A Few Dozen Visits A Month, And The AI Conversation Stops There
You’ve seen this bar. The AI referral line in the monthly report comes to a few dozen visits, and next to the others it sits flat on the floor.
Everyone at the review reads that bar the same way, and I read it that way too. I’ve stood there with the chart and nothing else behind it, watching the room settle on waiting to see how things developed. Then the item closed. I had no material to argue back with, which is a different problem from being wrong.
But the material exists.
The material is a count of more than two million AI conversations, matched against what those same people browsed afterwards, and it takes the reading apart. Eligible mentions were followed by site visits at a higher rate than the forecast baseline.
The referral report simply never found out about that week.
Those conversations are American, they run from January to June 2026, and each mention is followed for seven days — then set against that same person’s earlier weeks.
Do AI Mentions Actually Send People To Your Site?
They do. On Gemini, 5.42% of eligible mention exposures were followed within seven days by a visit to the same brand’s site. The forecast baseline, built from three earlier seven-day windows for the same users and brands, was 2.21%, so the post-mention rate was about two and a half times as high (Profound, “The AI Mention Effect”, 1 July 2026).
“Eligible” matters here. The brand appeared in the AI response but not in the user’s prompt; the main population also required observable browsing in the prior week and excluded exposures where that user had searched for or visited the brand during that week.
Three platforms are in that count: ChatGPT, Gemini and Google AI Overviews, over the first six months of 2026, in the United States. Browsing came from a privacy panel whose members opted in twice, explicitly, before anything about them was recorded.
And what I’d want to know before quoting 5.42% anywhere is what it gets compared against. Profound used a : the seven days after a mention are set against a forecast made from three earlier seven-day windows for the same person and brand.
Since the baseline is that person’s own past rather than some other group of people, the standing difference drops out. It’s the difference between asking whether you got heavier and asking whether you’re heavier than the person next to you. Somebody who visits a brand most weeks was always going to score high, and their earlier weeks score just as high, so the gap between those two closes on its own. What survives the subtraction is whatever moved when the mention landed, and that is the part I can take into a meeting.
The Gap Holds On All Three Assistants
The seven-day post-mention rate and forecast baseline came out at 7.79% against 4.83% for Google AI Overviews, and 6.39% against 4.33% for ChatGPT. Line the three up by the distance from each platform’s own baseline and Gemini leads:
- Gemini +3.21 points
- Google AI Overviews +2.96 points
- ChatGPT +2.07 points
The industry cuts are where I stopped scrolling. Narrow to one of them and the distances stretch: financial services on Gemini opened a gap of +5.7 points, a 132% lift on that category’s baseline, and retail opened +5.6 points, or 140%.
But points and percentages don’t share a ruler, and retail is where that shows. Smaller in points, larger in percent. Points say how many more people out of a hundred turned up; the percentage says how much bigger that is than what the category was already getting. If the baseline is low, a modest gap in points reads as an enormous multiple — which is why I’d read the points column first and the percentage column second.
Either way, the size of the prize turns on two things you already know about your own company: which assistant your buyers open, and which industry you sell into.
42% Of First Visits Arrive Within 24 Hours
Across all three platforms, 20.5% of the first brand-site visits after an exposure happened inside the first hour. Widen the window to 24 hours and it’s 42%.
A weekly line combines that timing distribution into one point. It cannot show when within those seven days the visits arrived.
My own dashboard still defaults to weekly. To inspect the arrival timing, I have to switch the display to daily while keeping the seven-day outcome window.
Where Do These Visits Show Up In Your Analytics?
Directly attributable visits are scarce. For ChatGPT, the share of first post-exposure visits carrying a product-matched was roughly 1.0% from January through April, 1.79% in May and 2.47% in June.
June is partial, so Profound says to read 2.47% as directional. The “about 2.5%, or one in forty” line therefore describes that partial June ChatGPT slice, not an all-platform rate for the full study period. The source separately reports that more than 97% of these brand-site visits did not include a UTM; it does not classify every untagged visit as either direct traffic or branded search.
That is still a case for counting somewhere else. Put the number of times the models say your name on its own line, outside the referral report, and the effect stops depending on an analytics tool happening to catch it.
Who Ran This Count, And What It Leaves Out
Profound sells AI visibility measurement, so a finding that mentions bring visits is good for the business that produced it. I’d want that on the table before anybody quotes the number (which doesn’t make the finding wrong, only worth reading with the seller in view). Two of the boundaries come from the publisher itself, and stating them counts in its favor:
- it measures site visits, not purchases or contracts
- it isn’t a randomized experiment, and the selection difference between people who met the name and people who didn’t isn’t fully removed
Three more sit around those. Results are published by category and by platform, so no individual brand’s figures are in there and nobody can check whether a company shaped like yours moves the same way. The window is January to June 2026 and the market is the United States, with nothing measured outside it. And the browsing came from a double opt-in panel (people who agreed twice, explicitly, to be measured), which is a self-selected group rather than AI users in general.
There’s also a reading the design can’t rule out. Maybe the mention and the visit both come out of something further upstream: the week somebody starts shopping a category. If that’s what happened, the answer pushed nobody anywhere and merely turned up in the same seven days as the shopping. A back-placebo design subtracts the standing difference between people; a one-off change in what one person is doing that week goes straight through it.
Maybe I’m leaning on this harder than it can hold. What I’d defend, though, is narrower than the 2.5x: the partial-June ChatGPT result shows how much a referral-only view can miss in this US panel, not a universal attribution rate. I’d treat both that share and the visit-rate multiplier as hypotheses to test at home. The measurement gap is the part to look for in your own reports.
Show Days, Keep A Seven-Day Outcome Window, Then Count The Mentions
The first-visit timing needs daily resolution, while the effect estimate needs its seven-day outcome window.
Fixing that is the small change: show daily values wherever AI visibility gets watched, but judge the post-mention visit rate over the same seven-day window Profound used.
The change people argue about is the second one. Take the referral report out of the headline slot and put the count of AI mentions there instead — how often the models said your brand’s name last month, against the month before. I’d keep the referral line on the page underneath it: the partial June ChatGPT result is still a warning that the directly attributable slice can be small.
And then the question that comes back across the table changes. “AI traffic is what, a few dozen a month?” has no move behind it. “Did we get named more often than last month?” points at your coverage, at third-party pages, at wherever other people write the name down.
What being named does before anybody clicks anything has its own numbers in the piece on the searches a model runs before it answers — that one covers the step upstream of this. Of the two changes here, I’d change the display first: a daily chart is what gives the second number something to land on, without shrinking the seven-day outcome window.
Sources
- Profound, “The AI Mention Effect”, 1 July 2026 (limitations: Profound sells AI visibility measurement and holds a commercial interest in the finding. The study measures site visits, not purchases or contracts. It is not a randomized experiment, and the selection difference between users who saw a mention and users who did not is not fully removed. Results are published at category and platform level only; no individual brand figures are released. The window is January to June 2026 in the United States, with no measurement outside that market. Browsing data comes from a double opt-in privacy panel, so the sample is not the general population of AI users.)
