Tracking AI Referral Traffic in GA4 Before Your Attribution Model Notices It Exists

Somewhere in your GA4 property right now there are sessions that began with an AI assistant recommending you. Most teams have never isolated them, so the channel that increasingly starts buying journeys reports as a rounding error inside "Referral" or vanishes into "Direct." Fixing the measurement takes an afternoon.

Where AI traffic hides

When a user clicks a citation in ChatGPT, Perplexity, Gemini, or Copilot, the session usually arrives with a referrer from that platform's domain. In GA4 these land in Referral, mixed in with every other site that links to you. Some arrivals (apps, certain privacy configurations) strip the referrer entirely and land in Direct, which means whatever you measure is a floor, not a ceiling.

The segmentation

Build a custom channel group (or an exploration segment) that matches session source against the known AI referrer domains: chatgpt.com and chat.openai.com, perplexity.ai, gemini.google.com, copilot.microsoft.com, claude.ai, and you.com cover the bulk. Add new ones as they appear in your referral report; the population changes quarterly.

Two implementation notes that save grief later:

• Use contains-style matching rather than exact strings, because subdomains and path variants multiply.
• Backfill your analysis with the same definitions in explorations, since channel groups apply going forward. The historical data is already there; only the labeling is new.

Google's AI Overviews remain the blind spot: clicks from Overviews arrive tagged as ordinary google organic, indistinguishable at the session level. Directional inference comes from Search Console (impression and click patterns on queries where you can verify an Overview exists), not from GA4.

Reading the numbers honestly

Expect the volume to look small: low single-digit percentages of sessions for most sites. Then look at quality metrics before dismissing it. Across the properties we monitor, AI-referred sessions consistently show stronger engagement and conversion behavior than average organic, which matches the mechanism: these visitors arrived pre-sold by a recommendation, often deep in evaluation. A session that starts from "ChatGPT told me you were the best option" is not a top-of-funnel click.

Also expect undercounting to be structural. The whole point of answer engines is that many users get their answer without clicking anything. Referral traffic measures the overflow, not the influence. This is why session data pairs with prompt-universe visibility tracking and a plain "how did you hear about us" field: the three together triangulate a channel that no single instrument sees fully.

The reporting move

Give AI referrals their own line in whatever dashboard leadership reads. Not because the number is big, but because its trajectory is the argument. A channel growing steadily quarter over quarter, with above-average conversion, purchased by zero ad dollars, tends to answer the "why are we investing in AEO" question before anyone finishes asking it.