Build log/GEO

How to measure traffic from ChatGPT, Perplexity and AI Overviews

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.

The short answer

  • 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.
  • 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.
  • 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.

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.

What you can and cannot see

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.

SourceWhat it showsWhat it misses
Analytics referrersClicks from AI tools that pass a referrerClicks from apps or settings that strip it
UTM parametersClicks from engines that tag their linksEngines that don’t tag, and copied links
Server logsCrawlers and fetchers reading your pagesWhether anyone saw or clicked the answer
Search ConsoleGoogle search clicks and impressions, AI features includedA separate figure for AI Overviews
Prompt trackingWhether answers mention or cite youHow many people asked those questions

Referrers in analytics

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.

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 chatgpt|openai|perplexity|copilot|gemini|claude on the session source is a reasonable start. Review it every few months, because new tools and domains appear.

The limits of referrers

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.

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.

UTM parameters some engines add

A UTM parameter is a tag on the end of a link, such as ?utm_source=chatgpt.com, that analytics tools read to attribute a visit. ChatGPT adds utm_source=chatgpt.com 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.

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.

Server logs and AI crawlers

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.

CompanyAgents to look for
OpenAIGPTBot, OAI-SearchBot, ChatGPT-User
AnthropicClaudeBot, Claude-SearchBot, Claude-User
PerplexityPerplexityBot, Perplexity-User
MicrosoftBingbot, which feeds Bing and Copilot
GoogleGooglebot, which feeds Search, including AI Overviews

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.

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 AI crawlers and robots.txt.

What Google Search Console separates and what it doesn’t

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.

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 Google AI Overviews covers what they change for business sites.

Tracking mentions in answers

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.

  1. Write a list of 20 to 50 questions buyers ask before they choose a supplier like you, in their own words.
  2. Run each question in the engines your customers use, signed out where possible, on a fixed day each month.
  3. Record whether your business is mentioned, whether your site is cited, and which other sources are cited.
  4. Save the answer text and the date, because answers change from one run to the next.
  5. Compare month on month, noting which of your pages are cited and which competitors appear in your place.

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.

Putting it into one monthly report

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.

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.

Want AI search measured properly? Our GEO guide explains how we set up tracking and content for answer engines.

Questions people ask

Why does ChatGPT traffic show as direct in my analytics?

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.

Can I see AI Overview clicks in Google Search Console?

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.

Which AI crawlers should I look for in my logs?

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.

Is it worth paying for an AI visibility tool?

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.

Written by Flavoureak

We build and run Estimark, CurioKids and CalcStack, and we build web and mobile software for other businesses. Start a project.