How to track AI traffic in GA4 — and what it misses.

GA4 already records visits from ChatGPT, Perplexity, and Gemini — it just doesn't label them. A click from ChatGPT arrives as source chatgpt.com, medium referral, filed under the generic Referral channel with every other link on the web. Fifteen minutes of setup — one referrer pattern, one exploration report, one custom channel group — turns that into a per-engine view. This tutorial walks through the exact steps, then covers what most tutorials skip: the documented majority of AI-driven visits that no analytics setup can attribute, and what it takes to measure those.

What GA4 does with an AI visit today

As of this writing, GA4's default channel definitions predate answer engines — there is no “AI” category out of the box. So AI-assistant visits scatter into three buckets, each hiding them a different way:

• Buried in Referral. Clicks from chatgpt.com, perplexity.ai, gemini.google.com, claude.ai, and copilot.microsoft.com land as ordinary referrals, mixed in with directories, blogs, and partner links.
• Blended into Organic Search.Clicks from Google AI Overviews carry a Google referrer, so they're indistinguishable from classic blue-link clicks in a source/medium report. Copilot answers inside Bing have the same problem — the source is just bing.com, and only the URL path would tell you it was the chat surface.
• Stripped to Direct. A large share of AI-assistant visits arrives with no referrer at all — more on that measured share below, because it changes how you should read everything else on this page.

The setup below fixes the first bucket completely and the second partially. Nothing inside GA4 fixes the third — which is why measurement in the AEO playbook starts from the answers engines give, not from analytics. Both halves matter; this page builds the analytics half properly.

Step 1 — the AI referrer pattern

Everything that follows runs on one regular expression matched against the session source. This list covers the major assistant domains, including the legacy hostnames older links still carry:

chatgpt\.com|chat\.openai\.com|perplexity\.ai|gemini\.google\.com|bard\.google\.com|copilot\.microsoft\.com|claude\.ai

Two practical notes before you paste it. First, GA4 offers both full and partial regex matching depending on where you use it — pick the partial option where it exists, or wrap the pattern as .*(pattern).* for a full-match field. Second, this list decays: engines rename domains (ChatGPT moved from chat.openai.com to chatgpt.com; Gemini absorbed bard.google.com) and change how they tag outbound links. Recheck the pattern quarterly, or use a tool that maintains the classification for you.

Deliberately absent: bing.com and google.com. Matching those hostnames would swallow your ordinary search traffic, and the AI-versus-classic distinction on both lives in the URL, not the domain — the Copilot chat path on Bing, and signature parameters like srsltid on AI Overview links. A source-level filter simply cannot make that cut; we come back to it in the blind-spot section.

Step 2 — a five-minute exploration report

The fastest way to look at AI traffic — today, without changing any configuration — is a free-form exploration:

1. Open Explore and start a blank exploration. Set the date range to at least the last 30 days — AI referral streams are thin, and short windows read as noise.
2. Import the dimensions Session source and Landing page + query string, and the metrics Sessions, Engaged sessions, and Key events.
3. Put Session source in rows and the three metrics in values.
4. Add a filter: Session source → matches regex → paste the pattern from Step 1.
5. Read the table once, then add Landing page + query string as a second row dimension. This is the payoff view: the exact pages each engine links when it cites you.

The landing-page cut matters more than the totals. It shows which of your pages the engines actually read and link — and therefore which content styles are earning citations. If the list is empty or thin, that's a finding too: engines may be naming you without linking, or not naming you at all. The five-prompt manual check settles which, in about ten minutes.

Step 3 — make it permanent with a custom channel group

Explorations are for looking; a custom channel group puts an “AI” row in your everyday acquisition reports:

1. In Admin, find Channel groups (under the data-display settings) and create a new group — leave the default group untouched.
2. Add a channel named AI with the condition: source matches the Step 1 pattern.
3. Drag it above Referral. Channel rules evaluate top-down, and the AI sources would otherwise be claimed by Referral first.
4. In acquisition reports, switch the channel dimension to your new group. AI now sits beside Organic Search, Direct, and Referral as a first-class row.

One honest caveat: a channel group changes labeling, not detection. It reclassifies the visits GA4 could already attribute — it cannot recover the ones whose referrer never arrived.

What the numbers mean

Expect a thin stream. Even for businesses that AI engines cite regularly, referral counts are small next to organic search — the stream is growing fast (ChatGPT referral traffic grew 206% across 2025 — Semrush, via Omnibound), but it starts from a low base. What makes it worth the setup is who these visitors are: they arrive pre-briefed, having already read a synthesized comparison before clicking.

So read your new AI row for three things, in order: which engines appear at all (that's your citation footprint), which pages they land on (that's what the engines consider quotable — the ChatGPT playbookcovers why certain page shapes win), and whether key events fire (that's the quality claim above, tested on your own funnel).

The blind spot: what GA4 cannot show you

Here is the part that changes how you read everything above. The visits your new AI channel catches are a minority of the AI-driven visits that actually occur:

Stripped: the referrer never arrives

Assistant apps and in-app browsers frequently drop the referrer in the app-to-web handoff, and readers copy links out of answers and paste them into a fresh tab. Both behaviors land in GA4 as Direct. The visit is real, the AI origin is gone, and no filter downstream can resurrect it.

Blended: AI clicks wearing search clothes

Clicks from Google AI Overviews arrive with Google referrers and file into your search traffic. Signature URL parameters — srsltid on result links, udm=50 on some AI Mode links — are among the few distinguishing marks, and reading them takes a script on your own site, not a source/medium report.

Silent: the answer that never links

The largest influence channel produces no click event anywhere. An engine names your business in an answer — often with no link — the reader closes the tab, and hours later types your name into a browser. GA4 books a branded search or a direct session. Multiply that across every answer where you or a competitor gets named, and the gap between AI's real contribution and the AI row in your reports becomes the story. This is why proving AEO's value can't rest on referral counts alone.

What closes the gap

The blind spot has a shape, so the fix has a shape: measure both ends of the funnel, and stop expecting either end to carry the whole answer.

• Upstream — measure the answers themselves.Ask the engines your buyers' questions on a schedule and record who gets named and cited. This catches the silent channel that no analytics tool can: the recommendation is observable even when the click never happens. If the answers don't name you, start with why before spending another hour on attribution.
• Downstream — classify at your own front door.A first-party script sees the full landing URL, so it can read the signatures a source/medium report can't — srsltid on AI Overview clicks, the chat path on Bing — and it can be updated centrally when engines change their tagging, instead of waiting for your quarterly regex audit.

Disclosure: this tutorial is published by AEO Action, and the paragraph below describes our own product — weigh it accordingly. Both halves above exist as product surfaces. The scans query the engines daily — ChatGPT and Gemini through the live consumer interface — and record who gets named. The AEO Pixel classifies each visit on the server from its referrer and URL signature (it deliberately doesn't trust client-side labels), keeping a 30-day per-engine view with AI landing paths and conversion events. And a GA4 connection through Google sign-in reads your existing property and buckets its source/medium pairs into the same engine categories — the Step 1 pattern, maintained for you. None of that recovers a visit whose referrer was stripped, and nothing downstream measures the reader who never clicked. The scan is the primary instrument; analytics corroborate it. If the question behind this setup is really whether to run the work in-house or pay someone, what AEO services cost and cover lays out both columns with the reporting each one actually hands you.