In short
Traffic from AI assistants arrives with identifiable referrers, so it can be segmented in analytics like any other channel. Expect the volume to look small and the quality to be high, since visitors arrived after a system judged you a good answer. The larger and harder question is citation without click — being named in an answer that satisfies the user entirely, which no analytics tool records.
Segment it in analytics
Visitors arriving from assistants carry referrer values identifying the source, so the first step is straightforward: create a channel grouping or segment that captures the known AI assistant domains and treat it as its own channel.
Once segmented, compare it against your other channels on the metrics that matter — conversion rate, pages per session, and whatever your primary action is.
The pattern most sites see: low volume, unusually good engagement. That combination makes sense, since someone arriving from an AI answer was directed there by a system that judged your page a good response to their specific question. That is a stronger qualification signal than a keyword match.
Be aware referrer data is incomplete. Some assistants pass no referrer, some traffic appears as direct, and mobile apps often lose it. Treat the number as a floor.
Check your server logs too
Analytics misses what analytics cannot see. Server logs show two things worth knowing.
Which AI crawlers fetch your pages, and how often. This tells you whether you are being retrieved at all, which is upstream of being cited.
Which pages they fetch most. That is a signal about what content is considered relevant, and it frequently differs from what you assumed.
This also lets you distinguish training crawlers from retrieval crawlers by user agent, which matters if you are making decisions about blocking either.
The measurement problem nobody solves cleanly
Here is the honest part: most AI citations never produce a click.
Someone asks a question, the assistant answers using your page, cites it, and the user is satisfied. No visit, no analytics event, no record on your side. You were useful and you cannot tell.
That is a genuine measurement gap, and it will not be closed by better tracking, because the interaction happened somewhere you have no visibility into.
Which means anyone promising complete measurement of AI visibility is overstating what is possible. The realistic position is to measure what you can and be explicit about what you cannot.
Measuring visibility rather than traffic
Since clicks understate the picture, measure presence directly.
Ask the assistants. Take twenty questions your customers actually ask — the ones your content answers — and put them to the main assistants periodically. Record whether you are mentioned, whether you are cited with a link, and who else appears.
This is manual, and it is the most useful signal available. Doing it monthly gives you a trend, and it tells you which competitors are winning the answers you want.
Track the share. Across your question set, what proportion of answers cite you? That number moving is a real result, and it is the closest thing to a ranking equivalent in this channel.
Note the shape of the answer. Being one of three cited sources is different from being the only one, and being quoted directly is different from being listed at the bottom.
Tools exist that automate this. They are worth evaluating, with the caveat that answers vary between runs, so any single measurement is noisy and the trend matters more than the number.
What to do with the finding
If you are cited: identify which pages, and what about them made them citable. It is nearly always a page with a direct answer near the top, sourced figures, and clear structure — the things that earn citation.
If a competitor is cited instead: read their page against yours for that specific question. Usually they answered it more directly, or had a number you did not.
If nobody is cited and the assistant answers generically: that is an opening. A question with no good source is a page worth writing.
Reporting it honestly
This matters, because the temptation to overstate is real.
Report referral traffic as a floor, noting that attribution is incomplete.
Report citation share separately, with the method — how many questions, which assistants, when tested.
Do not present citation share as traffic. They measure different things, and conflating them produces a number nobody can act on.
Show the trend, not the snapshot. Individual measurements are noisy enough that a single reading means little.
Being straightforward that this channel is partly unmeasurable is more credible than a dashboard implying otherwise — and clients generally respond well to being told which numbers are solid and which are directional.
The practical routine
Segment AI referrers in analytics and watch conversion rate rather than volume. Check server logs quarterly for crawler activity. Run a twenty-question citation check monthly and record the share. Investigate the pages that get cited and the questions where competitors win.
That is perhaps two hours a month and gives you a defensible picture of a channel most sites are not measuring at all.
If you want a baseline citation check run against your actual customer questions, book a call.
Common questions
How do I track traffic from ChatGPT and other AI assistants?
Segment by referrer in your analytics, since assistants pass identifiable referrer values, and treat it as its own channel. Compare conversion rate rather than volume — the usual pattern is low traffic with unusually good engagement, because the visitor was directed by a system that judged your page a good answer.
Why does AI referral traffic look so small?
Partly because referrer data is incomplete — some assistants pass none, some traffic appears as direct, and mobile apps often lose it. More importantly, most citations never produce a click at all: the assistant answers using your page and the user is satisfied without visiting.
How do you measure AI citations that don't produce clicks?
By asking the assistants directly. Take twenty questions your customers actually ask, put them to the main assistants monthly, and record whether you are mentioned, whether you are cited with a link, and who else appears. Track that share over time — it is the closest equivalent to a ranking measure.
Can AI visibility be measured completely?
No, and anyone claiming otherwise is overstating what is possible. The interaction happens somewhere you have no visibility into, so citations without clicks leave no record on your side. Report referral traffic as a floor and citation share separately, with the method stated.
What should I do if competitors are cited instead of me?
Read their page against yours for that specific question. Usually they answered it more directly near the top of the page, or included a sourced figure you did not. If nobody is cited and the assistant answers generically, that is an opening — a question with no good source is worth writing for.
