How to Read Profound's 'AI Mention Effect' Study

本文目前只有英文版。站点其余部分已汉化，这篇内容尚未翻译。

Short answer: the study finds that when an AI assistant brings up a brand the user didn't mention, that user visits the brand's site more often over the next seven days than their own earlier behavior predicts. It reports a lift of roughly 2 to 3 percentage points depending on the platform. The design is careful, but it's observational, US-only, and published by a company that sells AI visibility tools. Read it as strong evidence that AI mentions are followed by visits your analytics mostly won't attribute, not as a forecast of what a mention is worth to your site.

The study is "The AI mention effect" by Profound, dated 2026-07-01 (read 2026-09-24). Everything below about its method and results comes from that page. We summarize it rather than reproduce its tables, so open the original for the full per-industry breakdown.

One disclosure: TriRank sells AI visibility measurement too, as does Profound, the study's publisher. Profound says as much in the introduction: the question the study tests is the bet the company is built on. That doesn't make the result wrong. It's a reason to read the method carefully, which is what this post does.

What was measured

The data. A double-opt-in panel of consenting US users that records both their AI conversations (prompts and responses) and their subsequent web browsing. The study covers more than 2 million AI conversations from January through June 2026, across ChatGPT, Gemini and Google AI Overviews.

The treatment. An AI-exposure is a brand that appears in the AI's response but not in the user's prompt. The study leaves out cases where the user named the brand first, because those users were already interested. This is the most important design choice in the study: it isolates brands the assistant introduced.

The outcome. Whether the same user visited that brand's own website in the seven days after the response.

Filters. The main results require the user to have observable browsing in the prior week. They drop any exposure where the user had searched for or visited the brand in the prior week, and they exclude common-platform and generic high-traffic domains.

How the baseline works

Some of those visits would have happened anyway. The study handles that with what it calls a forecasted backward-placebo design:

1. For each exposure, take the seven days after the response as the treated window.
2. Take three seven-day windows before the exposure for the same user and the same brand. Those are the placebo windows: nothing was mentioned, so any visits there reflect the user's existing behavior.
3. Fit a linear forecast through those three pre-period rates to predict what the post-exposure rate would have been without the mention.
4. The lift is the treated rate minus that forecast.

Using the same user and the same brand as their own control is a strong choice. It removes a lot of what makes observational comparisons unreliable, such as some users simply browsing more or some brands simply getting more traffic. The linear forecast also allows for a trend, so a user whose interest was already rising doesn't inflate the result as much.

How to read the results

The headline table reports, per platform, the treated rate, the forecast baseline and the lift with a 95% confidence interval. Three things to keep in mind when reading it:

Percentage points vs. multiples. The study's takeaways describe the effect as 1.5 to 2.5 times the forecast baseline. That's the relative figure. The point lifts are closer together: about +3.2 points for Gemini, +3.0 for AI Overviews and +2.1 for ChatGPT. Gemini's multiple looks largest partly because its baseline was lowest, at 2.21%. When comparing platforms, compare the point lifts.

One wrinkle: the takeaways say AI Overviews has the largest absolute lift, but in the table Gemini's point lift is slightly higher, and the two intervals overlap. AI Overviews does have the highest treated visit rate (7.79%) and, per the study, the most exposures. That may be what "absolute" refers to. The page doesn't say, so we go by the table.

Confidence intervals. The intervals come from a user-clustered bootstrap: the analysis was rerun 2,000 times, resampling users rather than individual exposures, so one heavy user can't dominate. None of the three platform intervals crosses zero. ChatGPT's, for example, runs from +1.67 to +2.48 points. So under this design, each lift is unlikely to be pure noise. Gemini's interval is noticeably wider than the other two, so its point estimate is the least precise of the three.

Timing. Across platforms, 20.5% of first visits came within an hour and 42% within a day. Most came later, which is the study's argument for a seven-day window over same-session clicks.

What it can't tell you

The authors list the key limitation themselves. The study is "not a purchase or conversion study, and not a randomized experiment." Here's what follows from that:

• Association, not proven cause. The backward-placebo design controls for each user's own history, but the authors note it doesn't fully eliminate selection bias. One likely source, in our reading: a user who asks an assistant about a category is already in the market. The assistant's answer and the later visit could both follow from that intent. The design reduces this problem but can't remove it.
• Visits, not value. A visit isn't a sale. The study doesn't claim to measure revenue.
• One country, one panel. US only, and drawn from people who opted into having their AI use and browsing recorded. They may not behave like everyone else.
• Aggregates only. Results are published at platform and industry level, with individual brands excluded. There's no breakdown by brand size, and very large generic domains are excluded from the headline results.
• Three surfaces. ChatGPT, Gemini and AI Overviews. Other assistants aren't covered, so the study can't tell you about them either way.

What a small site can take from it

1. Referral data undercounts AI's effect. The study found that only about 2.5% of downstream visits carried a trackable AI-referral parameter, even after ChatGPT made links easier to click in May. If you judge AI visibility by AI referral sessions in your analytics, you're probably seeing a small part of it. It fits a gap we see from the other direction: a click from Google's AI Overviews arrives as ordinary Google search traffic, not as a separate AI referral. For what Google's own reporting does and doesn't separate, see our comparison of Google's generative AI report with our own sweeps.

2. The mention is the unit to measure. If visits follow mentions even without a click, then whether an assistant names you is worth knowing directly. Our free AI visibility checker gives a dated reading per engine of whether your domain appears among the cited sources, with no account needed. A cited source isn't exactly the same as the study's brand mention, but it's the closest thing you can check yourself.

3. Don't borrow the numbers. A 2 to 3 point lift averaged across a US panel tells you nothing specific about a small or new brand. The direction is the transferable part. The size isn't.

4. Measure over a week, not a session. If you compare traffic before and after a change in your AI visibility, use a window of several days. The study's own timing data shows most visits don't happen in the first hour. For how we record readings so that a later comparison means something, see how TriRank measures AI citations.

FAQ

Does the study prove that AI mentions cause website visits?

No, and it says so. The authors describe it as a site-visit study, not a purchase or conversion study, and not a randomized experiment, and they note that their design doesn't fully eliminate selection bias between exposed and non-exposed users. It shows that visits rose after AI-introduced mentions compared with a forecast built from the same users' earlier behavior. That's good evidence of association, not proof of cause.

Which platform showed the biggest lift?

It depends which figure you read. Gemini has the largest multiple, about 2.5 times its baseline, partly because its baseline was the lowest of the three. In the study's own table Gemini also has the largest percentage-point lift, just ahead of AI Overviews, with overlapping confidence intervals, even though the takeaways call the AI Overviews lift the largest in absolute terms. Point lifts are the more comparable figure, and there the three are closer together than the multiples suggest.

Does this apply to a small or new brand?

Not directly. Results are reported only at platform and industry level, US only, and the study excludes common-platform and generic high-traffic domains. It doesn't report results by brand size. The direction of the finding, that mentions are followed by visits your referral data doesn't show, is the part worth taking. The size of the lift for your brand is unknown until you measure it.