Two very different things can happen to a brand inside an AI answer.
It can be named in passing, as one item in a list of options. Or it can be recommended, with the model attaching some degree of confidence to the choice.
Most AI visibility reporting treats those as the same event and adds them together. New research from June 2026 suggests that is a costly simplification.
What the research measured
A preprint posted to arXiv on June 9, 2026, titled From Prompt to Purchase, took an unusual approach. Rather than counting appearances inside AI answers, it joined opt-in clickstream data to users’ actual conversations with ChatGPT, Claude, and Gemini, then tracked what those users did next on the open web.
The design included several controls worth noting. Pre-trend event studies compared activity before and after a brand came up. The analysis conditioned on users with no recent observed engagement with the brand, which removes existing customers from the picture. A stance classifier separated genuine recommendations from incidental name-drops. Backward placebo windows at 14, 21, and 28 days prior provided a comparison baseline.
Among users with no recent observed engagement, the results split cleanly.
After a recommendation, same-name Google searches rose 4.3 percentage points. After a neutral mention, 1.8.
Brand site visits rose 2.4 percentage points after a recommendation, against 1.1 after a mention.
Retailer page visits rose 1.0 percentage points after a recommendation, against 0.3 after a mention.
Depending on the behavior, a recommendation moved roughly two to three times more than a mention did.
A second study pointing the same direction
Similarweb published parallel findings on June 23, 2026, drawn from finance, travel, and beauty.
Users who received an AI recommendation were 2.5 times more likely to visit that brand’s website within seven days. Those visitors viewed nearly twice as many pages and stayed roughly twice as long as standard visitors.
And one number in that study deserves more attention than it has received. More than 56 percent of AI-influenced traffic arrived through branded search, compared with about 40 percent for standard visits.
The influence happened inside the AI conversation. The visit was recorded as branded search. Two different systems, two different stories, one buyer.
Why averaging the two is expensive
If both studies are directionally right, then a dashboard reporting your AI mentions as one figure is averaging two events with very different commercial weight and labeling the result visibility.
Consider two brands with identical mention counts. Brand A is named twenty times, almost always in list form, rarely with any language of endorsement. Brand B is named twenty times, half of those as an explicit recommendation with confident framing.
Same number on the dashboard. Substantially different downstream behavior, if this research holds.
This is also why recommendation language is worth reading rather than counting. There is a real difference between “a strong option for teams that need X” and “you could also consider X.” Both are appearances. Only one carries confidence, and confidence appears to be what moves people.
What this means for reporting
Three practical adjustments follow.
Separate the two categories in your reporting. Any report that gives you one appearance number should be able to break it into mentions and recommendations. If it cannot, you are looking at a blended figure whose composition could shift entirely without the total changing.
Read the language, not just the presence. Confident framing and hedged framing are different diagnostic signals. Tracking which one a model uses about your brand over time tells you something a count never will.
Stop expecting AI referral traffic to tell the story. If the majority of AI-influenced visits arrive as branded search, then a report showing negligible AI referral traffic is not evidence that AI is not influencing your buyers. It is evidence that the influence is happening upstream of your attribution system, which is exactly where the research says it happens.
A caution about the evidence itself
There is a wrinkle worth stating plainly, because it matters for how much weight to put on these numbers.
The authors of the arXiv preprint decline to report panel size or per-cell sample counts, describing those figures as commercially sensitive. They are an AI visibility vendor. They state that all estimates clear an internal minimum-disclosure threshold by a wide margin.
The methodology looks more careful than most vendor research. The controls are real. But six weeks after that paper posted, the IAB defined decision-grade measurement as requiring reproducibility, disclosed variation ranges, and methodology transparency.
Which means the strongest available evidence that AI recommendation moves buyers does not currently meet the standard the industry just set for itself.
That is worth holding in mind. Treat these findings as strong directional evidence rather than settled fact, and watch for replication. The direction is corroborated by an independent study from a different company using a different method, which is the kind of convergence that should raise confidence. The precise magnitudes are the part to hold loosely.
The takeaway
Mentioned and recommended are not interchangeable, and the gap between them appears to be where the commercial value sits.
If your reporting collapses them into a single number, the most important thing your AI visibility data could tell you is the thing it is currently hiding.
FAQ
What is the difference between an AI mention and an AI recommendation?
A mention is any appearance of your brand name in an AI answer, including a passing list entry. A recommendation is when the model attaches endorsement or confidence to your brand as a suggested choice. Research from June 2026 found recommendations move downstream buyer behavior roughly two to three times more than mentions.
How much does an AI recommendation increase branded search?
The From Prompt to Purchase preprint found branded search rose 4.3 percentage points after a recommendation, compared with 1.8 percentage points after a neutral mention, among users with no recent observed engagement with the brand. The study joined opt-in clickstream data to real ChatGPT, Claude, and Gemini conversations.
Why does AI influence show up as branded search instead of referral traffic?
A buyer often reads an AI recommendation and then searches the brand name separately rather than clicking a link. Similarweb found more than 56 percent of AI-influenced traffic arrived through branded search, against about 40 percent for standard visits. The recommendation drove the visit, but the search engine received the attribution.
Should I track AI referral traffic as my main AI visibility metric?
No. Referral traffic captures only the portion of AI influence that produces a direct click, which the June 2026 research suggests is a minority of it. Branded search growth, direct traffic, and high-intent page visits are better indicators of AI influence.
Is this research reliable enough to act on?
Treat it as strong directional evidence rather than settled fact. Two independent studies using different methods pointed the same direction, which raises confidence in the direction. The precise magnitudes are less certain because the arXiv authors withhold sample size figures as commercially sensitive.
How do I tell whether AI is recommending my brand or just naming it?
Run your key buying questions across ChatGPT, Claude, Gemini, and Perplexity and read the language rather than counting appearances. Confident phrasing such as “a strong option for” signals recommendation. Hedged phrasing such as “you could also consider” signals a mention with low conviction.
Start here: axissuite.ai
Axis Suite by TrendAxis