AI Traffic Analytics: Track ChatGPT and AI Referrals in GA4

To track AI traffic in GA4, read the built-in AI Assistant channel (Google added it on 13 May 2026) and build a custom channel group with a hostname regex of your own, then compare the two. Either way the result is a floor: GA4 can credit an assistant only when a click carries a referrer or UTM tag, and many AI answers never produce a click.

What counts as AI traffic?

AI traffic is a session that starts when someone clicks a link inside an assistant's answer (ChatGPT, Perplexity, the Gemini app, Copilot or Claude) and GA4 can attribute it to that assistant. It is an ordinary referral: the browser reports a referrer, the address of the page that held the link, and a UTM parameter in the URL can name the source.

Two AI-related things are not AI traffic:

• Google AI Overviews and AI Mode. Google counts their clicks as Organic Search, not AI Assistant, and Search Console files them under the Web search type (Google Search Central), so neither tool breaks them out. For the click impact, see AI Overviews and your traffic.
• Crawler and fetcher hits. ChatGPT can fetch a page for a user with the ChatGPT-User agent (OpenAI). That is the assistant reading your page, not a person visiting; count it from server logs, never as AI sessions.

How do you see AI traffic in GA4?

Open Reports → Acquisition → Traffic acquisition and read the AI Assistant channel, then build a custom channel group that matches assistant hostnames with a regex you control. Run both for a month; where they disagree, one of them is missing something.

Google's channel. Its rule sets the medium to ai-assistant when the referrer matches a recognized assistant. As of October 2026 Google names ChatGPT, Gemini, Claude, DeepSeek, Copilot and Grok as examples, does not name Perplexity, and publishes no full list. To audit it for dates after 13 May 2026, switch Traffic acquisition to Session source / medium and search for chatgpt, perplexity, gemini, copilot and claude: any hit whose medium is not ai-assistant was missed.

Your channel, step by step.

1. Open Admin → Data display → Channel groups and choose Create new channel group (Google's steps).
2. Add a channel named differently from Google's, such as AI referrals (regex).
3. Add a condition group: Source, matches regex, then paste the pattern below.
4. Click Reorder, drag the channel above Referral, then Apply and Save group.
5. In Traffic acquisition, switch the primary dimension to your new group.

^(.*\.)?(chatgpt(\.com)?|chat\.openai\.com|perplexity\.ai|gemini\.google\.com|bard\.google\.com|copilot\.microsoft\.com|claude\.ai|deepseek\.com|grok\.com)$

Unlike Google's list, this one is yours to extend (review your Session source list quarterly) and retroactive: custom channel groups apply to past data, and two older hostnames (chat.openai.com, bard.google.com) cover earlier months. GA4 regex is full-match and case-sensitive by default, so the pattern covers the whole value, in lowercase, with an optional prefix such as www.. We checked it with an RE2 engine (the syntax Google's regex help points to) against 15 hostnames it should catch and 27 look-alikes it should not; all 42 behaved. We have not run this setup in a live property, so check the first week's rows.

The ChatGPT tag. Seer Interactive's tests (as of June 2025) found ChatGPT tends to add utm_source=chatgpt.com to links it supplies as sources, but not to every link. The tag travels in the URL, so a Source rule catches those visits even when the referrer is stripped.

Why is AI referral traffic undercounted?

Because GA4 can credit an assistant only for clicks that arrive with a referrer or UTM tag. Many AI answers produce no click, some clicks lose their referrer on the way, and the visit an answer causes may come days later. Read the AI line as a floor, never the total.

• No click happens. In Pew Research Center's March 2025 study of 900 U.S. adults' browsing, people clicked a link inside a Google AI summary in 1% of visits to pages showing one, and a regular result in 8% of visits with a summary versus 15% without. That is Google; we found no public click rates for ChatGPT, Claude or Perplexity answers.
• The referrer is dropped. Browsers send a referrer by default, but a link's rel="noreferrer" or a page's Referrer-Policy can omit it. Google lists redirects, missing UTM parameters, URL shorteners and ad blockers behind (direct) / (none); Seer adds in-app browsers and copied links.
• The visit comes later. Someone who reads your name in an answer and searches for you tomorrow is recorded as brand search or Direct.
• You never see the question. By default only the origin travels on cross-origin HTTPS links, so GA4 shows which assistant sent a visit and where it landed, not the prompt.

Two checks probe the leak. In a private window, ask each assistant a question about your own site, click your citation on desktop and phone, then read the first-user-source card in Reports → Realtime (last 30 minutes). Then break (direct) / (none) down by landing page: few people type long article URLs, but newsletters, chat apps and bookmarks land there too, so treat it as a pool of suspected dark AI traffic, never a count.

Which numbers should you report?

Report sessions, engaged sessions, key events and revenue for the AI channel, plus the rule that produced them, and label every figure a lower bound. Quote rates only when sessions are numerous enough to trust.

Metric	How to read it	Caveat
Sessions	AI-channel sessions in a stated date range	A floor, not a total
Engaged sessions	Over 10 seconds, or a key event, or 2+ page views; engagement rate = engaged ÷ sessions	A short visit can mean the answer was complete
Key events	Actions you marked as important; key event rate = sessions with one ÷ sessions	Say which event counts, such as sign_up
Revenue per session	Total revenue ÷ sessions	Not lifetime value
Landing pages	Where AI sessions start	A proxy for the question, not the question
Share of traffic	AI sessions ÷ all sessions; same for revenue	Observed share, not AI's influence

Stay in Traffic acquisition, which is session-scoped. User acquisition credits a user's first source, so someone first found through organic search who returns from ChatGPT counts as organic.

Do not import someone else's conversion multiple; compute your AI key event rate beside Organic Search over the same dates. Below about 100 AI sessions, quote counts instead of rates, because one extra purchase moves the rate by a full percentage point. (That threshold is our rule of thumb, not Google's.)

If checkout or login runs on another domain, add it under List unwanted referrals so it does not take credit for the visit. Never add an assistant's domain there; you would erase what you are measuring.

Save each export with its date, property, channel rule and filters, because rules move. Footnote every report: AI Assistant excludes Google AI Overviews and AI Mode; visits with no referrer or tag appear as Direct.

How does AI traffic relate to AI visibility?

AI traffic counts visits that already happened. AI visibility is whether assistants name you when buyers ask. Traffic is a lagging, partial readout of visibility, so pair it with direct testing of your buyers' questions.

In AEOeye's audit records (127 completed audits, 26 June to 7 September 2026), the audited brand was named in 157 of 892 engine answers to buyer questions: 17.6%. The brands are self-selected with repeat audits, engine samples are unequal and not comparable, and four of five engines have under 10 mentions, so read this as scale, not a market estimate; the aggregates are in our benchmark file. Even so, most answers did not name the audited brand. A handful of AI sessions can mean you are among the few named, while a zero can mean you were named without a click or not named at all, and GA4 cannot say which.

To find out which, ask your buyers' questions in ChatGPT, Gemini, Claude, Perplexity and Google AI and record whether you or a competitor is named: AI citation tracking in its simplest form. AEOeye's free audit runs three questions in ChatGPT and shows the answers in full; as of October 2026 the $29 report covers ten questions across all five engines. For daily monitoring across hundreds of prompts, trackers such as Profound, Peec AI and Otterly are built for it; we are not.

Microsoft offers one first-party view: AI Performance in Bing Webmaster Tools (introduced in public preview in February 2026) counts how often your pages appear as sources in Copilot, Bing's AI summaries and select partner integrations, not clicks. Set it beside your Copilot sessions in GA4 to see the gap.

So optimize for being named and treat traffic as the confirmation. A flat AI line can mean the work has not landed yet, not that AI does not matter to you; the full metric set is in Measuring AI Visibility.