AI assistants send referral traffic. By default GA4 scatters it across "Referral" and "Direct," which makes it invisible in exactly the reports where you would look for it.
Separating it out takes twenty minutes and gives you a number worth tracking — provided you are clear about what it measures.
The referrers
The main sources, as they appear in referrer data:
| Referrer | Product |
|---|---|
chatgpt.com, chat.openai.com | ChatGPT |
perplexity.ai, www.perplexity.ai | Perplexity |
gemini.google.com | Gemini |
copilot.microsoft.com, bing.com/chat | Microsoft Copilot |
claude.ai | Claude |
you.com | You.com |
The list changes. Check your full referrer report periodically for hostnames you do not recognize rather than relying on a fixed list.
Building the segment
In GA4, create a comparison or audience filtered on Session source matching a regex:
chatgpt\.com|chat\.openai\.com|perplexity\.ai|gemini\.google\.com|copilot\.microsoft\.com|claude\.ai|you\.comBetter: create a custom channel group. Admin → Data settings → Channel groups → create a new group with an "AI Assistants" channel defined by that regex on Source. New channel groups apply going forward and, in most properties, retroactively to the data retention window. This puts AI traffic in the standard acquisition reports rather than in a segment you have to remember to apply.
In Looker Studio, the same regex as a filter on Session source works directly against the GA4 connector.
The measurement caveats, which are substantial
Undercounting is guaranteed. Some AI interfaces strip referrer headers, some route through link wrappers, and mobile app traffic frequently arrives with no referrer at all. Traffic that loses its referrer lands in Direct. Your measured AI referral number is a floor, not a total.
Referrals are a small fraction of exposure. The much larger effect is being cited without a click — your brand named in an answer the user acts on. That is real exposure and it does not appear in analytics at all. Reporting referral sessions as "our AI search performance" understates it in a way that is hard to correct for and easy to forget.
Attribution is unresolved. A user who reads about you in an AI answer, then searches your brand name the next day, converts as branded organic. The AI interaction caused it and gets no credit. This is the same last-click problem that has always distorted upper-funnel measurement, and it applies with full force here.
What the data does tell you
Despite all of that, the segment is worth building, because a few things it measures are reliable:
Direction over time. Even a systematically undercounted number tracks its own trend accurately, provided the undercounting is roughly stable.
Which pages get cited. Landing page data within the AI segment tells you which content these systems are pointing people at. That is directly actionable — it tells you what kind of content earns citations from your site specifically, and you can write more of it.
Engagement quality. Compare engagement rate, pages per session, and conversion rate for AI referrals against organic search. Many sites find AI referral traffic engages at least as well, which makes sense: the user arrived after reading a summary and clicked because they wanted more depth.
Which assistants matter for you. The distribution varies significantly by industry and audience. Knowing whether your traffic comes from Perplexity or ChatGPT tells you where to focus manual citation checks.
Reporting it honestly
The framing that survives scrutiny:
AI assistant referrals were 1,840 sessions this quarter, up from 1,190 last quarter. This is a floor — some AI traffic arrives without a referrer and is counted as Direct — and it excludes the larger effect of being cited without a click, which we cannot measure. Engagement rate on this traffic is 68% versus 54% for organic search.
That paragraph gives a stakeholder the number, its direction, its known limitations, and a genuine quality signal. It does not claim to measure something it cannot.
What to pair it with
Because referral data misses citation-without-click entirely, pair it with manual citation tracking: a fixed list of questions your customers ask, run against the major assistants monthly, recording whether you are cited. Small sample, manual, imperfect — and the only view you have of the part of the funnel analytics cannot see.
Two imperfect measurements pointing the same direction is a considerably stronger basis for a decision than one precise measurement of the wrong thing.