copies and hurried figures. Send handwritten fields to review more readily, and have a person confirm quantities and signatures.</p>
<h3>Can AI extract requirements from tender documents?</h3>
<p>Yes, though tenders are harder than invoices because they are long and every one is different. Extraction there means finding the deadline, submission method, evaluation criteria and required forms, each with a page reference. It works best combined with search over the document, so a person can ask follow-up questions and see the source.</p>
<h3>Is it safe to send invoices to an AI provider?</h3>
<p>Invoices often contain personal data, such as a sole trader’s name and address, so UK GDPR applies. Use a provider that acts as your processor under a written contract, check where the data is processed and whether it is used for training, and keep the original documents for as long as HMRC requires, which for VAT records is generally six years. Our guide to <a class="link" href="https://flavoureak.co.uk/blog/ai-and-uk-gdpr/">AI and UK GDPR</a> covers the rest.</p>]]></content:encoded>
</item>
<item>
<title>Discovery phase in software development: what happens and what you get</title>
<link>https://flavoureak.co.uk/blog/discovery-phase-explained/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/discovery-phase-explained/</guid>
<pubDate>Tue, 10 Mar 2026 08:00:00 GMT</pubDate>
<category>Working with us</category>
<description>A discovery phase is short, fixed work before a build that turns an idea into a scope, a prototype, an estimate and a risk list, so you decide with evidence before committing the full budget.</description>
<content:encoded><![CDATA[<ul><li>A discovery phase, sometimes called a discovery sprint, is a short piece of work before development that defines what to build, for whom, how, and roughly what it will cost.</li><li>At the end you should have a written scope, a clickable prototype of the main journeys, a technical approach, a ranged estimate and a list of risks with a plan for each.</li><li>For a new product, discovery usually takes a few weeks, and it saves money by finding expensive surprises while they are still cheap to change.</li></ul><p>A discovery phase is a short, fixed piece of work at the start of a software project that turns an idea into a plan you can build from. It answers what to build, for whom, how, and roughly what it will cost, and it ends with a scope, a prototype, an estimate and a list of risks. It saves money because changing a decision on paper costs far less than changing working software.</p>
<h2>What a discovery phase is</h2>
<p>UK government services use discovery as the first of four phases, followed by alpha, beta and live, as described in the GOV.UK Service Manual. Studios and agencies use the same word for the work before a build. You may also hear it called a discovery sprint, an inception phase or a scoping phase.</p>
<p>Discovery is a few weeks of interviews, sketching, prototype testing and technical checks, done with the people who will use and pay for the product. Its outputs are short and practical, written so that a team can build from them.</p>
<h2>What happens during discovery</h2>
<ol><li><b>Kick-off.</b> Agree the goal, the decision the discovery will support, the budget range and who decides.</li><li><b>Interviews and observation.</b> Talk to the people who will use the product, and watch the current process where you can. For construction software, that can mean standing on site while someone fills in a paper diary.</li><li><b>Map the current process.</b> Gather the spreadsheets, forms and emails people use today. They show the real data and the real exceptions.</li><li><b>Define users and priorities.</b> List the types of user and what each needs to do, then sort features into must have, should have, could have and won’t have for now, a method often called MoSCoW.</li><li><b>Prototype and test.</b> Build a clickable prototype of the main journeys and put it in front of real users.</li><li><b>Technical investigation.</b> Check integrations, data migration, hosting, security, and constraints such as working without signal (<a class="link" href="https://flavoureak.co.uk/blog/offline-first-mobile/">offline-first mobile</a>) or rules for children’s data.</li><li><b>Estimate and plan.</b> Turn the agreed scope into an estimate, a release plan and a risk list.</li></ol>
<h2>What you should get at the end</h2>
<p>A discovery phase should leave you with documents you own and could hand to any competent team.</p>
<table><thead><tr><th>Output</th><th>What it contains</th><th>Why it matters</th></tr></thead><tbody><tr><td>Scope</td><td>Users, features for the first release, what is out of scope, acceptance criteria</td><td>Sets out what you are paying for</td></tr><tr><td>Prototype</td><td>Clickable screens for the main journeys, tested with users</td><td>Shows the product before it is built</td></tr><tr><td>Technical approach</td><td>Platforms, architecture, hosting, integrations and an outline data model</td><td>Exposes hard problems early</td></tr><tr><td>Estimate</td><td>A range of cost and time for the first release, with its assumptions</td><td>Lets you set a budget and choose a contract model</td></tr><tr><td>Risk list</td><td>Each risk, its likely impact and how to reduce it</td><td>Shows where the estimate could move</td></tr><tr><td>Release plan</td><td>What ships first and what follows</td><td>Gets something useful live sooner</td></tr></tbody></table>
<p>The estimate should be a range with its assumptions written out. A single precise figure at this stage hides uncertainty that still exists.</p>
<p>A tested prototype is useful beyond the build, too. You can show it to investors, use it in early sales conversations, or put it in front of a few target customers to check they would pay.</p>
<h2>How long discovery takes</h2>
<p>For a new product, discovery usually takes a few weeks. A small product with one type of user and no integrations might need a week or two. A platform with several types of user, payments, integrations with existing systems or regulated data can take six weeks or more. The things that change the length most are these.</p>
<ul><li>The number of user types and journeys.</li><li>Integrations, especially with older systems that have little documentation.</li><li>Regulation, such as UK GDPR, the ICO’s Children’s code for services children are likely to use, or CIS and VAT rules in construction.</li><li>How quickly decision makers and users are available.</li></ul>
<p>A discovery that is short for the size of the product still produces a useful estimate, with wider ranges and more assumptions. That is fine, as long as everyone reads the estimate that way.</p>
<h2>Why discovery saves money</h2>
<p>The later a wrong assumption is found, the more it costs to fix. A changed decision during discovery means redrawing a screen. The same change after launch can mean migrating data, rewriting features and retraining users.</p>
<p>Discovery finds the expensive surprises while they are still cheap. Typical examples are an accounting system without the API the product depends on, users who work underground with no signal, a feature the founder assumed was essential that users don’t want, and a requirement that brings extra legal obligations. Estimark works offline on site, saving to the phone first and syncing later. A requirement like that shapes the whole architecture, so it has to be known before the build starts. CurioKids takes bookings in each family’s own time zone, which is straightforward to plan for from the start and fiddly to add to a booking system that assumed a single time zone.</p>
<p>Discovery also gives you a clean point to decide. You can go ahead, cut the scope to fit the budget, or stop, having spent a small part of what the build would cost. A defined scope also makes fixed-price quotes possible (<a class="link" href="https://flavoureak.co.uk/blog/fixed-price-or-time-and-materials/">fixed price or time and materials</a>).</p>
<h2>How to get the most from discovery</h2>
<ul><li>Make the person who will make decisions available every week.</li><li>Bring real material, such as spreadsheets, forms, quotes, invoices and emails.</li><li>Arrange access to the people who do the work day to day, as well as their managers.</li><li>Share your budget range, so the scope can be shaped to it.</li><li>Say what you already know is fixed, such as a launch date or a system you must keep.</li><li>Agree how outputs will be shared, so you can review the scope and prototype as they develop and comment early.</li></ul>
<h2>Who owns the discovery outputs</h2>
<p>Make sure the contract gives you ownership of everything discovery produces, including the prototype, documents and designs. Then you can take them to any supplier, use them to raise money, or build the product in-house. <a class="link" href="https://flavoureak.co.uk/blog/who-owns-the-code-uk/">Who owns the code</a> covers how ownership of software and designs works in the UK.</p>
<div class="callout"><b>Have an idea to test before you build?</b> <a href="https://flavoureak.co.uk/founders/">See how we scope an MVP</a>.</div>
<h2>Questions people ask</h2>
<h3>Is a discovery phase the same as a design sprint?</h3><p>A design sprint is a specific five-day format for exploring and testing one idea with a prototype. A discovery phase is broader, covering users, scope, technical approach, estimate and risks. A discovery phase may include a design sprint as one part of it.</p>
<h3>Can I skip discovery?</h3><p>You can, if the product is small and well understood, or if you already have a tested prototype and a detailed specification. Skipping it on a larger or less certain project usually moves the same questions into the build, where answering them costs more.</p>
<h3>Who should take part in a discovery phase?</h3><p>On your side, the person who makes decisions about the product, someone who knows the current process in detail, and a few future users. On the supplier’s side, usually a product or delivery lead, a designer and a senior developer, so that design and technical questions are answered together.</p>
<h3>Is a discovery phase usually fixed price?</h3><p>Usually, yes. Discovery has a defined length and a defined set of outputs, which suits a fixed price. The build that follows may then be priced as fixed price or as time and materials, depending on how settled the scope is.</p>
<h3>What happens if discovery shows the idea won’t work?</h3><p>Then discovery has done its job at the lowest possible cost. The outputs often point to a smaller or different product that would work, or show that the idea needs more testing with customers before anything is built. Either way, you keep the research, prototype and documents.</p>]]></content:encoded>
</item>
<item>
<title>Single sign-on for SaaS: SAML vs OpenID Connect, SCIM and enforced SSO</title>
<link>https://flavoureak.co.uk/blog/sso-for-saas/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/sso-for-saas/</guid>
<pubDate>Wed, 04 Mar 2026 08:00:00 GMT</pubDate>
<category>SaaS</category>
<description>Enterprise buyers expect staff to sign in with their own identity provider. Support SAML 2.0 and OpenID Connect, add SCIM provisioning, and tie every SSO connection to one tenant.</description>
<content:encoded><![CDATA[<ul><li>Single sign-on (SSO) lets a customer’s staff sign in to your SaaS with the work account they already have, managed in an identity provider such as Microsoft Entra ID, Okta or Google Workspace.</li><li>SAML 2.0 is the older XML-based standard that many enterprise IT teams still ask for, and OpenID Connect is the newer standard built on OAuth 2.0 and JSON tokens; many SaaS products support both.</li><li>Enterprise buyers usually ask for SSO, SCIM provisioning so accounts are created and removed automatically, and the option to enforce SSO so staff cannot sign in with a password.</li></ul><p>Single sign-on (SSO) lets people sign in to your SaaS using the account their employer already manages, instead of a separate username and password. For a B2B product it means supporting SAML 2.0 and OpenID Connect, the two standards identity providers use, and linking each customer’s identity provider to that customer’s tenant. Larger buyers will also ask for SCIM provisioning and the option to make SSO the only way in.</p>
<h2>How single sign-on works</h2>
<p>Two parties are involved. The identity provider (IdP) holds the customer’s staff accounts. Examples include Microsoft Entra ID, Okta, Google Workspace and JumpCloud. The service provider is your product. The IdP proves who the person is, and your product trusts that proof because the two were connected in advance.</p>
<ol><li>The user enters their work email address on your sign-in page.</li><li>Your product finds the SSO connection for that email domain and sends the browser to the customer’s IdP.</li><li>The user signs in there, including any multi-factor authentication their company requires.</li><li>The IdP sends the browser back with a signed statement of who the user is.</li><li>Your product checks the signature, the intended audience and the expiry time, then finds or creates the user in that tenant and starts a session.</li></ol>
<p>Your product never sees the user’s password. When the company disables the person’s account in its IdP, they can no longer start new sessions in your product.</p>
<h2>SAML 2.0 and OpenID Connect compared</h2>
<p>Both standards do the same job. SAML 2.0 became a standard in 2005 and is still common across enterprise IT. OpenID Connect, finalised in 2014, is an identity layer on top of OAuth 2.0 and is the usual choice for newer services and for mobile apps.</p>
<table><thead><tr><th></th><th>SAML 2.0</th><th>OpenID Connect</th></tr></thead><tbody><tr><td>Message format</td><td>Signed XML assertions</td><td>Signed JSON Web Tokens, called ID tokens</td></tr><tr><td>Built on</td><td>Its own XML protocol</td><td>OAuth 2.0</td></tr><tr><td>Setup</td><td>Exchange metadata containing URLs and certificates</td><td>Client ID, client secret and the IdP’s issuer URL</td></tr><tr><td>Typical use</td><td>Enterprise web applications</td><td>Newer IdPs, web and mobile apps</td></tr><tr><td>Key rotation</td><td>Often manual, by swapping certificates</td><td>Usually automatic, by fetching the IdP’s published keys</td></tr><tr><td>Where bugs tend to appear</td><td>XML signature validation</td><td>Token checks and redirect URL rules</td></tr></tbody></table>
<p>The main enterprise IdPs support both, so build the one your first enterprise customers ask for and add the other later. Whichever you choose, use a well-maintained library. XML signature validation in particular has a long history of serious security bugs, and it is not code to write from scratch.</p>
<h2>What enterprise buyers ask for</h2>
<p>A security team reviewing your product rarely stops at “do you support SSO”. Expect these requests to arrive together.</p>
<h3>SSO with their own identity provider</h3>
<p>The customer’s admin should be able to set up the connection from a settings screen, by uploading SAML metadata or entering OpenID Connect details, and test it before switching it on for everyone.</p>
<h3>SCIM provisioning</h3>
<p>SCIM (System for Cross-domain Identity Management) is a standard REST API that lets the IdP create, update and deactivate users and groups in your product. Its main value is offboarding. When someone leaves and IT disables them centrally, SCIM removes their access to your product as well, without anyone having to remember.</p>
<h3>Enforced SSO</h3>
<p>The customer can require everyone with their email domain to sign in through SSO, with password sign-in switched off. Keep a break-glass route, such as one owner who can still sign in another way, so a broken IdP configuration doesn’t lock the whole company out.</p>
<h3>Just-in-time provisioning</h3>
<p>Just-in-time provisioning creates an account the first time someone signs in through SSO, using the name and email address the IdP sends. It is simpler than SCIM and enough for many smaller customers. It can’t remove anyone, though, because a person who has left never signs in again to trigger it.</p>
<h3>Group-to-role mapping</h3>
<p>Groups in the IdP, such as Finance or Site managers, map to roles in your product, so access follows the customer’s own structure. This depends on having a clear model of <a class="link" href="https://flavoureak.co.uk/blog/saas-roles-permissions/">roles and permissions</a> first.</p>
<h2>Build it or use an identity service</h2>
<p>There are three routes. You can add SAML and OpenID Connect to your own authentication using open-source libraries. You can run an open-source identity server such as Keycloak. Or you can use a hosted identity service such as Auth0, WorkOS, AWS Cognito or Microsoft Entra External ID, which handles the protocol work and often gives your customers’ admins a ready-made setup screen.</p>
<p>Hosted services save a lot of protocol work and keep up with the quirks of different IdPs. They also add a supplier to your sub-processor list and a running cost, often charged per SSO connection or per active user, so check that cost against your plan prices. Building it yourself makes sense when authentication is central to your product or you have specific data residency needs. For most early SaaS products we’d use a hosted service or a mature library, and spend the time saved on provisioning and permissions.</p>
<p>Whichever route you take, test against the identity providers your customers actually use. Microsoft Entra ID, Okta and Google Workspace each name and format user attributes and groups in their own way, and those differences tend to surface during the first enterprise rollout.</p>
<h2>How SSO fits multi-tenancy</h2>
<p>In a multi-tenant product, every SSO connection belongs to exactly one tenant (<a class="link" href="https://flavoureak.co.uk/blog/multi-tenancy-explained/">multi-tenancy explained</a>). That rule drives several design decisions.</p>
<ul><li><b>Verify domains.</b> Before a tenant can route an email domain to its IdP, it proves it owns the domain, usually by adding a DNS TXT record. Otherwise one customer could claim another company’s staff.</li><li><b>Never link users across tenants on email alone.</b> A sign-in from tenant A’s IdP only signs the user into tenant A.</li><li><b>Allow people to belong to more than one tenant.</b> A consultant might use SSO for their employer’s account and a password or email link for a client’s.</li><li><b>Store connection settings per tenant</b>, including certificates and their expiry dates, and warn admins well before a certificate expires.</li><li><b>End sessions when access ends.</b> Keep sessions reasonably short, and when SCIM deactivates a user, revoke their sessions and API tokens too.</li></ul>
<h2>Common mistakes with SSO</h2>
<p>Most SSO problems come from shortcuts in validation and in linking accounts. These are the ones to check for in review.</p>
<ul><li>Accepting an assertion without checking its signature, issuer, audience and time window.</li><li>Allowing no tolerance for small clock differences between servers, so sign-ins fail at random.</li><li>Linking an SSO identity to an existing password account without proof that the same person controls both.</li><li>Leaving password sign-in open after the customer has asked to enforce SSO.</li><li>Not recording changes to SSO settings, which are among the most sensitive actions in the product, in the audit log.</li></ul>
<div class="callout"><b>Selling to larger organisations?</b> <a href="https://flavoureak.co.uk/saas/">See how we build SaaS platforms</a> with sign-in, provisioning and tenant isolation planned from the start.</div>
<h2>Questions people ask</h2>
<h3>Is SAML or OpenID Connect more secure?</h3><p>Both are secure when implemented with a maintained library, signed messages and strict validation. Real-world SSO flaws usually come from implementation mistakes, such as skipped signature checks or loose redirect URL rules. Choose the standard your customers’ identity providers and IT teams prefer.</p>
<h3>Does SSO replace multi-factor authentication?</h3><p>SSO moves authentication to the customer’s identity provider, which is where multi-factor authentication is then applied for those users. Your product should still offer its own multi-factor authentication for people who sign in with a password, such as small customers without an identity provider.</p>
<h3>What is the difference between SSO and SCIM?</h3><p>SSO handles signing in. SCIM handles the user list, creating, updating and deactivating accounts in your product when they change in the customer’s identity provider. Without SCIM, a person removed from the IdP can’t sign in again, but their account, open sessions and API tokens may remain in your product.</p>
<h3>Should SSO be a paid feature?</h3><p>Many B2B SaaS products include SSO only in higher tiers, because the customers who need it are larger and it carries support work. Others include it on every plan to encourage good security. Either is reasonable if it is a deliberate choice that fits your <a class="link" href="https://flavoureak.co.uk/blog/saas-pricing-models/">pricing model</a>.</p>]]></content:encoded>
</item>
<item>
<title>Headless commerce explained: when it helps and when it is overkill</title>
<link>https://flavoureak.co.uk/blog/headless-commerce-explained/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/headless-commerce-explained/</guid>
<pubDate>Fri, 27 Feb 2026 08:00:00 GMT</pubDate>
<category>E-commerce</category>
<description>Headless commerce separates the storefront customers see from the engine that runs products, carts and orders. It suits custom experiences and many channels, and costs more to build and run.</description>
<content:encoded><![CDATA[<ul><li>Headless commerce is an architecture in which the storefront is built separately from the commerce engine, with the two connected through APIs.</li><li>It helps when you need a shopping experience a theme cannot deliver, sell through several channels from one catalogue, or want full control of the front end.</li><li>For a standard shop with one website, a themed platform is usually cheaper to build and run, because headless makes you responsible for the storefront, its hosting and its SEO.</li></ul><p>Headless commerce is an architecture in which the storefront, the part customers see and click, is built and hosted separately from the commerce engine that holds products, prices, carts, checkout and orders. The two communicate through APIs, so one engine can serve a website, an app, an in-store screen and a trade portal at the same time. It is worth the extra cost when the shopping experience or the number of channels outgrows what a themed store can do, and it is overkill for most shops that sell a standard catalogue through one website.</p>
<h2>What headless means</h2>
<p>In a traditional e-commerce platform, the engine and the storefront come as one product. You choose a theme, customise it with the platform’s templates and the platform renders the pages. The storefront is called the “head”. Going headless means removing it and building your own, which then asks the engine’s APIs for everything it needs.</p>
<p>A typical journey through a headless shop works like this.</p>
<ol>
<li>A customer opens a product page on your storefront, often a JavaScript framework application hosted on a platform such as Vercel, Netlify or Cloudflare.</li>
<li>The storefront asks the commerce engine’s API for the product, its variants, its price and its stock.</li>
<li>It asks a content management system (CMS) for the editorial content around the product, such as the story, the size guide and the lifestyle images.</li>
<li>When the customer adds the product to their basket, the storefront calls the engine’s cart API.</li>
<li>At checkout, the customer is often handed to the engine’s own hosted checkout, which takes payment and keeps card data away from your code.</li>
</ol>
<p>A related term is composable commerce, which means assembling the whole stack from separate specialist services, such as search, content, payments, reviews and the commerce engine, each connected by APIs. Composable set-ups are almost always headless, but a headless shop can still use a single engine for everything behind the storefront.</p>
<h2>Commerce engines that support headless</h2>
<p>Most modern engines offer APIs suitable for a headless build. They differ in how much they include and who runs them.</p>
<ul>
<li><b>Established platforms with headless options</b>, such as Shopify, BigCommerce, Adobe Commerce and Salesforce Commerce Cloud. Shopify, for example, offers a Storefront API and its own React-based framework, Hydrogen.</li>
<li><b>API-first engines</b> designed for headless from the start, such as commercetools.</li>
<li><b>Open-source engines</b> that you host yourself, such as Medusa, Saleor and Vendure.</li>
</ul>
<p>The right engine depends more on your catalogue, pricing rules and integrations than on the storefront, so choose the engine first and the front-end framework second.</p>
<h2>When headless helps</h2>
<h3>Custom shopping experiences</h3>
<p>Product configurators, made-to-measure ordering, editorial content woven through product pages, or a shop that is one part of a larger web product all push against theme systems. When the theme gets in the way of the experience you need, owning the front end removes the limit.</p>
<h3>Several channels from one catalogue</h3>
<p>A business selling through a website, a mobile app, in-store screens and a trade portal can run them all from one engine, with one set of products, prices and stock levels. Each channel is its own front end, built for its own users. Our article on <a class="link" href="https://flavoureak.co.uk/blog/b2b-ordering-portals/">B2B ordering portals</a> covers one common second channel.</p>
<h3>Performance and control</h3>
<p>A headless storefront can pre-render pages and serve them from servers close to the customer, which helps page speed and Core Web Vitals. A well-built theme can also be fast, though, so speed alone is rarely a strong enough reason. The case is stronger when you also need the control for one of the reasons above.</p>
<h2>When it is overkill</h2>
<p>Headless adds a second system to build, host, monitor and keep up to date. It is likely to be the wrong choice when:</p>
<ul>
<li>the catalogue is standard and a good theme already does what you need;</li>
<li>there is no developer available to maintain the storefront after launch;</li>
<li>the business relies on apps or plugins that add features to theme pages, many of which do not work on a headless front end;</li>
<li>marketing needs to change pages without a developer and there is no budget for a CMS with preview;</li>
<li>the real goal is a redesign, which a new or customised theme can deliver for much less.</li>
</ul>
<h2>Cost and team implications</h2>
<p>Going headless moves cost around more than it removes it. The table compares where effort goes.</p>
<table><thead><tr><th>Cost driver</th><th>Themed store</th><th>Headless store</th></tr></thead><tbody>
<tr><td>Initial build</td><td>Theme purchase and customisation</td><td>A full front-end application, plus integration with the engine and a CMS</td></tr>
<tr><td>Hosting</td><td>Included in the platform</td><td>Separate front-end hosting, plus the engine</td></tr>
<tr><td>Content editing</td><td>The platform’s built-in editor</td><td>A headless CMS, with page preview to set up</td></tr>
<tr><td>Apps and plugins</td><td>Installed from the platform’s app store</td><td>Often need custom integration work</td></tr>
<tr><td>SEO</td><td>Mostly handled by the platform</td><td>Your job: rendering, metadata, sitemaps and redirects</td></tr>
<tr><td>Ongoing work</td><td>Platform updates and theme changes</td><td>Framework upgrades, security patches and monitoring</td></tr>
<tr><td>Team</td><td>Merchandisers and an occasional developer</td><td>Front-end developers on call, and someone who owns the integrations</td></tr>
</tbody></table>
<p>SEO needs particular care. A storefront that only builds product pages in the browser can be slow to index, so render pages on the server or at build time. Our guide to <a class="link" href="https://flavoureak.co.uk/blog/seo-for-javascript-sites/">SEO for JavaScript sites</a> covers how. When you move an existing shop to headless, keep the same URLs where you can and redirect any that change, or you risk losing the search traffic the old shop had earned.</p>
<p>Caching is the other hidden decision. Pre-rendered pages are fast because they are built in advance, but prices and stock change during the day. Decide for each part of a page how long it can be cached, and fetch the parts that must be live, such as stock levels or a trade customer’s own prices, when the page is requested. Getting this wrong shows customers products they cannot buy.</p>
<h2>A middle path</h2>
<p>Headless does not have to be all or nothing. Several half-steps get most of the benefit for less cost:</p>
<ul>
<li>keep the themed store and build one custom part, such as a configurator, as an embedded app;</li>
<li>run a separate content site for guides and campaigns that links into the themed shop;</li>
<li>use the engine headless for a second channel, such as an app or a trade portal, while the main website stays themed.</li>
</ul>
<p>We compare these routes on one popular platform in <a class="link" href="https://flavoureak.co.uk/blog/shopify-or-custom/">Shopify or a custom build</a>.</p>
<div class="callout"><b>Weighing up a headless build?</b> <a href="https://flavoureak.co.uk/ecommerce/">See how we approach e-commerce projects</a>.</div>
<h2>Questions people ask</h2>
<h3>What is headless commerce?</h3>
<p>Headless commerce is an architecture where the customer-facing storefront is built separately from the commerce engine that manages products, carts, checkout and orders. The two connect through APIs. It lets one engine serve several front ends, such as a website, an app and a trade portal.</p>
<h3>Is headless commerce better for SEO?</h3>
<p>Not automatically. A headless storefront can be very fast, but it also makes you responsible for rendering, metadata, sitemaps and redirects that a themed platform handles for you. Pages should be rendered on the server or at build time so search engines can read them easily.</p>
<h3>What is the difference between headless and composable commerce?</h3>
<p>Headless commerce separates the storefront from the commerce engine. Composable commerce goes further and builds the whole stack from separate specialist services, such as search, content, payments and the engine, each connected through APIs. A composable set-up is almost always headless, while a headless shop may still use one engine for everything behind the storefront.</p>
<h3>Is headless commerce worth it for a small business?</h3>
<p>Usually not. A small business with a standard catalogue and one website will normally get more from a well-built theme, at lower build and running costs. Headless becomes worth considering when the experience or the number of channels outgrows what the theme can do.</p>
<h3>Can you build a headless store on Shopify?</h3>
<p>Yes. Shopify offers a Storefront API and a React-based framework called Hydrogen for building custom storefronts, while Shopify still runs the catalogue, checkout and orders. Many theme apps will not work on a headless storefront, so check which features you rely on first.</p>]]></content:encoded>
</item>
<item>
<title>Building an online tutoring platform: features, safeguarding and payments</title>
<link>https://flavoureak.co.uk/blog/building-a-tutoring-platform/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/building-a-tutoring-platform/</guid>
<pubDate>Sat, 21 Feb 2026 08:00:00 GMT</pubDate>
<category>Education</category>
<description>A tutoring platform vets tutors, matches them with learners, schedules lessons across time zones, runs them live and pays tutors. When learners are children, safeguarding shapes every feature.</description>
<content:encoded><![CDATA[<ul><li>An online tutoring platform needs tutor vetting, matching, time-zone-aware scheduling, a live classroom, payments and payouts, reviews and, for children, parent accounts.</li><li>In England and Wales, criminal record checks for tutors come from the Disclosure and Barring Service, while Disclosure Scotland and AccessNI cover Scotland and Northern Ireland.</li><li>Store each lesson as an exact time plus the time zone it was booked in, and calculate repeat lessons in that zone, so clock changes do not move lessons by an hour.</li></ul><p>An online tutoring platform is software that connects tutors with learners, books lessons, runs them live over video and moves the money between the two. To work, it needs tutor profiles and vetting, matching, scheduling across time zones, a live classroom, payments and payouts, and reviews. When the learners are children, it also needs parent accounts and safeguarding built into each of those parts.</p>
<h2>Tutor profiles and vetting</h2>
<p>A tutor profile is what a family uses to choose, so it should show subjects, levels, availability, languages, a short introduction video and verified qualifications. The vetting behind the profile is what lets a family trust it. A practical vetting process has seven steps.</p>
<ol>
<li>Identity check against photo ID, ideally with a liveness check, so the person on camera is the person on the documents.</li>
<li>Right to work checks, where you employ tutors in the UK.</li>
<li>A criminal record check. In England and Wales this comes from the Disclosure and Barring Service (DBS). Scotland uses Disclosure Scotland and Northern Ireland uses AccessNI.</li>
<li>References, preferably from previous teaching or childcare work.</li>
<li>Qualification checks, such as a teaching qualification or degree certificate.</li>
<li>An interview or recorded trial lesson, reviewed by a person.</li>
<li>Safeguarding training, with a record of when it was completed.</li>
</ol>
<p>The level of DBS check depends on the role. An enhanced check with a check of the children’s barred list is only available for eligible roles, such as regulated activity with children, and whether a tutoring role qualifies depends on what the tutor does and how often. Individuals cannot request an enhanced check for themselves. It has to come through a registered body or an umbrella body acting for the organisation. Tutors can join the DBS Update Service, which lets an organisation check online, with the tutor’s consent, that a certificate is still current. DBS checks do not cover records held abroad, so tutors who have lived overseas need certificates from those countries too. Eligibility is a legal question, so take advice on how it applies to your platform.</p>
<h2>Matching tutors to learners</h2>
<p>Matching can be done by the family browsing profiles, by the platform assigning a tutor, or by a mix in which the platform shortlists and the family picks. Browsing suits older learners and adults looking for a specialist. Assigned matching suits younger children, whose parents tend to care more about reliability and fit than about reading profiles. In CurioKids, teachers are matched to children.</p>
<p>A good match uses the signals that predict a good lesson: subject and level, the learner’s age, overlapping availability, language, and for children, interests that give the tutor something to talk about. Keep the matching explainable, so support staff can tell a parent why a tutor was suggested, and make it easy to change tutor without a difficult conversation.</p>
<h2>Scheduling across time zones</h2>
<p>Online tutoring is often international. A tutor in one country teaches a learner in another, and the two countries may change their clocks on different dates or not at all. The classic bug is a weekly lesson that moves by an hour twice a year.</p>
<p>Take a tutor in London and a family in Dubai. The United Arab Emirates stays on UTC+4 all year. The UK is on GMT (UTC+0) in winter and on BST (UTC+1) from the last Sunday in March to the last Sunday in October.</p>
<table><thead><tr><th>Lesson booked as</th><th>Winter, UK on GMT</th><th>Summer, UK on BST</th></tr></thead><tbody>
<tr><td>16:00 London time</td><td>20:00 in Dubai</td><td>19:00 in Dubai</td></tr>
<tr><td>20:00 Dubai time</td><td>16:00 in London</td><td>17:00 in London</td></tr>
</tbody></table>
<p>A weekly lesson can stay fixed for one person only. The software has to decide whose clock the lesson follows, and warn the other person in advance when their local time will change. The rules we build to:</p>
<ul>
<li>Store each lesson as an exact instant in UTC, together with the time zone it was booked in, using IANA names such as <code>Europe/London</code> or <code>Asia/Dubai</code>.</li>
<li>Generate repeat lessons from a rule in the booking time zone, then convert each one, so clock changes are handled lesson by lesson.</li>
<li>Show every time in the viewer’s own time zone, with the zone named on screen.</li>
<li>Send reminders that state the local time of the person receiving them.</li>
</ul>
<p>CurioKids takes bookings in the family’s own time zone, which is the right default when the customer is a parent planning around school and bedtime.</p>
<h2>The live lesson</h2>
<p>The classroom is the product. It needs reliable two-way video and audio, a shared whiteboard both people can draw on, a way to show lesson material and a chat for links and spellings. It should cope with poor connections by lowering video quality before audio suffers. In CurioKids, lessons run on live video with a shared whiteboard, and the classroom opens 15 minutes before the lesson, so a family can get a camera or microphone working before the teacher needs them. The technical choices are covered in <a class="link" href="https://flavoureak.co.uk/blog/live-video-for-lessons/">live video for online lessons</a>.</p>
<p>Recording lessons helps with safeguarding and complaints, but a recording is personal data, often of a child. Decide the purpose, tell families clearly, limit who can watch recordings and delete them on a fixed schedule.</p>
<h2>Payments, payouts and reviews</h2>
<p>Most tutoring platforms take payment from the family and pay the tutor after the lesson, keeping a commission. That makes the platform a marketplace, and holding money on behalf of tutors brings payments regulation into play. That is why most platforms use a payment provider with connected accounts for tutors, as we explain in <a class="link" href="https://flavoureak.co.uk/blog/marketplace-payments-uk/">marketplace payments in the UK</a>. Decide early:</p>
<ul>
<li>whether families pay per lesson, buy packages of lessons or subscribe monthly;</li>
<li>what happens when a family cancels late or a tutor does not turn up;</li>
<li>when tutors are paid, and whether payment waits until the lesson has happened;</li>
<li>whether tutors are self-employed or employed, which has tax and employment law consequences that need advice.</li>
</ul>
<p>Reviews help families choose, but they must be genuine. Under the Digital Markets, Competition and Consumers Act 2024, fake reviews and reviews that conceal an incentive are banned, and businesses that publish reviews must take reasonable steps to prevent and remove fake ones. Only let a family review a tutor after a completed lesson, and explain how ratings are collected.</p>
<h2>Parent accounts and safeguarding</h2>
<p>When learners are children, the parent is the customer and the account holder, while the child uses the classroom. The design has to serve the parent and protect the child. Measures to build in from the start:</p>
<ul>
<li>Parents book, pay and see lesson history and feedback. In CurioKids, parents book, join and follow lessons in an app, and teachers leave feedback after each lesson.</li>
<li>Children use a restricted space. CurioKids has a child mode for homework, opened with the parent’s PIN.</li>
<li>All contact between tutor and family happens on the platform, with no personal phone numbers, email addresses or social media.</li>
<li>Children, parents and tutors have a clear way to report a concern, routed to a named safeguarding lead.</li>
<li>Messages and attendance are logged and kept for a defined period.</li>
</ul>
<p>Because the service is likely to be used by children, the ICO’s Age Appropriate Design Code applies to how it handles their data, including high privacy settings by default and no nudges to share more. We summarise it in <a class="link" href="https://flavoureak.co.uk/blog/childrens-code-for-edtech/">the Children’s code for edtech</a>. In the UK, a child must be 13 or over to give their own consent to an online service where consent is the lawful basis, so younger children need a parent’s consent. The Department for Education also publishes safeguarding guidance for out-of-school settings, including tuition providers.</p>
<div class="callout"><b>Building a tutoring or live lessons product?</b> <a href="https://flavoureak.co.uk/curiokids/">See how CurioKids works</a>.</div>
<h2>Questions people ask</h2>
<h3>Do online tutors need a DBS check?</h3>
<p>There is no general legal requirement for every private tutor to hold a DBS certificate. It is, however, an offence for a barred person to work in regulated activity with children, and for an organisation to knowingly let them. Platforms that work with children should carry out the highest level of check each role is eligible for, and take advice on eligibility.</p>
<h3>How should a tutoring platform handle time zones?</h3>
<p>Store each lesson as an exact time in UTC together with the time zone it was booked in, and calculate repeat lessons in that zone. Show every time in the viewer’s own time zone. Warn families and tutors in advance when a clock change will move their local lesson time.</p>
<h3>How do tutoring platforms make money?</h3>
<p>Most take a commission on each lesson, sell packages of lessons or charge a monthly subscription, and pay tutors their share after lessons take place. Holding money on tutors’ behalf is usually handled by a regulated payment provider with connected accounts for each tutor.</p>
<h3>Can children have their own account on a tutoring platform?</h3>
<p>Children under 13 cannot give their own consent to an online service in the UK where consent is the lawful basis, so the parent should hold the account. A common pattern is a parent account for booking and payment with a limited, PIN-protected space for the child. The ICO’s Children’s code applies to any data the platform holds about the child.</p>]]></content:encoded>
</item>
<item>
<title>Offline event check-in: scanning QR tickets without Wi-Fi</title>
<link>https://flavoureak.co.uk/blog/offline-event-check-in/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/offline-event-check-in/</guid>
<pubDate>Sun, 15 Feb 2026 08:00:00 GMT</pubDate>
<category>Events</category>
<description>Offline check-in works by loading the guest list onto every scanner before doors open, validating QR codes on the device, and syncing scans when a connection returns.</description>
<content:encoded><![CDATA[<ul><li>Offline check-in means scanning and validating tickets on the device itself, using a guest list downloaded before the event, so entry continues without internet.</li><li>Signed QR codes can be verified on a scanner with no connection, while lookup codes need the guest list on the device to check them.</li><li>Preventing double entry across several offline scanners needs a local network between them, separate entrances by ticket type, or reconciliation after the event.</li></ul><p>Offline check-in for events means scanning and validating tickets on the scanning device itself, without a connection to the internet, and syncing the results once a connection is available. It works by preloading the guest list onto every scanner before doors open, checking each QR code against that list or against a digital signature, and recording each scan on the device. The hard part is stopping the same ticket being used twice at two scanners that cannot talk to each other.</p>
<h2>Why check-in needs to work offline</h2>
<p>Venue Wi-Fi is often slow or overloaded once a crowd arrives. Mobile data struggles in basements, concrete halls and fields. And the moment a queue forms is the moment every phone in it is trying to use the same network. A check-in system that needs a round trip to a server for every scan will stall exactly when it matters.</p>
<p>The rules are the same as for any offline-first app, which we cover in <a class="link" href="https://flavoureak.co.uk/blog/offline-first-mobile/">offline first: building a mobile app for a basement with no signal</a>. Save to the device first, sync when possible, and tell the person scanning the truth about what has synced.</p>
<h2>Preloading the guest list</h2>
<p>Before doors open, each scanner downloads the current guest list. That includes ticket IDs, ticket types, names where they are needed for ID checks, and each ticket’s status, such as valid, cancelled, refunded or already checked in. The scanner then validates each code against its local copy.</p>
<ol>
<li>Pause changes to the list as late as practical, and download it to every device.</li>
<li>Check that each device shows the same ticket count as the server.</li>
<li>Download the revocation list, meaning tickets cancelled, refunded or transferred since they were issued.</li>
<li>Keep the guest list encrypted on the device, because it contains personal data, and wipe it after the event.</li>
<li>Refresh the list whenever a connection is available during the event, to pick up late sales and cancellations.</li>
</ol>
<h2>Signed QR codes or lookups</h2>
<p>A QR code on a ticket can work in two ways. A lookup code holds a random ticket ID and nothing else, and the scanner checks it against the guest list. A signed code holds the ticket ID, event and ticket type together with a digital signature created by the ticketing system, and the scanner checks that the signature is genuine.</p>
<table>
<thead><tr><th>Factor</th><th>Lookup code</th><th>Signed code</th></tr></thead>
<tbody>
<tr><td>What the QR holds</td><td>A random ticket ID</td><td>Ticket details plus a cryptographic signature</td></tr>
<tr><td>Needs the guest list on the device</td><td>Yes</td><td>No, though it helps</td></tr>
<tr><td>Stops forged tickets</td><td>Yes, if the ID isn’t on the list</td><td>Yes, if the signature fails</td></tr>
<tr><td>Handles cancellations</td><td>Status in the list</td><td>Needs a revocation list on the device</td></tr>
<tr><td>Stops double entry</td><td>Only on the same device or with sync</td><td>Only on the same device or with sync</td></tr>
<tr><td>Ticket transfers</td><td>Update the list</td><td>Issue a new code and revoke the old one</td></tr>
<tr><td>QR size</td><td>Small and quick to scan</td><td>Larger and denser, so keep the payload short</td></tr>
</tbody>
</table>
<p>When codes are signed with a public-key scheme, each scanner holds only the public key, which can check signatures but cannot create them. A lost or stolen scanner therefore cannot be used to make tickets. Many systems combine the two approaches, with signed codes so that any device can recognise a genuine ticket and a preloaded list for status and names.</p>
<p>Neither approach on its own stops someone sharing a screenshot of a valid ticket. Catching that needs the scan record, which is why double entry is the real problem.</p>
<h2>Preventing double entry across scanners</h2>
<p>Each scanner knows which tickets it has scanned. Without a connection, it doesn’t know what the scanner at the other door has scanned. If a ticket is screenshotted and shared, the first copy can get in at door A and the second at door B. There are four practical ways to close the gap.</p>
<ul>
<li><b>A local network.</b> “No Wi-Fi” usually means no internet. A portable router creates a local network that scanners can use to share scans with each other, or with a laptop acting as a local hub, without any internet connection.</li>
<li><b>Split entrances.</b> Send each ticket type or surname range to one entrance, so each ticket is only valid on one set of devices.</li>
<li><b>Opportunistic sync.</b> Scanners sync whenever any connection appears, even briefly, and pick up other devices’ scans.</li>
<li><b>Reconcile afterwards.</b> For low-risk events, accept the gap, record every scan with a device ID and time, and review duplicates once synced.</li>
</ul>
<p>When a duplicate is detected, the scanner should show when and where the ticket was first scanned, so staff can deal with it calmly. A customer whose ticket was copied and used first needs a different response from someone trying to get a friend in.</p>
<h2>Syncing afterwards</h2>
<p>Every scan is a record with a ticket ID, a device, a time and a result, which is admitted, rejected, duplicate or manual override. When devices reconnect, they upload their scans and download everyone else’s. Because scans are only ever added and never edited, merging them is simple. The earliest scan of each ticket counts as the entry, and later ones are flagged as duplicates to review.</p>
<p>After the event, the synced record gives the attendance figures, a list of no-shows for follow-up and an audit trail for any disputed entry. It also feeds refunds and transfers, which are covered in <a class="link" href="https://flavoureak.co.uk/blog/ticket-refunds-and-transfers/">ticket refunds and transfers</a>.</p>
<h2>Hardware</h2>
<p>Most events can check in with ordinary smartphones running a scanning app. Phone cameras read QR codes well in normal light, and staff already know how to use them. Dedicated scanners, which are rugged handheld devices with a built-in scan engine, are usually quicker with cracked or dim screens, cope better with bright sunlight and survive being dropped.</p>
<ul>
<li>Bring at least one spare scanner per entrance.</li>
<li>Bring power banks or chargers, because scanning with the screen on all evening drains batteries.</li>
<li>Plan for lighting at night and glare in sunshine.</li>
<li>Use lanyards or wrist straps.</li>
<li>Lock devices into the check-in app so they can’t be used for anything else.</li>
<li>Bring a portable router if you plan to sync locally.</li>
</ul>
<h2>Rehearse before doors open</h2>
<p>Test the whole process in the conditions of the event. Office Wi-Fi hides every problem that matters.</p>
<ol>
<li>Put every device in airplane mode and scan a set of test tickets.</li>
<li>Scan the same ticket on two devices and check what each one shows.</li>
<li>Cancel a ticket after download and check it is rejected once the revocation list refreshes.</li>
<li>Scan from a cracked screen, a printed ticket and a phone with low brightness.</li>
<li>Reconnect and confirm that all scans sync and duplicates are flagged.</li>
<li>Time how long each scan takes, and work out how many scanners your arrival pattern needs.</li>
</ol>
<p>The last step is simple arithmetic. If each check-in takes 6 seconds and 1,200 guests arrive in the hour before the start, that is 1,200 × 6 = 7,200 seconds of scanning, or two hours of work in one hour, so you need at least 2 scanning points working without a break. If half of those guests arrive in the last 15 minutes, 600 × 6 = 3,600 seconds of scanning has to fit into 900 seconds, which needs 4 scanners, plus spares.</p>
<div class="callout"><b>Planning event software?</b> We build for the busiest five minutes of the event. See our <a class="link" href="https://flavoureak.co.uk/events/">events sector page</a> or read about <a class="link" href="https://flavoureak.co.uk/blog/event-software-on-sale-day/">event software on sale day</a>.</div>
<h2>Questions people ask</h2>
<h3>Can you scan QR tickets without internet?</h3>
<p>QR tickets can be scanned without internet if the scanning app validates them on the device. The app either checks each code against a guest list downloaded beforehand or verifies a digital signature in the code. Scans are stored on the device and uploaded when a connection returns.</p>
<h3>How do you stop the same ticket being used twice at different entrances?</h3>
<p>Scanners at different entrances need to share their scans to catch a ticket used twice. Without internet, a portable router can create a local network for this. Alternatively, each ticket type can be assigned to one entrance, or duplicates can be reviewed after the event once scans have synced.</p>
<h3>Do I need dedicated scanners or will phones do?</h3>
<p>Smartphones with a scanning app are enough for most events. Dedicated handheld scanners are usually quicker with poor screens and bright light, are more durable and suit very large or outdoor events. Whichever you use, bring spares and power.</p>
<h3>What happens to tickets sold after the guest list was downloaded?</h3>
<p>Tickets sold after the download are not on a device’s list until it refreshes. Signed QR codes can still be recognised as genuine without the list. Otherwise, refresh the list whenever a connection is available, or send late purchases to a help desk for a manual check.</p>]]></content:encoded>
</item>
<item>
<title>What a construction site diary should record, and why it matters</title>
<link>https://flavoureak.co.uk/blog/site-diary-what-to-record/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/site-diary-what-to-record/</guid>
<pubDate>Mon, 09 Feb 2026 08:00:00 GMT</pubDate>
<category>Construction</category>
<description>A site diary is the daily record of what happened on a job. Record people, work, weather, deliveries, instructions, delays and incidents, with photos, on the day itself.</description>
<content:encoded><![CDATA[<ul><li>A construction site diary is a dated daily record of who was on site, what work was done, the conditions, deliveries, instructions, delays and incidents.</li><li>Diary entries made on the day carry more weight in disputes, delay claims and payment discussions than records written up later from memory.</li><li>A site diary app must work without signal, because the people making entries are often in places with no connection.</li></ul><p>A construction site diary should record, for every day on site, the date and weather, who was on site and which trades, the work done, deliveries and plant, visitors and inspections, instructions received, delays and their causes, and any incidents, supported by photos. A site diary is a dated, daily record of what happened on a construction job, made at the time by the people running it. It matters because in a dispute about delay, cost or payment, the record made on the day is usually the most reliable evidence anyone has.</p>
<h2>Why the site diary matters</h2>
<p>Most construction disputes are about what happened, when, and who caused it. Months later, memories differ and emails tell only part of the story. A diary written on the day fills the gaps. It shows that the scaffold arrived two days late, that the client’s representative asked for a change on a Tuesday, and that rain stopped the pour on Thursday.</p>
<p>Contemporaneous records, meaning records made at the time of the events they describe, are given more weight than accounts pieced together later. The Society of Construction Law’s Delay and Disruption Protocol puts heavy emphasis on keeping good records for this reason. Contracts depend on them too. Under JCT contracts, exceptionally adverse weather is a Relevant Event that can entitle the contractor to more time. Under NEC contracts, a contractor that does not notify a compensation event within eight weeks of becoming aware of it can, in many cases, lose the right to extra time and money. In both, the diary is where the facts behind the notice come from.</p>
<h2>What to record every day</h2>
<table>
<thead><tr><th>Entry</th><th>What to write</th><th>Why it matters</th></tr></thead>
<tbody>
<tr><td>Date and times</td><td>Date, site open and close times, who made the entry</td><td>Ties every other fact to a day</td></tr>
<tr><td>Weather</td><td>Conditions morning and afternoon, rain, wind, temperature where it affects work</td><td>Supports weather-related delay claims and explains stopped work</td></tr>
<tr><td>People and trades</td><td>Own operatives, each subcontractor and their headcount</td><td>Shows resources on site and supports labour costs</td></tr>
<tr><td>Work done</td><td>Activities by area or plot, what was started and finished</td><td>Shows progress against the programme and supports valuations</td></tr>
<tr><td>Deliveries</td><td>Supplier, materials, quantity, time, any shortages or damage</td><td>Records late or short deliveries</td></tr>
<tr><td>Plant</td><td>Plant on site, hired in or off-hired, breakdowns</td><td>Supports plant costs and explains lost time</td></tr>
<tr><td>Visitors and inspections</td><td>Client, designers, building control, warranty inspectors, HSE</td><td>Records sign-offs, comments and who saw what</td></tr>
<tr><td>Instructions</td><td>Who gave the instruction, when, what it was, whether it was in writing</td><td>Starting point for variations and changes to the price</td></tr>
<tr><td>Delays and causes</td><td>What stopped or slowed work, for how long, and why</td><td>The core evidence for extensions of time</td></tr>
<tr><td>Incidents</td><td>Accidents, near misses, damage, security issues</td><td>Links to safety records and insurance</td></tr>
<tr><td>Photos</td><td>Progress, problems, deliveries, work before it is covered up</td><td>Visual proof with a date and time</td></tr>
</tbody>
</table>
<h2>How to write a useful entry</h2>
<p>A diary entry should be factual, specific and short. Write down what happened and leave out opinions about it. “Electrician not on site, second day running; first fix to flats 3 and 4 on hold” is useful. “Electricians messing us about again” tells nobody anything, and it reads badly if the diary is ever disclosed in a dispute.</p>
<ul>
<li><b>Write it on the day.</b> An entry written a week later from memory is worth much less.</li>
<li><b>Be specific about place and quantity.</b> Name the plot, floor, room or gridline, and give numbers.</li>
<li><b>Record the cause of each delay as well as its length.</b> “No work to roof, 3 hours, waiting for revised truss drawings from the engineer” tells the whole story.</li>
<li><b>Note verbal instructions straight away</b>, with the name of the person who gave them, and confirm them in writing the same day.</li>
<li><b>Photograph work before it is covered.</b> Drainage, insulation, fire stopping and reinforcement are hard to prove once hidden.</li>
<li><b>Leave old entries alone.</b> Add a dated correction instead of editing, so the original stays visible.</li>
</ul>
<h2>How the diary supports payment and valuations</h2>
<p>A valuation, or application for payment, claims money for the work done up to a date. The diary shows what was done and when, and the photos show it in place. When a quantity surveyor questions how far the work had progressed at the valuation date, a dated entry with photos of the finished first fix settles it faster than an argument.</p>
<p>Variations follow the same path. Most disputes about extra work come down to whether the client asked for it. A diary that records the instruction, followed by a written confirmation sent the same day, turns a disagreement into a record. The guide to <a class="link" href="https://flavoureak.co.uk/blog/job-costing-for-builders/">job costing for builders</a> shows how variations feed into the margin on a job.</p>
<h2>Incidents and health and safety</h2>
<p>The site diary sits alongside the health and safety records and doesn’t replace them. It should mention that an incident happened and say where the full record is. Accidents go in the accident book, and those that are reportable under RIDDOR 2013 must be reported to the HSE. Toolbox talks, inductions and permits have their own records too. A short diary line such as “Near miss 10.40, unsecured board fell from scaffold lift 2, no injury, reported to site manager, see incident report 14” connects them.</p>
<h2>Paper or app</h2>
<table>
<thead><tr><th>Factor</th><th>Paper diary</th><th>Site diary app</th></tr></thead>
<tbody>
<tr><td>Cost to start</td><td>Low</td><td>Subscription or build cost</td></tr>
<tr><td>Photos</td><td>Separate, often on personal phones</td><td>Attached to the entry with date and time</td></tr>
<tr><td>Legibility</td><td>Depends on handwriting</td><td>Typed and searchable</td></tr>
<tr><td>Office visibility</td><td>When the book comes back</td><td>As soon as it syncs</td></tr>
<tr><td>Risk of loss</td><td>Rain, fire, a stolen van</td><td>Backed up once synced</td></tr>
<tr><td>Prompts</td><td>None</td><td>Fields for weather, trades, deliveries and delays</td></tr>
<tr><td>Searching later</td><td>Reading every page</td><td>Search by date, trade or keyword</td></tr>
</tbody>
</table>
<p>A paper diary is far better than no diary, and some people will always prefer a pen. An app is stronger where the diary needs to be read in the office, searched in a dispute or linked to the rest of the job, such as the programme, variations, valuations and photos.</p>
<h2>Why a site diary app must work offline</h2>
<p>Site diaries are written in basements, plant rooms, rural plots and steel-framed buildings, where mobile signal is weak or absent. If the app needs a connection to save, people stop using it, and the diary ends up on the back of a delivery note. A site diary app should save every entry and photo to the phone first, show clearly what has and hasn’t synced, and upload when a connection returns.</p>
<p>In Estimark, site diaries sit in the same record as the job’s estimate, schedule and health and safety paperwork, and the app saves to the phone first and syncs later. We wrote up the engineering behind that in <a class="link" href="https://flavoureak.co.uk/blog/offline-first-mobile/">offline first: building a mobile app for a basement with no signal</a>.</p>
<div class="callout"><b>Running jobs from paper and group chats?</b> See how <a class="link" href="https://flavoureak.co.uk/estimark/">Estimark</a> keeps the diary, the job and the paperwork in one record, or read about our <a class="link" href="https://flavoureak.co.uk/construction/">construction software work</a>.</div>
<h2>Questions people ask</h2>
<h3>Is a site diary a legal requirement?</h3>
<p>There is no general legal requirement in the UK to keep a site diary. Some contracts and clients require one, and many contract procedures depend on records of the kind a diary holds. In practice, a contractor without a diary is at a disadvantage in any dispute about delay or payment.</p>
<h3>Who should fill in the site diary?</h3>
<p>The site manager or the person running the job day to day should fill in the site diary, so one person is accountable for it. Other people can add photos and notes, as long as each entry shows who made it. On small jobs this is often the builder or the lead tradesperson.</p>
<h3>How long should I keep site diaries?</h3>
<p>Keep site diaries at least as long as a claim could be brought about the work. Under the Limitation Act 1980 that is generally six years for a simple contract and twelve years for a contract executed as a deed, and work on homes can face longer periods under the Defective Premises Act 1972 as extended by the Building Safety Act 2022. Take legal advice on the right period for your work.</p>
<h3>Are photos in a site diary good evidence?</h3>
<p>Photos are strong evidence when they are clearly dated, show where they were taken and are linked to a diary entry describing them. Photos stored loose on personal phones are harder to rely on and easy to lose. An app that attaches photos to the entry with the date and time keeps them together.</p>]]></content:encoded>
</item>
<item>
<title>llms.txt explained: what it is and whether your site should have one</title>
<link>https://flavoureak.co.uk/blog/llms-txt-explained/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/llms-txt-explained/</guid>
<pubDate>Wed, 04 Feb 2026 08:00:00 GMT</pubDate>
<category>GEO</category>
<description>llms.txt is a proposed markdown file at the root of a website that summarises the site for language models. It is cheap to add, but no major AI provider has committed to using it.</description>
<content:encoded><![CDATA[<ul><li>llms.txt is a proposed convention, published in 2024, for a markdown file at a website’s root that gives language models a short summary of the site and links to its key pages.</li><li>Major AI providers have not committed to reading llms.txt when they crawl sites or answer questions, so any benefit is uncertain.</li><li>robots.txt controls which crawlers may visit, a sitemap lists URLs for discovery, and llms.txt is a short, curated guide to the site written for a model to read.</li></ul><p>llms.txt is a proposed convention for a plain markdown file, placed at <code>/llms.txt</code> on a website, that gives language models a concise summary of the site and links to its most useful pages. It was put forward in 2024 by Jeremy Howard of Answer.AI, and many documentation sites have since published one. None of the major AI providers has committed to using it in search or in answers, so adding one costs little and its benefit is uncertain.</p>
<h2>Why llms.txt was proposed</h2>
<p>Language models have a limited context window, which is the amount of text they can take in at once. Web pages are full of material that wastes it, such as navigation, scripts, cookie banners and adverts. The idea behind llms.txt is that a site owner offers a short, clean map of the site in markdown, a simple text format that models handle well, so a model or tool can find the important information quickly.</p>
<p>The proposal was aimed mainly at the moment a question is asked. A developer working with a coding assistant, for example, can point it at a library’s llms.txt to bring the right documentation pages into the conversation. That remains its clearest use today.</p>
<h2>How an llms.txt file is structured</h2>
<p>The specification at llmstxt.org sets out a simple order of sections. Only the first is required.</p>
<ol>
<li><b>A title.</b> A level-one heading with the name of the site or project, written as <code># Name</code>.</li>
<li><b>A summary.</b> A blockquote, written as a line starting with <code>&gt;</code>, giving a short description with the key facts a reader needs to understand the rest of the file.</li>
<li><b>Optional detail.</b> Paragraphs or lists with more context, such as who the site is for or how to use the sections that follow. This part contains no headings.</li>
<li><b>Sections of links.</b> Each starts with a level-two heading, such as <code>## Services</code> or <code>## Guides</code>, followed by a markdown list of links in the form <code>- [Page name](url): one-line note</code>.</li>
<li><b>An Optional section.</b> A section headed <code>## Optional</code> holds secondary links that a tool can skip when it needs a shorter context.</li>
</ol>
<p>For a small business, a file might open with <code># Example Roofing Ltd</code> and a one-line summary of what the company does and where it works. Sections would then list the service pages, the areas covered, the most useful guides and the contact page, each with a short note on what the page contains. A line in the services section might read <code>- [Flat roofs](https://example.com/flat-roofs): EPDM and GRP flat roofs, repairs and replacements</code>.</p>
<p>The proposal also suggests offering a clean markdown version of individual pages at the same address with <code>.md</code> added. Some documentation platforms also publish an <code>llms-full.txt</code> file that contains the full text of the documentation in one place.</p>
<h2>llms.txt, robots.txt and sitemaps compared</h2>
<p>The three files are often mentioned together, but they do different jobs for different readers.</p>
<table><thead><tr><th></th><th>robots.txt</th><th>sitemap.xml</th><th>llms.txt</th></tr></thead><tbody>
<tr><td>Purpose</td><td>Tells crawlers which paths they may fetch</td><td>Lists URLs for search engines to discover</td><td>Summarises the site and points to key pages</td></tr>
<tr><td>Format</td><td>Plain text rules</td><td>XML</td><td>Markdown</td></tr>
<tr><td>Status</td><td>Formal standard, RFC 9309</td><td>Long-established protocol supported by the major search engines</td><td>Community proposal, not formally adopted</td></tr>
<tr><td>Who reads it</td><td>Crawlers, before fetching pages</td><td>Search engine crawlers</td><td>Tools and models, usually when someone points them to it</td></tr>
<tr><td>Coverage</td><td>Rules for the whole site</td><td>Every URL you want indexed</td><td>A curated selection</td></tr>
<tr><td>Controls access</td><td>Yes, for crawlers that follow it</td><td>No</td><td>No</td></tr>
</tbody></table>
<p>The last row matters most. llms.txt has no power to allow or block anything. If you want to control whether AI crawlers can fetch your pages for training or for search, that is done in robots.txt, as explained in <a class="link" href="https://flavoureak.co.uk/blog/ai-crawlers-robots/">AI crawlers and robots.txt</a>.</p>
<h2>Whether AI providers use llms.txt</h2>
<p>As of late 2026, none of the major AI providers has publicly committed to reading llms.txt when its crawlers visit a site, or when its answer engine chooses what to cite. Google’s guidance on appearing in its AI search features doesn’t ask for one. You can check your own server logs to see whether anything requests the file, and which user agents do.</p>
<p>Some AI companies publish llms.txt files for their own developer documentation, which shows the format is useful for that purpose. It doesn’t mean their crawlers look for the file on other people’s sites. Coding assistants and other tools that can fetch web pages will read it when a user asks them to, which is the use the proposal had in mind.</p>
<p>This could change. If a major provider announces support, the file will matter more, and a site that already has an accurate one will have lost nothing by being early.</p>
<h2>Deciding whether to add one</h2>
<p>For most business websites it is a low-cost, low-certainty addition. Writing a good one takes an hour or two for a small site, and it does no harm. It won’t replace the things that make you easier for answer engines to find and quote, which are clear pages, content in server-rendered HTML, consistent facts about your business across the web, and passages that answer questions directly. Those are covered in <a class="link" href="https://flavoureak.co.uk/blog/geo-become-the-source/">how to become the source an AI answer cites</a>.</p>
<p>For a site with only a handful of pages, the file adds little that a model couldn’t get from the pages themselves. The case grows with the size of the site and the amount of material a reader has to sift through.</p>
<p>It is a clearer yes for software products with developer documentation or a public API. Developers already paste documentation into AI tools, and a well-kept llms.txt makes that quicker and more accurate.</p>
<h2>How to write an llms.txt that helps</h2>
<ul>
<li>Write the summary as plain facts covering what you do, for whom, where, and what sets you apart.</li>
<li>Link only to the pages that answer the questions customers ask, and leave the rest to your sitemap.</li>
<li>Give each link a one-line note saying what the page contains.</li>
<li>Use the same company name and description as on your site and your Companies House record.</li>
<li>Keep it in step with the site. Generate it from your content when the site is built, if you can, so it never describes a page that has gone.</li>
<li>Serve it as plain text at the root, and check that it loads at <code>yourdomain.com/llms.txt</code>.</li>
<li>Make it a summary of the site, and never add claims that the site itself doesn’t make.</li>
</ul>
<p>Review it whenever you add a service, retire a product or restructure the site, in the same way you would update the sitemap. An out-of-date llms.txt is unlikely to be read often, but when it is read, it should be right.</p>
<div class="callout"><b>Want your site ready for AI answer engines?</b> We’d generate llms.txt, sitemaps and structured data from the same content as the pages, so they always agree. <a class="link" href="https://flavoureak.co.uk/geo/">See our GEO guide</a>.</div>
<h2>Questions people ask</h2>
<h3>Is llms.txt an official standard?</h3>
<p>No. It is a community proposal published in 2024, with a specification at llmstxt.org. It hasn’t been adopted by a standards body, unlike robots.txt, which is defined in RFC 9309.</p>
<h3>Do ChatGPT and Google read llms.txt?</h3>
<p>Neither OpenAI nor Google has committed to using llms.txt for search or for choosing sources. A tool that can browse may read the file if you give it the address, in the same way it would read any page.</p>
<h3>Will llms.txt improve my rankings or AI citations?</h3>
<p>There is no evidence that it does today. It has no known effect on Google rankings, and no major answer engine has said it uses the file to choose sources. Treat it as a small, low-risk addition.</p>
<h3>Can llms.txt stop AI companies using my content?</h3>
<p>No. llms.txt contains no rules for access or training. To control AI crawlers, use robots.txt directives for each crawler’s user agent, and bear in mind that robots.txt relies on crawlers choosing to follow it.</p>]]></content:encoded>
</item>
<item>
<title>SEO for JavaScript and single-page app websites: what Google sees</title>
<link>https://flavoureak.co.uk/blog/seo-for-javascript-sites/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/seo-for-javascript-sites/</guid>
<pubDate>Thu, 29 Jan 2026 08:00:00 GMT</pubDate>
<category>SEO</category>
<description>Google can render JavaScript, but in a separate step, and many AI crawlers don’t run it at all. Server-side rendering or static generation is the safest way to be found and quoted.</description>
<content:encoded><![CDATA[<ul><li>Google crawls a page, queues it for rendering in a headless Chromium browser, and indexes what the rendered page contains.</li><li>Server-side rendering or static generation puts content and links in the HTML the server sends, so crawlers can read it without running JavaScript.</li><li>Each page needs its own URL using the History API, real links with an href, and its own title, description and canonical tag.</li></ul><p>A JavaScript website can rank well in Google, as long as search engines can get the content, links and metadata of every page. Google renders JavaScript in a separate step after crawling, while many AI crawlers read only the HTML the server sends and never run scripts. A single-page app, or SPA, is a site that loads one HTML shell and then uses JavaScript to draw each page in the browser, which is why it needs extra care.</p>
<h2>How Google processes JavaScript</h2>
<p>Google’s documentation describes three stages for JavaScript pages: crawling, rendering and indexing.</p>
<ol>
<li><b>Crawling.</b> Googlebot checks robots.txt, fetches the URL and reads the HTML the server returns. It collects any links it finds in that HTML.</li>
<li><b>Rendering.</b> The page joins a queue. When resources allow, a headless version of Chromium loads it, runs the JavaScript and produces the finished page.</li>
<li><b>Indexing.</b> Google reads the rendered page, including content and links added by JavaScript, and uses it for indexing.</li>
</ol>
<p>Google queues pages that return a 200 status for rendering unless they tell it not to index them. The risks are in timing and completeness. Rendering happens after crawling, scripts that fail or time out leave content missing, and anything that appears only after a click or a scroll may never be seen, because Googlebot doesn’t interact with a page the way a person does.</p>
<p>There is also one trap that catches many teams. If the HTML the server sends contains a noindex robots tag, Google may skip rendering altogether. JavaScript that removes the tag later never gets the chance to run.</p>
<h2>Rendering options compared</h2>
<p>How your pages are built decides what a crawler receives on its first request. This is the most important SEO decision on a JavaScript site.</p>
<table><thead><tr><th>Approach</th><th>What the crawler first receives</th><th>Good for</th><th>Trade-offs</th></tr></thead><tbody>
<tr><td>Client-side rendering (CSR)</td><td>An almost empty HTML shell and a script</td><td>Logged-in apps and dashboards</td><td>Content depends on rendering; weak for search and AI crawlers</td></tr>
<tr><td>Server-side rendering (SSR)</td><td>Full HTML built on the server for each request</td><td>Pages that change often or vary by visitor</td><td>Needs a running server; pages are only as fast as the server</td></tr>
<tr><td>Static generation (SSG)</td><td>Full HTML built in advance at deploy time</td><td>Marketing pages, blogs, documentation</td><td>Content changes need a rebuild or incremental regeneration</td></tr>
<tr><td>Dynamic rendering</td><td>Pre-rendered HTML for bots, the app for people</td><td>A stopgap for older single-page apps</td><td>Google describes it as a workaround and doesn’t recommend it long term</td></tr>
</tbody></table>
<p>Most modern frameworks, including Next.js, Nuxt, SvelteKit and Astro, can mix these approaches page by page. A common pattern is static generation for public marketing pages and client-side rendering behind the login, where search engines don’t need to go. Static pages also tend to load quickly, which helps with <a class="link" href="https://flavoureak.co.uk/blog/core-web-vitals-static/">Core Web Vitals</a>.</p>
<h2>Hash routing and why each page needs a real URL</h2>
<p>Older single-page apps often used addresses like <code>example.com/#/pricing</code>. Everything after the <code>#</code> is a fragment. Browsers don’t send fragments to the server, and Google generally treats addresses that differ only by fragment as the same URL. A site built this way can look to Google like a single page.</p>
<p>The fix is the History API, which lets an app change the address bar to a normal path such as <code>example.com/pricing</code> without reloading the page. The server must also respond correctly when someone requests that path directly, returning the right content with a 200 status, or a 404 status when the page doesn’t exist.</p>
<p>That last point catches many SPAs. If the server returns the app shell with a 200 status for every path, the address of a deleted product looks like a live page. Google calls these soft 404s. Return a real 404 status from the server, or have the error view redirect to a URL that does, or add a noindex robots tag to it.</p>
<h2>Links that crawlers can follow</h2>
<p>Google follows links only when they are <code>a</code> elements with an <code>href</code> attribute holding a URL it can resolve. A button or a <code>div</code> with a click handler that changes the page works for people and is invisible to crawlers. The link components in most frameworks render a proper <code>a</code> element with an <code>href</code>, so use them instead of navigating inside a click handler.</p>
<p>Links are also how search engines judge how pages relate and which matter most, which we cover in <a class="link" href="https://flavoureak.co.uk/blog/internal-linking/">internal linking for SEO</a>. An XML sitemap helps discovery, but it doesn’t replace links between pages.</p>
<h2>Metadata for every page</h2>
<p>Each URL needs its own title, meta description and canonical tag, and pages meant to be shared need their own Open Graph tags. In client-rendered apps these are often set once in the shell and never changed, so every page claims to be the home page.</p>
<ul>
<li>Set the title and meta description per route, ideally in the server-rendered HTML.</li>
<li>Point each canonical tag at the page’s own clean URL.</li>
<li>Put any noindex directive in the server response, never in JavaScript alone.</li>
<li>Render structured data with the page, so it is present in the first HTML.</li>
<li>Don’t block JavaScript or CSS files in robots.txt, because Google needs them to render the page.</li>
<li>Use real image elements with alt text for images that matter, and avoid loading key content only on scroll.</li>
</ul>
<h2>Testing what Google and AI crawlers see</h2>
<p>The URL Inspection tool in Google Search Console shows how Google sees a page. Run a live test, then view the tested page to see the rendered HTML, a screenshot and any resources that failed to load. If your content or links are missing from the rendered HTML, Google is missing them too. Google’s Rich Results Test also shows rendered HTML, and works for public URLs on sites you haven’t verified.</p>
<p>For AI crawlers, check the raw HTML. Many AI crawlers fetch a page and read the HTML the server returns without running JavaScript, so a client-rendered page can look empty to them. Use your browser’s view source, or fetch the page with a command-line tool such as <code>curl</code>, and confirm the main text and links are there. Whether you let those crawlers in at all is a separate decision, covered in <a class="link" href="https://flavoureak.co.uk/blog/ai-crawlers-robots/">AI crawlers and robots.txt</a>.</p>
<p>Search Console’s Page indexing report is the longer-term check. Watch for public pages listed as soft 404s, as duplicates without a canonical, or as crawled but not indexed, and inspect a few of each to find the pattern.</p>
<div class="callout"><b>Building a site on a JavaScript framework?</b> We choose the rendering approach page by page, so public pages arrive as finished HTML and the app stays fast behind the login. <a class="link" href="https://flavoureak.co.uk/seo/">See our technical SEO guide</a>.</div>
<h2>Questions people ask</h2>
<h3>Can Google index a React or Vue website?</h3>
<p>Yes. Google renders JavaScript and can index content that React, Vue and other frameworks add to the page. Indexing is more dependable when the content is already in the server’s HTML through server-side rendering or static generation, because it doesn’t then rely on rendering working perfectly.</p>
<h3>Is client-side rendering bad for SEO?</h3>
<p>It is riskier for public pages that need to rank. Content and links depend on the rendering step, metadata mistakes are common, and many AI crawlers won’t see the content at all. It is fine for logged-in areas that search engines don’t need to index.</p>
<h3>Do AI crawlers run JavaScript?</h3>
<p>Many don’t. Several AI crawlers read only the HTML the server sends, so content drawn by JavaScript can be invisible to them. Putting your main content in server-rendered or static HTML is the safest way to be read and quoted.</p>
<h3>Does Google index URLs with a hash?</h3>
<p>Google generally ignores the part of a URL after the hash, so pages that differ only by fragment are treated as one URL. Use normal paths through the History API for any page that should appear in search on its own.</p>]]></content:encoded>
</item>
<item>
<title>How to pass Apple App Store and Google Play review first time</title>
<link>https://flavoureak.co.uk/blog/app-store-review-first-time/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/app-store-review-first-time/</guid>
<pubDate>Fri, 23 Jan 2026 08:00:00 GMT</pubDate>
<category>Engineering</category>
<description>Apps are rejected for a familiar set of reasons, including crashes, missing information, no demo login, payment rules and privacy forms. Here is a checklist to work through before you submit.</description>
<content:encoded><![CDATA[<ul><li>Common reasons for rejection include crashes, placeholder content, incomplete store information, a login without a demo account, and breaches of payment or privacy rules.</li><li>Apps that sell digital content or features inside the app generally have to use Apple’s and Google’s in-app purchase systems, while physical goods and real-world services do not.</li><li>Both stores require apps that let people create an account to let them delete it, and both require accurate privacy disclosures before an app is published.</li></ul><p>To pass Apple App Store and Google Play review first time, submit a finished app that does not crash, give reviewers a working demo login, fill in every privacy and store form accurately, follow each store’s payment rules and let people delete their accounts. App review is the check each store makes against its published rules before an app or update is made available to the public. The common reasons for rejection are predictable, and this article goes through each of them.</p>
<h2>How the two review processes work</h2>
<p>Apple reviews every app and every update against its App Review Guidelines before release, and its reviewers try the app as a customer would. If something is wrong, Apple names the guideline the app breaks in App Store Connect. You can reply, fix and resubmit, or appeal to the App Review Board if you think the decision is wrong.</p>
<p>Google Play checks apps against its Developer Program Policies, using a mix of automated and human review. First submissions from new developer accounts can take longer than later updates. Google also currently requires new personal developer accounts to run a closed test with real testers for a set period before they can publish to production, which is one more reason to register the Play Console account as your organisation.</p>
<p>For both stores, allow time. Plan a launch around a review that may need a second attempt, and never announce a date that depends on the first submission being approved.</p>
<h2>Common reasons apps are rejected</h2>
<p>Apple publishes guidance on the issues it sees most, and Google’s policies cover similar ground. These are the ones that catch first-time submissions.</p>
<table><thead><tr><th>Reason</th><th>Store</th><th>How to avoid it</th></tr></thead><tbody><tr><td>Crashes, broken links or placeholder content</td><td>Both</td><td>Test the exact build you submit on real devices, and remove unfinished screens and sample text</td></tr><tr><td>A login with no way for the reviewer to get in</td><td>Both</td><td>Provide a working demo account or a full demo mode, with any instructions needed</td></tr><tr><td>Screenshots or descriptions that do not match the app</td><td>Both</td><td>Show the real app in use, and leave other platforms out of Apple listings</td></tr><tr><td>Digital goods sold without in-app purchase</td><td>Both</td><td>Use the store’s billing for digital content and features used in the app</td></tr><tr><td>Missing or inaccurate privacy information</td><td>Both</td><td>Declare what the app and every SDK in it actually collects</td></tr><tr><td>No way to delete an account</td><td>Both</td><td>Add account deletion in the app, plus a web link for Google Play</td></tr><tr><td>Sensitive permissions without justification</td><td>Both</td><td>Explain clearly why the app needs the camera, location or contacts, and complete Google’s declaration for restricted permissions</td></tr><tr><td>Social login without a private equivalent</td><td>Apple</td><td>Offer a login that meets guideline 4.8, such as Sign in with Apple</td></tr><tr><td>A thin wrapper around a website</td><td>Apple</td><td>Give the app features that make sense as an app, under guideline 4.2</td></tr></tbody></table>
<h2>Demo accounts and access for reviewers</h2>
<p>If your app needs a login, the reviewer needs one too. Apple asks for an active demo account or a fully featured demo mode in the App Review Information section, and Google Play asks for access instructions in the App access section of Play Console. Create an account that stays active, holds realistic sample data and is not protected by a two-factor code sent to someone’s personal phone. If a feature needs a particular location, device or piece of hardware, explain how the reviewer can see it, or attach a video.</p>
<p>Business software catches people out here. An app for a firm’s own staff, or one that only works after a sales call, still has to be reviewable. Apple also has routes for apps that are not meant for the general public, such as unlisted distribution and custom apps for businesses, so check those before you submit a private app to the public store.</p>
<h2>In-app purchase rules for digital goods</h2>
<p>Both stores generally require digital content and features that are bought and used inside the app, such as subscriptions, premium features or in-game currency, to be sold through their own in-app purchase systems. Physical goods, and services delivered in the real world, such as a delivery, an event ticket or a booked tradesperson, are paid for with ordinary methods such as card or Apple Pay.</p>
<p>Live lessons need a careful read. Apple’s guidelines allow ordinary payment for real-time person-to-person services between two individuals, such as one-to-one tutoring, while one-to-few and one-to-many live services must use in-app purchase. If your app teaches, coaches or consults, read the current wording closely before you design the checkout.</p>
<p>The rules on telling users about prices elsewhere, or linking to a website to pay, differ between countries and have changed several times recently, including in the United States and the European Union. Check the current rules for every storefront you sell in.</p>
<h2>Privacy forms and account deletion</h2>
<p>Both stores ask you to declare what data your app collects and why. On the App Store these are the privacy details shown on the product page, and on Google Play this is the Data safety form. The declarations must cover data collected by third-party code in your app as well as your own, including analytics, crash reporting and advertising SDKs. Apple also requires a privacy manifest that explains why the app uses certain system APIs, and an app that tracks people across other companies’ apps and websites must ask permission through App Tracking Transparency. Both stores require a link to a privacy policy.</p>
<p>Any app that lets people create an account must also let them start deleting it from inside the app. Google Play additionally requires a web link where people can ask for their account and data to be deleted without reinstalling the app. Deleting an account should remove the personal data that goes with it, apart from records you are legally required to keep. UK GDPR gives people a right to erasure in many circumstances in any case, with one month to respond to a request.</p>
<p>Apps for children face further rules, through Apple’s Kids Category and Google Play’s Families policy. In the UK, the ICO’s Children’s code applies to online services likely to be accessed by children, as covered in our guide to <a class="link" href="https://flavoureak.co.uk/blog/childrens-code-for-edtech/">the Children’s code for edtech</a>.</p>
<h2>A checklist before you submit</h2>
<ol><li>Test the exact release build on several real devices, including an older one, and on a poor connection.</li><li>Remove test screens, placeholder text and links that go nowhere.</li><li>Create a demo account with realistic data and put the details in the review notes.</li><li>Check that every permission prompt has a clear, specific explanation.</li><li>Confirm how each purchase is paid for and whether store billing applies.</li><li>Complete the privacy details, the Data safety form, the age rating questionnaires and any content declarations.</li><li>Add the privacy policy and support URLs, and check that they load.</li><li>Make sure account deletion works in the app, and from the web link for Google Play.</li><li>Prepare screenshots from the real app at the sizes each store asks for.</li><li>Send the build to TestFlight external testers and a Google Play closed test first. Apple reviews the first build sent to external testers, and Google’s pre-launch report runs the app on real devices and flags crashes.</li></ol>
<div class="callout"><b>Launching an app?</b> Plan the store accounts, demo login and privacy declarations alongside the build, well before the launch date. <a class="link" href="https://flavoureak.co.uk/mobile/">See how we approach mobile apps</a>.</div>
<h2>Questions people ask</h2>
<h3>How long does App Store review take?</h3>
<p>It varies, and Apple does not promise a fixed time. Many submissions are reviewed within a day or two, but first submissions, complex apps and busy periods can take longer, and a rejection adds another round. Build a buffer of at least a week into any launch plan.</p>
<h3>Can I appeal an App Store rejection?</h3>
<p>Yes. You can reply to the reviewer in App Store Connect to explain or ask for clarification, and if you believe the decision is wrong you can appeal to the App Review Board. Google Play also has an appeals process for policy decisions. Fixing the issue and resubmitting is usually quicker than an appeal.</p>
<h3>Do I need Sign in with Apple?</h3>
<p>Only in some cases. If your app uses a third-party or social login such as Google to set up or sign in to the main account, Apple’s guideline 4.8 generally requires an equivalent option that limits data collection to name and email and lets people keep their email address private. Sign in with Apple meets that requirement. Apps that use only their own accounts do not need it.</p>
<h3>Why was my app rejected for being a website in an app?</h3>
<p>Apple’s minimum functionality guideline, 4.2, expects an app to offer more than a repackaged website. An app that only shows a website in a web view is often rejected. Add features that justify an app, or consider whether a <a class="link" href="https://flavoureak.co.uk/blog/pwa-or-native-app/">progressive web app</a> fits better.</p>]]></content:encoded>
</item>
<item>
<title>Retrieval-augmented generation (RAG) explained for business software</title>
<link>https://flavoureak.co.uk/blog/rag-explained/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/rag-explained/</guid>
<pubDate>Sat, 17 Jan 2026 08:00:00 GMT</pubDate>
<category>AI</category>
<description>RAG finds the passages in your own documents that answer a question and gives them to a language model, so the answer comes from your sources and can cite them.</description>
<content:encoded><![CDATA[<ul><li>Retrieval-augmented generation (RAG) searches your own documents for relevant passages and gives them to a language model with the question, so the answer is based on your sources.</li><li>Hybrid search, which combines meaning-based vector search with keyword search, often finds the right passages more reliably than either method alone.</li><li>Permissions must be applied when passages are retrieved, before anything reaches the model, so no customer or user sees documents they are not allowed to.</li></ul><p>Retrieval-augmented generation (RAG) is a way of answering questions with a language model by first searching your own documents for the passages most relevant to the question, then giving those passages to the model with instructions to answer from them and cite them. It lets a model answer from information it was never trained on, such as your policies, contracts or product data, and it lets people check every answer against its source. RAG suits questions whose answer is written down somewhere in your documents, and it is the wrong tool for totals, calculations and tasks that need a whole document at once.</p>
<h2>Why RAG exists</h2>
<p>A language model knows what was in its training data and nothing about your business. It has not read your health and safety procedures, your price book or last year’s tender responses. Asked about them, it will either say it does not know or, worse, write a plausible answer that is wrong.</p>
<p>There are two ways to give a model your information. You can paste the documents into the prompt, which works when they are short. Or you can search them and send only the relevant parts, which is RAG. Once you have more than a handful of documents, searching is cheaper and faster, and usually more accurate, because the model is not distracted by thousands of pages that have nothing to do with the question.</p>
<h2>How a RAG system works</h2>
<ol><li><b>Ingest.</b> Documents are collected from where they live, such as a file store, a document management system or the application’s own database, and turned into clean text.</li><li><b>Chunk.</b> Each document is split into passages small enough to be precise and large enough to make sense on their own, often a few hundred words. Good chunking follows the document’s headings and keeps the title and section name with each passage.</li><li><b>Embed.</b> Each chunk is passed through an embedding model, which turns text into a list of numbers called a vector. Texts with similar meanings get vectors that sit close together, even when they use different words.</li><li><b>Index.</b> The vectors are stored in an index along with the text, the source document, the page or section, and who is allowed to see it. The index can live in an ordinary database, such as PostgreSQL with the pgvector extension, or in a dedicated vector database.</li><li><b>Retrieve.</b> When someone asks a question, it is embedded in the same way and the closest chunks are found. Keyword search usually runs alongside, and a reranking step can reorder the candidates by how well they answer the question.</li><li><b>Generate.</b> The best few chunks go to the language model with the question and instructions to answer only from them, cite them, and say when the answer is not there.</li></ol>
<h2>Vector, keyword and hybrid search</h2>
<p>Retrieval often matters more to the quality of a RAG system than the choice of model. If the right passage is not retrieved, the model cannot use it.</p>
<table><thead><tr><th>Method</th><th>How it matches</th><th>Good at</th><th>Weak at</th></tr></thead><tbody><tr><td>Keyword search, such as BM25</td><td>Passages that contain the words in the question, weighted by how rare those words are</td><td>Exact terms, codes, part numbers, names and regulation references</td><td>Different wording for the same idea</td></tr><tr><td>Vector search</td><td>Passages whose meaning is close to the question</td><td>Paraphrases and questions in everyday language</td><td>Exact codes and rare terms, which can be lost in the overall meaning</td></tr><tr><td>Hybrid search</td><td>Runs both and merges the results</td><td>Real questions, which tend to mix everyday language with specific terms</td><td>Needs tuning, and is more to build and run</td></tr><tr><td>Reranking</td><td>A second model scores the top candidates against the question</td><td>Choosing the best few from a longer list</td><td>Adds time and cost to every question</td></tr></tbody></table>
<p>Business documents are full of exact terms. A question about “CDM 2015 duties of the principal contractor” needs keyword matching for “CDM 2015” and meaning-based matching for the rest, which is why we’d recommend hybrid search as the default.</p>
<h2>Permissions and tenants</h2>
<p>In business software, not everyone may see every document. In a SaaS product, each customer’s documents belong to that customer, and within a customer some documents are restricted to certain roles. A RAG system must respect both.</p>
<p>The rule is to filter at retrieval. Every chunk is stored with its tenant and access information, and every search includes a filter for the current tenant and user, applied before results are ranked and long before anything reaches the model. An instruction in the prompt telling the model not to reveal certain documents is not a security control, because a passage the model has been given can always end up in an answer. When permissions change, the index has to change with them, and a deleted document must be removed from the index as well as from storage. Our guide to <a class="link" href="https://flavoureak.co.uk/blog/multi-tenancy-explained/">multi-tenancy</a> covers how each customer’s data is kept separate more generally.</p>
<h2>Citations people can check</h2>
<p>Every answer should show where it came from. Because each chunk carries its source document and location, the system can show a citation such as “Hot works procedure, section 4” beside each claim and link straight to that passage.</p>
<p>Citations also make errors visible. If the model cites a passage that does not support what it said, a reviewer can see the problem. You can check this automatically too, by confirming that each cited chunk was among those retrieved and that any quoted text appears in it. Where the documents do not contain an answer, the system should say so plainly. Designing for that case is covered in <a class="link" href="https://flavoureak.co.uk/blog/designing-for-wrong-answers/">designing for wrong answers</a>.</p>
<h2>A worked example over a company’s documents</h2>
<p>Take a building contractor with 2,000 documents in its shared drive, including health and safety policies, method statements, risk assessments, product data sheets and past tender responses. Say they average eight pages of 400 words each.</p>
<ul><li>That is 2,000 × 8 × 400 = 6.4 million words, or roughly 8.5 million tokens at about 1.33 tokens per word.</li><li>Split into chunks of around 500 tokens, the collection becomes about 17,000 chunks. Each is embedded once and embedded again only when its document changes.</li><li>A site manager asks, “What do we need in place before hot works on a flat roof?” Hybrid search returns 30 candidate chunks, already filtered to this company and to documents the site manager may see.</li><li>A reranker picks the best five. With instructions and the question, the model receives about 3,000 input tokens and writes an answer of about 300 tokens.</li><li>The answer lists what the company’s own hot works procedure requires, such as the permit and the fire watch, with a citation to the section each point comes from.</li></ul>
<p>The cost per question follows the same sum as any model call. It is 3,000 input tokens at the input rate plus 300 output tokens at the output rate, plus a small cost for embedding the question and reranking. Our guide to <a class="link" href="https://flavoureak.co.uk/blog/ai-cost-per-call/">what an AI feature costs per call</a> shows how to put rates into that sum.</p>
<h2>When RAG is the wrong tool</h2>
<ul><li><b>Totals and calculations.</b> “How much did we invoice in March?” is a database query. Give the model a tool that runs the query, instead of searching invoice PDFs.</li><li><b>Whole-document tasks.</b> Summarising one contract or comparing two versions needs the whole text. If it fits in the model’s context window, send it.</li><li><b>Small, stable knowledge.</b> A two-page policy can go in the prompt every time, with no index at all.</li><li><b>Out-of-date or contradictory documents.</b> RAG repeats what your documents say. If the drive holds three versions of a procedure, tidy it first or mark which version is current.</li><li><b>Answers that must be exact every time.</b> Where a wrong answer has legal or safety consequences, use RAG to find and show the source, and let a person read it.</li></ul>
<div class="callout"><b>Thinking about AI over your own documents?</b> Start with a test set of real questions and the passages that answer them. <a class="link" href="https://flavoureak.co.uk/ai/">See how we build AI features</a>.</div>
<h2>Questions people ask</h2>
<h3>What is the difference between RAG and fine-tuning?</h3>
<p>RAG gives a model your information at the moment it answers, by retrieving relevant passages. Fine-tuning changes the model itself by training it further on examples, which suits teaching a style or format better than adding facts. For answering from documents that change, RAG is usually the better fit, because updating a document updates the answers.</p>
<h3>What is an embedding?</h3>
<p>An embedding is a list of numbers that represents the meaning of a piece of text, produced by an embedding model. Texts with similar meanings have embeddings that are close together, which lets a system find relevant passages even when they use different words from the question.</p>
<h3>Does RAG stop AI from making things up?</h3>
<p>It reduces the problem without removing it. A model can still misread a passage or combine two passages wrongly. Citations, instructions to say when the answer is not in the sources, and a test set of questions with known answers keep the risk visible and measurable.</p>
<h3>How do you measure whether a RAG system works?</h3>
<p>Build a test set of real questions with the passages that answer them. Measure how often the right passage appears in the retrieved results, then how often the final answer is correct and properly cited. Run the test set again after every change to chunking, search or prompts.</p>]]></content:encoded>
</item>
<item>
<title>Fixed price vs time and materials: choosing a software contract model</title>
<link>https://flavoureak.co.uk/blog/fixed-price-or-time-and-materials/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/fixed-price-or-time-and-materials/</guid>
<pubDate>Mon, 12 Jan 2026 08:00:00 GMT</pubDate>
<category>Working with us</category>
<description>Fixed price puts the risk of overruns on the supplier and suits well-defined work. Time and materials puts it on the client and suits products that will change. A retainer buys ongoing capacity.</description>
<content:encoded><![CDATA[<ul><li>A fixed-price contract sets one price for an agreed scope, so the supplier carries the risk if the work takes longer, and any change to the scope goes through a change request.</li><li>Time and materials bills for the time actually spent at agreed day or hourly rates, so the client carries the risk of overruns and keeps the freedom to change direction.</li><li>A retainer pays for an agreed amount of capacity each month and suits ongoing improvements, support and maintenance once a product is live.</li></ul><p>Fixed price, time and materials, and retainers are the three common ways to pay for software work, and the main difference between them is who carries the risk when work takes longer than expected. A fixed-price contract sets one price for a defined scope. Time and materials charges for the hours or days actually worked. A retainer buys a set amount of a team’s time each month.</p>
<h2>How fixed price works</h2>
<p>The supplier estimates the work, adds a contingency for the risk it is taking on, and quotes one price for an agreed scope. Payments are usually tied to milestones, such as signing the contract, delivering a prototype and acceptance of the finished release. If the work takes longer, the supplier absorbs the cost. If it goes faster, the supplier keeps the difference.</p>
<p>Take an illustrative day rate of £500, chosen to make the sums easy. The supplier estimates 40 days, which is £20,000, and adds 25% contingency for a fixed price of £25,000. If the work takes 32 days, the client pays £25,000. If it takes 55 days, worth £27,500, the client still pays £25,000 and the supplier absorbs the extra £2,500 of its own time.</p>
<p>Fixed price only works when the scope is written precisely enough that both sides would agree whether a given request is in or out. That needs a specification with acceptance criteria, which is one reason fixed-price builds usually follow a <a class="link" href="https://flavoureak.co.uk/blog/discovery-phase-explained/">discovery phase</a>.</p>
<h2>How time and materials works</h2>
<p>Under time and materials, the client pays for the time worked at agreed rates, usually reported weekly or at the end of each sprint, plus any agreed costs such as licences. The supplier gives an estimate, often as a range, and the final cost follows the actual work.</p>
<p>In the same example, 32 days would cost £16,000 and 55 days would cost £27,500. The client gains if the work is quicker and pays more if it is slower. In exchange, the client can change priorities at any point without renegotiating, which suits a new product where you learn as you build. A budget cap, sometimes called capped time and materials, limits the exposure. When spending reaches the cap, work pauses for a decision.</p>
<p>To keep time and materials under control, ask for working software at the end of each sprint, a simple chart of work completed against budget spent, and a regular review where you reorder what comes next. Seeing real progress every week or two is the best early warning of an overrun.</p>
<h2>How a retainer works</h2>
<p>A retainer pays a fixed monthly fee for an agreed amount of capacity, such as a number of days of development, support and hosting management. It suits a live product that needs steady improvement, security updates and help when something goes wrong. Agree what happens to unused days, how urgent issues are handled, and how much notice either side needs to change the amount. <a class="link" href="https://flavoureak.co.uk/blog/software-maintenance-costs/">Software maintenance costs</a> explains what a live product needs each month.</p>
<h2>How each model shares risk</h2>
<table><thead><tr><th></th><th>Fixed price</th><th>Time and materials</th><th>Retainer</th></tr></thead><tbody><tr><td>Who carries overrun risk</td><td>Supplier</td><td>Client</td><td>Shared, within the monthly capacity</td></tr><tr><td>Cost certainty for the client</td><td>High, for the agreed scope</td><td>Low to medium, unless capped</td><td>High, per month</td></tr><tr><td>Freedom to change scope</td><td>Low, through change requests</td><td>High</td><td>High, within the capacity</td></tr><tr><td>Contingency in the price</td><td>Built in</td><td>None, you pay for actual time</td><td>None</td></tr><tr><td>Best for</td><td>Well-defined, bounded work</td><td>New products and evolving scope</td><td>Live products</td></tr><tr><td>Main risk</td><td>Disputes over what is in scope</td><td>Budget drift</td><td>Paying for capacity you don’t use</td></tr></tbody></table>
<p>Fixed price can cost more than time and materials for the same work, because the supplier prices in contingency. It can also cost less, if the work turns out harder than anyone expected. The trade is certainty against flexibility.</p>
<h2>When each model fits</h2>
<ul><li><b>Fixed price</b> fits a discovery phase, a bounded integration, a marketing site, or a build whose scope has been fully specified.</li><li><b>Time and materials</b> fits the first version of a product, where real users will change what you build, and any work with big unknowns, such as integrating with an undocumented system.</li><li><b>A retainer</b> fits a product after launch, when the work is a steady flow of small improvements and fixes.</li></ul>
<p>Many projects combine them. A common pattern is a fixed-price discovery phase, then time and materials for the build in fixed-length sprints with a budget cap, then a retainer once the product is live.</p>
<h2>How change requests work</h2>
<p>A change request is the formal way to alter an agreed scope. In a fixed-price contract it is the only way, and a clear process prevents most disputes.</p>
<ol><li>The client describes the change in writing.</li><li>The supplier assesses its effect on cost, timeline and other features, and replies in writing.</li><li>The supplier offers options, such as adding cost, swapping out another feature of similar size, or moving the change to a later phase.</li><li>The client approves one option in writing before work starts.</li><li>The scope document, plan and price are updated, so everyone works from the same version.</li></ol>
<p>For example, partway through a fixed-price build the client asks for invoices to be sent to their accounting package. The supplier estimates 4 days, or £2,000 at the example rate, and offers to drop a reporting screen of similar size instead. The client chooses the swap, and the price stays at £25,000.</p>
<p>Under time and materials, changes are handled by reordering the backlog. They still cost money, so the same habit of writing them down and agreeing their effect on the budget helps.</p>
<h2>What to put in the contract, whichever model you choose</h2>
<p>The pricing model is one clause among several that matter just as much.</p>
<ul><li>Who owns the code, designs and documents, and when ownership passes to you (<a class="link" href="https://flavoureak.co.uk/blog/who-owns-the-code-uk/">who owns the code</a>).</li><li>Acceptance criteria, and how long the client has to test each delivery.</li><li>Payment terms and what triggers each invoice.</li><li>A warranty period for fixing defects found after acceptance.</li><li>Termination terms, and handover of the repository, credentials and documentation if you part ways.</li></ul>
<p>Take legal advice on the contract itself, especially on intellectual property and liability.</p>
<div class="callout"><b>Comparing ways to pay for a build?</b> <a href="https://flavoureak.co.uk/pricing/">See how we price projects</a>.</div>
<h2>Questions people ask</h2>
<h3>Is fixed price cheaper than time and materials?</h3><p>Not necessarily. A fixed price includes contingency for the supplier’s risk, so if the work goes to plan it usually costs more than the same work on time and materials. If the work turns out much harder than expected, fixed price protects the client. In effect, you are paying for certainty.</p>
<h3>Can a fixed-price project change scope?</h3><p>Yes, through a change request. The supplier assesses the effect of the change on cost and timeline, offers options such as swapping out another feature, and the client approves one in writing. Without that process, changes lead to disputes about what the fixed price covered.</p>
<h3>What is capped time and materials?</h3><p>Capped time and materials bills for the actual time worked, like normal time and materials, and sets a maximum spend for a phase or sprint. When the cap is reached, work pauses until the client decides whether to increase the budget, reduce the scope or stop. It limits risk without fixing the scope.</p>
<h3>How is a retainer different from time and materials?</h3><p>Both charge for time. A retainer commits to a fixed amount of capacity each month for a set term, which gives the client a predictable cost and the supplier a reason to keep the same people on the product. Time and materials charges for whatever time is used, with no fixed monthly amount.</p>
<h3>Which contract model suits a startup MVP?</h3><p>We’d usually suggest a fixed-price discovery phase followed by capped time and materials. Discovery fixes the cost of deciding what to build, and capped time and materials lets you change priorities as you learn while keeping spending under control.</p>]]></content:encoded>
</item>
<item>
<title>SaaS pricing models: per-seat, tiered, usage-based, flat-rate and hybrid</title>
<link>https://flavoureak.co.uk/blog/saas-pricing-models/</link>
<guid isPermaLink="true">https://flavoureak.co.uk/blog/saas-pricing-models/</guid>
<pubDate>Tue, 06 Jan 2026 08:00:00 GMT</pubDate>
<category>SaaS</category>
<description>For most new B2B SaaS products we’d start with two or three simple tiers, priced per account or per seat. Here is how each model works, what it does to billing code and when to change.</description>
<content:encoded><![CDATA[<ul><li>A SaaS pricing model is the rule that turns a customer’s use of your product into a bill, usually charged per seat, by tier, by usage, at a flat rate or as a mix of these.</li><li>Per-seat and tiered pricing are the simplest to bill and forecast, while usage-based pricing suits products whose costs and value grow with volume and needs metering you can trust.</li><li>Choose one model, publish it, and review it once enough customers pay to show who gets the most value; new prices for new customers are easy, while existing customers need notice under their contract.</li></ul><p>A SaaS pricing model is the rule that turns a customer’s use of your product into a bill. The five common models are per-seat, tiered, usage-based, flat-rate and hybrid, and each one shapes your revenue, your sales conversations and your billing code. For a new B2B product we’d usually start with two or three tiers priced per account or per seat, because they are simple to explain, simple to build and easy to forecast.</p>
<h2>The five common SaaS pricing models</h2>
<p>Each model answers one question, which is what the customer pays for. The table compares them using one example product, a job management tool for small building firms. The prices are illustrations chosen to make the arithmetic clear.</p>
<table><thead><tr><th>Model</th><th>What the customer pays for</th><th>Example in £</th><th>Billing effort</th></tr></thead><tbody><tr><td>Per-seat</td><td>Each user with access</td><td>£15 per user per month, so 12 users pay £180</td><td>Low to medium</td></tr><tr><td>Tiered</td><td>A package of features and limits</td><td>Starter £49, Team £129, Business £299 a month</td><td>Low</td></tr><tr><td>Usage-based</td><td>Units consumed, such as documents processed</td><td>£0.40 per document, so 2,000 documents cost £800</td><td>High</td></tr><tr><td>Flat-rate</td><td>Access to the whole product</td><td>£99 a month whatever the size of the firm</td><td>Lowest</td></tr><tr><td>Hybrid</td><td>A base fee plus seats or usage</td><td>£150 a month including 10 users, then £12 per extra user</td><td>Medium to high</td></tr></tbody></table>
<p>Prices for UK businesses are normally shown excluding VAT. If you are VAT registered, VAT at the standard rate of 20% is added to UK customers’ invoices, so a £180 subscription becomes a £216 invoice.</p>
<h2>How per-seat pricing works</h2>
<p>Per-seat pricing charges a fixed amount for each user. It is easy to understand, and revenue grows as a customer’s team grows. A firm with 12 users at £15 a month pays £180 a month, or £2,160 a year. If it gives four more office staff access, the bill rises to £240 a month.</p>
<p>The weakness is that it charges for the very thing you want customers to do more of, which is invite colleagues. Teams share logins to save money, and people who only need to read a report never get access. Some products deal with this by charging only for people who edit and letting viewers in free.</p>
<p>In the billing code, per-seat pricing means keeping the seat count on the subscription in step with the users in the account. You need to decide what counts as a seat (invited, active, or signed in this month) and what happens when someone is added partway through a billing period. Billing providers handle this with proration, which charges only for the part of the period that remains. Adding a £15 seat halfway through a month charges about £7.50.</p>
<h2>How tiered pricing works</h2>
<p>Tiered pricing sells packages. Each tier has a price, a set of features and some limits. Starter might be £49 a month for up to 3 users and 20 active jobs, Team £129 for up to 10 users, and Business £299 with unlimited users and single sign-on. Customers pick the package that fits, and upgrade when they reach a limit or need a feature.</p>
<p>Tiers work well when different sizes of customer genuinely need different things. A sole trader needs quotes and invoices. A firm with three site teams needs scheduling, permissions and reports. Keep the number of tiers small, and make the reason to move up obvious.</p>
<p>In code, tiers become entitlements. Each plan maps to a list of features and limits, and the product checks the entitlement, never the plan name, before showing a feature. That way you can change what a plan includes, or give one customer an exception, without a release. Entitlements say what the account has paid for. Permissions, a separate layer, say what each person in it may do.</p>
<h2>How usage-based pricing works</h2>
<p>Usage-based pricing charges for what the customer consumes, such as API calls, documents processed, messages sent or storage used. It fits products where your own costs rise with use. AI features are the clearest case, because every model call has a price (<a class="link" href="https://flavoureak.co.uk/blog/ai-cost-per-call/">what an AI call costs</a>).</p>
<p>There are two common ways to price volume. With graduated pricing, each band of units has its own rate. With volume pricing, the band the total falls into sets the rate for every unit. Take a rate of £0.50 per document for the first 1,000 and £0.30 after that, and a customer who processes 2,500 documents in a month.</p>
<ul><li><b>Graduated:</b> 1,000 × £0.50 = £500, plus 1,500 × £0.30 = £450, so £950.</li><li><b>Volume:</b> all 2,500 documents at £0.30 = £750.</li></ul>
<p>Usage pricing is the hardest model to build. Every billable event has to be recorded exactly once, stored against the account with a timestamp, totalled at the end of the period and shown to the customer as it builds up. Customers also worry about surprise bills, so plan for a usage screen, alerts at thresholds and optional spending caps.</p>
<h2>Flat-rate and hybrid pricing</h2>
<p>Flat-rate pricing charges one price for everything. It is the easiest model to sell and to build, and it suits a product with one clear type of customer. It stops working when your largest customers cost far more to serve than your smallest ones and pay the same.</p>
<p>Hybrid pricing combines a base fee with seats or usage. A plan of £150 a month including 10 users, then £12 for each extra user, charges a 14-user firm £150 + (4 × £12) = £198 a month. Add a usage charge for something expensive to run, such as AI document extraction, and the price covers your costs at every size of customer. The trade-off is a pricing page that takes longer to read and billing code with more moving parts.</p>
<h2>What each model does to your billing code</h2>
<p>Pricing is a product decision that turns into code. Before you choose, check what each model asks of the system.</p>
<ol><li><b>A plan catalogue</b> with prices, currencies, billing intervals and the entitlements each plan grants.</li><li><b>Subscription changes</b>, meaning upgrades, downgrades, cancellations and proration, with clear rules for when each takes effect.</li><li><b>Seat syncing</b> for per-seat models, so the bill follows the user list.</li><li><b>Metering</b> for usage models, with events recorded so that a retried request never charges twice.</li><li><b>Invoices and tax</b>, including VAT, credit notes and failed payments.</li><li><b>Grandfathering</b>, so customers on an old price keep it until you decide to move them.</li></ol>
<p>Billing providers such as Stripe, Paddle and Chargebee cover much of this. Your product still has to tell them what changed and enforce the entitlements itself. The detail is in <a class="link" href="https://flavoureak.co.uk/blog/billing-for-saas/">billing for SaaS</a>.</p>
<h2>How to choose a pricing model, and when to change it</h2>
<h3>Choosing</h3>
<p>Start from the value metric, the unit that grows as the customer gets more from your product. For a job management tool it might be users or active jobs. For a document tool it is documents. Then check four things.</p>
<ul><li><b>Cost per customer.</b> If some customers will cost you far more to serve than others, include usage charges or limits.</li><li><b>How buyers budget.</b> Companies that set annual budgets prefer a predictable figure.</li><li><b>Who signs up.</b> Self-serve buyers need a price they can see and pay by card. Larger buyers expect to talk to someone and sign a contract.</li><li><b>What you can build now.</b> A model your billing code can’t support yet becomes a manual invoicing job every month.</li></ul>
<h3>Changing</h3>
<p>Review pricing when you have enough paying customers to see patterns. Signs it needs to change include customers on the cheapest plan who use the product more than those on the top plan, frequent requests for exceptions during sales calls, and accounts whose running costs exceed what they pay.</p>
<p>Changing prices for new customers is a change to your website and plan catalogue. Changing them for existing customers depends on your terms. Give clear notice, explain what changes and why, and consider keeping early customers on their price for a set period. On annual contracts, a new price usually applies at renewal. Check what your terms allow, and take legal advice if you are unsure.</p>
<div class="callout"><b>Planning a SaaS?</b> <a href="https://flavoureak.co.uk/founders/">See how we scope an MVP</a>, including the plans and billing rules for the first release.</div>
<h2>Questions people ask</h2>
<h3>What is a value metric in SaaS pricing?</h3><p>A value metric is the unit you charge for that grows as the customer gets more value, such as users, active projects, documents processed or orders. A good value metric is easy for the customer to understand and predict. Ideally it also moves in step with your own costs.</p>
<h3>Should a SaaS publish its prices?</h3><p>For products that small businesses sign up to on their own, published prices remove a step and build trust. Products sold to large organisations often publish starting prices and quote the top tier, because those contracts include terms such as single sign-on, support levels and invoicing that vary by customer.</p>
<h3>Is usage-based pricing harder to build than per-seat pricing?</h3><p>Yes. Usage pricing needs reliable metering, protection against counting the same event twice, running totals the customer can see and invoices calculated after the period ends. Per-seat pricing only needs the seat count kept in step with the user list, which billing providers support directly.</p>
<h3>Should I offer annual plans?</h3><p>Annual plans bring cash in up front, and many B2B buyers prefer one invoice a year. They are usually offered at a discount to twelve monthly payments. Count an annual plan as one twelfth of its value in monthly recurring revenue, as explained in <a class="link" href="https://flavoureak.co.uk/blog/saas-metrics/">SaaS metrics</a>.</p>]]></content:encoded>
</item>
</channel>
</rss>
