How to track AI search traffic — and why most of it never reaches your analytics.
Your dashboard says ChatGPT sent 40 visits last month. That number is real, and it is not the answer to the question you were asking. Every visit that starts in an AI answer lands in one of three states — and only one of them is a number you can read off a report.
AI search traffic arrives in three states. Attributable — a click carrying a recognisable referrer like chatgpt.com or perplexity.ai, which your analytics files under Referral next to every unrelated site. Blended — a real AI-driven visit whose referrer was stripped by an assistant app, or an AI Overviews click that Google reports inside its ordinary Search totals; it lands in direct or organic and is indistinguishable there. Not attributable — the answer was the destination, so no visit exists at all. The third state is the largest, and it is the one that decides whether you get shortlisted. Which is why the honest measurement isn't a traffic report: it's asking the engines what they say about you, and keeping unknown separate from zero.
Analytics was built for a web where every visit had a source. A person searched, clicked a blue link, and arrived carrying a referrer that said where they'd been. AI search breaks that chain in three separate places — and a dashboard that only counts arrivals can't tell you which break you're looking at.
So the "ChatGPT sent 40 visits" number in your report is doing something quietly dishonest: it's presenting the one slice that happens to be countable as though it were the whole. The right question isn't how much traffic did AI send. It's how many decisions did AI participate in — and those two numbers are further apart than any dashboard admits.
The three states of an AI visit
Before you can measure anything, you have to accept that "AI traffic" isn't one thing. A person who asks an assistant about your category and ends up as your customer can reach you through three completely different paths — and your analytics has a different relationship with each:
Attributable
The answer linked out, they clicked, and the visit arrived carrying a referrer you can recognise. This is the slice everyone reports, because it's the slice that counts itself.
Blended
A genuine AI-driven visit whose origin was erased in transit — an app that sends no referrer, or an AI Overviews click that Google reports as ordinary Search. Real traffic, filed under the wrong name.
Not attributable
The answer was the destination. Your brand was named, compared, recommended or skipped, and no visit followed — so there is nothing in analytics to find.
Almost every reporting mistake in AI search comes from collapsing those three into one. Counting only state one and calling it "AI traffic" understates the channel. Counting only state one and concluding AI doesn't matter gets the direction of the business wrong entirely — because state three is where shortlists are made.
| State | Where it lands | What you can honestly say |
|---|---|---|
| Attributable | Referral, under an engine hostname — mixed in with every other referring site | A hard floor on AI-driven traffic. Never the ceiling. |
| Blended — no referrer | Direct, alongside bookmarks, typed URLs, and links from email and chat | A pattern you can read, not a number you can prove. |
| Blended — AI Overviews | Organic search; Google folds it into the overall Web totals in Search Console | Pressure shows up as falling click-through on steady impressions. |
| Not attributable | Nowhere. No session was ever created | Only measurable by asking the engine what it said. |
| Unknown is not zero. Three of these four rows are invisible to a traffic report, and none of them are absences. | ||
What AI referral traffic actually looks like
Start with the part that does work. When an AI answer links out and someone clicks, the visit usually arrives carrying the assistant's own hostname as the referrer — chatgpt.com, perplexity.ai, and their equivalents across the other engines.
Two things about that set of hostnames matter more than the hostnames themselves. First, any list of them is incomplete by construction — surfaces launch, get renamed, and move between domains, an assistant embedded in a browser or an operating system may not present as an engine at all, and some engines split their traffic across more properties than you'd guess. Keeping a current list is maintenance, not a one-time lookup, which is a large part of why we don't publish ours.
Second, none of them are a channel. Out of the box, GA4 has no AI grouping: they land in Referral next to a partner blog and a forum thread, so nothing in the standard reports adds them together. Isolating them is something you have to build and then keep rebuilding — and a measurement that leans on a parameter the operator controls, like the tracking tags some engines currently append to outbound links, is one release note away from appearing to collapse overnight.
The traffic hiding in “direct”
A referrer is a courtesy, not a guarantee. Native desktop and mobile assistant apps frequently open links without sending one; some surfaces route the click through a redirector that drops it; privacy settings strip it. Analytics still has to file the visit somewhere, so it files it under direct — the bucket that already holds bookmarks, typed URLs, QR codes, and every link opened from an email or messaging app.
You can't recover those sessions individually, and no amount of staring at a direct-traffic total will separate them out. What you can do is look for direct behaving in ways real direct traffic doesn't — arrivals that only make sense as the answer to a question somebody just asked, rather than as somebody returning to a site they already knew. Classic direct tracks brand familiarity; assistant traffic doesn't, and the divergence is legible if you know which comparisons to run.
We deliberately don't publish the full set of tests we use here, because they're exactly what a look-alike tool would copy and then apply badly. The principle is the part that generalises: treat it as a signal and label it as one. "Direct is up 30% on comparison pages, consistent with assistant traffic we can't attribute" is a useful, defensible sentence. "AI drove 30% more traffic" is a number you invented.
Why AI Overviews clicks never show up as AI traffic
Google is the biggest AI surface most brands touch, and the hardest to isolate — by design. AI Overviews and AI Mode are assembled on a Google results page, so a click from one carries a Google referrer like any other organic click. There's no separate hostname to segment.
Search Console doesn't close the gap either: Google includes AI Overviews and AI Mode activity in the overall Web search type in the Performance report rather than breaking it out as its own filter. You get one blended organic number that contains both a traditional ten-blue-links click and an answer-panel click, with no way to split them.
What you can see is the pressure. When an answer gets assembled above your result, the query still shows your page — so impressions hold — but fewer people travel any further. Impressions and clicks coming apart is the shape to watch for, though which queries it shows up on, and how to tell it apart from an ordinary ranking slip, is the part that takes judgement rather than a filter. Either way it's inference, not measurement, and it's worth saying so out loud when you report it.
The visits that never happen
Now the state that matters most and reports least. Someone asks an assistant which firms in their city handle a problem they have. They get three names, a sentence about each, and a reason to prefer one. They never click anything. Two weeks later one of those three gets a call.
Nothing in that sequence appears in analytics — not the impression, not the comparison, not the recommendation, and not the reason you weren't in it. This is the part of the funnel that AI search moved, and it's precisely the part a traffic report can't observe. If your brand is missing, the number you see isn't a low number. It's no number, and it looks identical to a quiet month.
Which is why the measurement that matters isn't downstream of a click at all. You have to ask the engines directly — the questions your buyers actually ask — and record what comes back: whether you were named, whether you were cited with a working link, whether what was said about you was accurate, and who was named instead of you. That's a measurement of the answer itself, and it works whether or not a visit follows.
The click is the last honest thing AI search will hand you, and it's already the smallest part of the story.
What a real AI-traffic measurement covers
Put together, a measurement you can defend in a board meeting has to hold four things at once:
- The referred slice, named and maintained. The engine properties pulled out of the referral pile and tracked as their own line — with the list revisited as surfaces appear and rename themselves, because a stale list quietly reads as a decline.
- The blended slice, labelled as inference. The direct and organic patterns that assistant traffic leaves behind, reported as signals rather than promoted to facts.
- The answer itself, on a cadence. What the engines say about you when someone asks — named, cited, described accurately, or replaced by a competitor. This is the only view of state three that exists.
- Access, underneath all of it. None of the above is meaningful if engines can't read you in the first place. A quiet crawler block looks exactly like a content problem in every report you'll run.
And it has to repeat. A single snapshot tells you where you stand; it can't tell you whether you're gaining or losing, which is the only question anyone actually asks. Answers are regenerated per query, models get updated, competitors publish — the same prompt can return a different set of names next month with nothing on your side having changed.
The referral and inference work lives in your analytics, and AI Search Console is where we put the AI side of it. The part that can't be done from analytics — asking ten engines what they say about you and diffing it against last month — is what AI Pulse runs on a schedule.
Key takeaways
- AI visits arrive in three states — attributable, blended, and not attributable — and only the first one counts itself. Reporting that slice as "AI traffic" understates the channel every time.
- There is no AI channel by default. Engine referrals sit in Referral next to unrelated sites, assistant apps that strip the referrer land in direct, and Google folds AI Overviews into its ordinary Search totals.
- Zero-click is the point, not a flaw. The answer is the destination, so the decision that matters most leaves no trace in analytics — which means the only way to see it is to ask the engines what they say about you.
- Unknown is never zero. An unattributed visit is a measurement you didn't take; treat it as a signal, label it as one, and re-measure on a cadence so you're reading a trend instead of a snapshot.
Questions about tracking AI traffic.
Does ChatGPT send referral traffic?
Yes — when an answer links out and the person clicks. Those visits usually arrive with a chatgpt.com referrer, and some engines currently append tracking tags of their own, which makes those clicks easier to spot for as long as the operator keeps doing it. But referred visits are only the visible slice. Answers delivered in the desktop and mobile apps frequently arrive with no referrer at all and land in direct, and the majority of AI answers about a brand end without any click, because the answer was the destination. Treat ChatGPT referrals as a floor on your AI traffic, never as the total.
How do you see ChatGPT and Perplexity traffic in Google Analytics?
By source, not by channel. Out of the box, GA4 has no AI channel: visits from the engines fall into Referral alongside every unrelated site, so nothing in the standard reports adds them up. Isolating them means maintaining your own list of engine properties and grouping on it — and maintaining is the operative word, because surfaces launch and rename themselves, and a list that goes stale reads as a decline that never happened. Even done well, the number you get is the referred slice only: app sessions with a stripped referrer will still be sitting in direct.
Can you track AI Overviews clicks in Google Search Console?
Not separately. Google includes clicks and impressions from AI Overviews and AI Mode in the overall Web search type in the Performance report, and doesn't break them out as their own filter or search appearance. There's no distinct referrer either, because the click comes from a Google results page. In practice you can infer pressure rather than measure it — impressions and clicks coming apart is the shape to watch — but telling that apart from an ordinary ranking slip takes more than a filter, and neither one tells you whether you're still inside the answer.
Why does AI traffic show up as direct?
Because a referrer is a courtesy, not a guarantee. Native desktop and mobile assistant apps often open links without sending one, some surfaces route clicks through a redirector that drops it, and privacy settings strip it. Analytics has to file a visit with no referrer somewhere, so it files it under direct — the bucket that also holds bookmarks, typed URLs, and links from email and chat apps. You can't pull those sessions back out individually, but direct that behaves unlike real direct traffic — arrivals that only make sense as an answer to a question just asked, rather than a return to a known site — is a readable fingerprint of assistant traffic you're not being credited for.
Is AI search traffic worth tracking if the volume is small?
Yes, because the visits arrive late in the decision rather than early. Someone who clicks through from an AI answer has usually had the comparison made for them already and is checking a shortlist rather than starting one, which is why these sessions tend to convert at a different rate than generic organic traffic. The other reason is direction of travel: the volume is small and growing, and the brands being named now are the ones models and retrieval indexes will keep naming. Judge it on the quality of the sessions it produces, not on its share of a dashboard.
How do you measure AI visibility if you can't measure all the clicks?
You measure the answer, not just the click. Ask the engines the questions your buyers actually ask, and record whether you were named, whether you were cited with a link, what was said about you, and who was named instead — that's a measurement of the answer itself, and it works even when no visit follows. Read it next to your referral and direct trends: rising citations with flat referrals is a real, healthy pattern in a zero-click surface, and falling citations explain a traffic dip that analytics alone can only describe. A monitored scan like AI Pulse runs that check monthly across all 10 engines and reports what changed. Whatever you use, keep unknown separate from zero — a visit you couldn't attribute is a measurement you didn't take, not an absence of influence.
Related posts & guides.
See the AI traffic your analytics can't count.
Start with the free audit to see where you stand today. Then let AI Pulse measure the part no dashboard reaches — one scan a month across all 10 AI engines, a single GEO score, what changed since last cycle, and a ranked list of fixes in order of impact. $499/month, and you'll know the month you stopped being named instead of the quarter after.