You have probably had this argument. Someone pulls up the analytics, points at the AI referral line, and says the quiet part: this channel is a rounding error. And on the screen in front of them, it is. That is the problem with AI referral traffic attribution right now. The screen is honest and the screen is wrong.
Scrunch published panel research this month on how people actually behave after an AI conversation, and one pair of numbers should reset the whole conversation. Only 1.1% of publisher visits came in as AI referrals. But of the visits that happened after an AI conversation, roughly 75% arrived through direct navigation. About 9% came through traditional search.
Read those two sentences again. They are describing the same people, on the same day, doing the same thing. One number counts them and the other four numbers lose them.
The AI conversation happened. The visit happened. The connection between them evaporated somewhere in between, and what your dashboard recorded was somebody typing your name into a browser bar out of nowhere.
The 1.1% everyone is about to misread
The 1.1% is going to get quoted in a lot of meetings over the next month, and it is going to get quoted as evidence that AI search does not matter yet. I understand the appeal. It is a comfortable number.
It is also a measurement of one specific thing: how often a visit carried a referrer header from an AI product. That is not the same question as how often AI shaped a visit. Those two questions have been drifting apart for two years, and this data is the clearest picture yet of how far apart they have got.
| WHAT THE PANEL MEASURED | FIGURE | WHAT IT IS ACTUALLY TELLING YOU |
|---|---|---|
| Publisher visits arriving as AI referrals | 1.1% | The tracked channel. Small. |
| Post-conversation visits arriving direct | ~75% | The same behavior, filed elsewhere. |
| Post-conversation visits arriving via search | ~9% | Brand search, triggered by the chat. |
| Visit-likelihood lift in the following week | 20.5pp | The effect is delayed, not immediate. |
| Publisher clicks on news searches with AI Overviews | ~20% | Down from ~30% without them. |
There is a version of this argument that is too convenient, and I want to name it before somebody uses it on you. It goes: our AI numbers look bad, therefore the good numbers must be hiding in direct traffic, therefore the programme is working. That is not analysis. That is an alibi.
The difference between an alibi and a finding is whether you can show the mechanism. Here you can. Three quarters of post-chat visits arriving direct is a measured behavior, not an inference from a gap in a spreadsheet.
Where AI referral traffic attribution actually breaks
It breaks in four places, and only one of them is a tracking bug. That distinction decides whether this is a problem you can fix with engineering time or one you have to design your reporting around. Three of the four are user behavior, and user behavior does not care what your measurement plan assumed.
Notice that three of those four are behavior, not instrumentation. You cannot tag your way out of them. This is the same wall we hit when AI search reached its own not-provided moment, and the honest answer then is the honest answer now: some of this is permanently unmeasurable at the session level, so the measurement has to move up a level.
The lag is the finding
The 20.5 point figure is the one I keep coming back to, because it changes what kind of channel this is.
Readers who had a news-related AI conversation were 20.5 percentage points more likely to visit a news publisher in the following week than readers whose conversation was about something else. Not in the following minute. The following week.
That is not a referral channel. That is a demand channel. It behaves like a podcast ad or a well-placed mention in a newsletter: the effect is real, the timing is loose, and the click path is invisible. We have spent a decade building measurement for channels where the click is the event. This one does not have a reliable click.
“A referral channel you measure with a referrer. A demand channel you measure with a baseline. Getting AI wrong is mostly a matter of using the first tool on the second kind of problem.”
Which means the metric that actually tracks AI performance for most sites is branded search volume and direct traffic, measured against a baseline, correlated with citation share. Not clean. Not per-session. But it moves when the thing you care about moves, which is more than the referral line can say.
There is a second-order effect here that the AI Overviews half of the same research makes concrete. On news searches where an AI Overview appeared, publishers got a click about 20% of the time. Without one, about 30%. So the immediate click gets suppressed at the exact moment the delayed visit gets encouraged. Both things are happening to the same reader, in opposite directions, on different clocks.
If you only own the first clock, the story is a straight decline. If you own both, it is a channel shifting from instant response to delayed recall. Those call for completely different responses, and most teams are still running the first playbook because it is the only one their dashboard supports.
Who the follow-through goes to
One more number from the panel, and this is the one publishers should sit with. Major publishers accounted for 82% of publisher mentions in AI conversations, but captured 97% of the follow-through visits.
Being mentioned and being visited are not proportional. The big names convert a mention into a visit at a much higher rate, because the reader already knows the brand and will go find it. A smaller outlet gets named, the reader does not recognise it, and the mention dies in the chat window.
| PUBLISHER TIER | SHARE OF MENTIONS | SHARE OF FOLLOW-THROUGH VISITS |
|---|---|---|
| Major publishers | 82% | 97% |
| Everyone else | 18% | 3% |
That is brand equity doing the work, and it is a genuinely uncomfortable finding for anyone who has been sold AI visibility as the great leveller. Getting cited is not the win. Getting cited by a name the reader will go looking for is the win. If nobody recognises you, a citation is a compliment paid in private.
It also means AI visibility work and brand work are the same programme now, not two budgets. The same dynamic explains why publishers have been fighting over Google referral data instead of over citations: the citation was never the asset. The recognisable name was.
Fixing AI referral traffic attribution without new tools
You can close most of the gap this quarter with things you already own. None of this requires a vendor, a new tag, or a budget line. It does require agreeing, once and in writing, that you are moving from per-visit attribution to correlation, because half the arguments about AI measurement are really arguments about that unstated switch.
The other thing worth agreeing up front is what would falsify the story. If citation share rises for two quarters and direct traffic to those pages does not move at all, the channel is not working for you and the undercounting argument has run out of road. Write that test down before you start, so the correlation model cannot quietly become unfalsifiable.
The third one gets dismissed as unserious and it should not be. Self-reported attribution is noisy, biased, and the only instrument that survives contact with a user who retyped your name. Run it for a quarter alongside a proper reporting and analytics baseline and you will have a range you can defend, which is more than the referral line will ever give you.
Two caveats on the research itself, because I would rather you hear them from me than from a client. Scrunch sells AI visibility tooling, so the finding that AI visibility is undercounted is convenient for them. And the panel size is described as millions of search events without a stated number, which is not the same as a disclosed sample. Directionally strong, methodologically thin. Use it to reframe the question, not to size the answer.
What I would tell a publisher this week
Stop defending the AI referral number. It was never going to be the number. Go and look at what direct traffic did over the last six months on the pages that get cited most, and compare it against pages that get cited rarely. If the cited pages are pulling more direct visits than they used to and the uncited ones are flat, you have your answer, and you have it without a single new tool.
Then decide what you are optimizing for. If you are a name people recognise, citations convert and the work is to earn more of them. If you are not, the citation alone will not save you, and the honest priority is the unglamorous one: be worth typing into a browser bar. That is the same job it always was. The measurement got harder. The work did not change.
If you want the fuller version of tying this to revenue rather than sessions, we walked through the model in attributing pipeline to organic and AI-cited traffic. Source: Scrunch, Surviving the clickpocalypse, reported by Search Engine Land, August 13, 2026 (coverage).
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Josh leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.