Only 1.1% of Post-Chat Publisher Visits Carry an AI Referrer. Your Analytics Are Not Broken, They Are Measuring the Last Click.
Scrunch tracked millions of searches, AI conversations and website visits from a privacy-safe opt-in panel between February and June 2026, then looked at what people did after a conversation with an AI assistant. Of the publisher visits that followed, about three quarters arrived by direct navigation, about 9% came through traditional search, and 1.1% carried an AI referrer.
Every solo operator running a content site has stared at that 1.1% in their own analytics and drawn the obvious conclusion: AI sends nothing, so stop thinking about it. The study is a decent argument that the conclusion is wrong and the number is right, which is an uncomfortable combination to sit with.
The measurement cannot see what you want it to see
Referral data captures the last click before a visit. That is all it can capture. Scrunch says this explicitly, and it is the whole ballgame.
Picture the actual path. Someone asks an assistant about a topic. The assistant mentions three sources. The person reads the answer, closes the tab, and forty minutes later types your domain into the address bar, or searches your site name, or clicks a Google result for a query the conversation gave them the vocabulary for. Every one of those arrives as direct or organic. The conversation that caused it is invisible to your analytics by construction, not by accident.
So when your dashboard says AI referrals are a rounding error, it is telling you the truth about referrers and nothing at all about influence.
The study's attempt to get at influence is the interesting part. Readers were 20.5 percentage points more likely to visit a news publisher during the week after a news-related AI conversation than after a non-news conversation, comparing each reader's own browsing history. Scrunch is careful about what that does and does not mean: it does not show AI created new demand, only that news-related AI conversations often coincided with continued news reading. Somebody who asks an assistant about an election was probably going to read about the election.
That caution is why I trust the rest of the numbers more than I otherwise would.
The 97% is the number that should actually bother you
Here is the finding that got almost no coverage, and it is the one with teeth.
When an assistant volunteered a publisher the reader had not named in the prompt, visits to that publisher rose 10.6 percentage points the next day and nearly 20 points over the following week, compared with similar publishers not mentioned in the same response. So being named does move traffic. Real effect, not a rounding error.
Then: major publishers accounted for 82% of the publisher names assistants volunteered, and 97% of the resulting follow-through visits. Mid-sized publishers made up about 17% of mentions and received a much smaller share of the visits.
Read those two numbers next to each other. Being mentioned is concentrated. Converting a mention into a visit is more concentrated still. A mid-sized publisher getting 17% of mentions and a sliver of the visits means the mention alone does not do the work. The brand recognition that makes someone act on a mention does, and that is the thing a small site does not have.
Every "get cited by AI" guide sells you the first half of that mechanism. The 97% is the second half, and it says the mechanism pays off in proportion to the audience you already had.
What this does to the AI Overviews panic, specifically
The study also found that news searches showing an AI Overview produced publisher clicks about 20% of the time, against roughly 30% for searches without one. That looks like a ten-point catastrophe.
Then Scrunch compared the same query with and without an Overview, and the gap collapsed to about two percentage points.
The difference between those two comparisons is selection. AI Overviews appeared on about one in four news searches during the study period, and they showed up disproportionately on queries that already converted badly: weather, market prices, explainers. Sports searches, which had the highest publisher click-through rates, rarely triggered one. Google is putting Overviews where people were not clicking through anyway.
That is a meaningfully smaller effect than the raw comparison implies, and it is a good reminder that the aggregate before-and-after charts everyone quotes are comparing different query mixes.
Where I could be wrong, and where the study could be
Three honest caveats, and the first two are about the study rather than my reading of it.
It is a panel. Opt-in panels skew, and a privacy-safe panel skews further, because the people who install tracking software knowingly are not a random sample of readers. The direction of that skew is not obvious to me, which is worse than knowing it.
It is news publishers. Everything above is about news reading behaviour, and I am generalising it to content sites in general, including a personal tech blog, which is a different animal with different intent and different repeat-visit dynamics. Treat the shape as transferable and the exact figures as not.
And the counter-argument to my "measurement, not absence" framing deserves stating properly: it is entirely possible that AI genuinely sends very little traffic, that the 1.1% is close to the truth, and that the direct-navigation share is people who were coming anyway. Scrunch's own 20.5-point finding is deliberately framed as correlation. If you want to believe the pessimistic reading, this study does not stop you, and I cannot prove it wrong from a single panel.
What I would actually do
Stop trying to measure AI referrals. The number is real and it is not the number you want, and building a dashboard around it will make you optimise for something that does not exist.
Measure two things instead.
Direct and branded search, tracked over time. These are the channels an AI conversation actually converts into. You will not get attribution, but you will get a trend line, and a trend line that moves when nothing else changed is a signal. On this blog, direct is the only line I look at week to week.
Whether you get mentioned at all. Not with a paid AI-visibility tool, at least not yet. Take the ten questions a reader would plausibly ask that your best posts answer, run them through two or three assistants once a month, and write down whether your domain appears unprompted. That is fifteen minutes and it tells you something the referrer column structurally cannot.
Then adjust your expectations honestly. If you are a small site, the 82/97 split says a mention is worth less to you than to a large one, so build for the case where somebody arrives having already decided to trust you: a name they recognise, an archive that rewards a second visit, and something to subscribe to. That has been good advice for fifteen years. It is more true now, not less, because the discovery layer got worse at sending anonymous traffic and better at reinforcing names people already know.
Author
Lukas
@lukcombinator