AI Search Gives Different Buyers Different Brand Lists.

A brand can be visible in AI search and still miss the buyer it wants.

Profound ran 71,147 responses across ChatGPT, Claude and Gemini, then changed the age, income, gender and occupation attached to the same shopping prompts. Brand mentions shared only about 1 in 4 names across different personas. AEO Wire independently logged the study’s figures. Citations overlapped even less.

The finding matters because AI-search reporting usually rolls every prompt into one score. If that score rises, the team assumes it has gained visibility. It may only have gained visibility for a buyer it doesn’t serve well.

Profound’s own summary puts it plainly: “Answer Engines give different recommendations to different people, even when the underlying question stays the same.”

One prompt is not one market

The study covered clothing, furniture and credit cards. Across the three systems, Profound found that same-persona answers repeated more of the same brands than answers for different personas. It also found a larger gap in cited sources: the same personas shared 34% of cited domains, while different personas shared 20%.

That’s not a small measurement wrinkle. It changes what an AI-search dashboard is saying.

A SaaS company that asks, “What’s the best CRM for a startup?” is testing a bundle of assumptions about company size, budget, job title, technical maturity and urgency. The AI can make those assumptions itself. Or the buyer can give it enough context to make a different recommendation.

The winner in the generic test may not be the winner for the actual account you’re trying to close.

The practical implication is narrower than “personalize everything.” Start splitting the prompts that decide your category into the audiences your marketing already uses. Segment by company size, role, budget band, use case and switching trigger. Then compare who shows up, which claims recur and what source material the answer cites.

That work is closer to message testing than SEO. It asks whether the product appears for the buyer you built it for, not whether it appears in an average answer that nobody will ever see. Profound’s own documentation shows why the aggregation matters: its Visibility Score is built from answer-engine responses across a selected prompt set.

The caveat belongs in the result

Profound sells the monitoring product this research makes more necessary. The work is useful, but it isn’t an independent field study of real consumer behavior. It covers 3 categories and personas supplied as prompt attributes, not observed people making purchases.

It also measures model outputs from a defined collection period, August 11 to 24. Answer engines change. A different model version, prompt wording or source mix could move the result.

Those limits don’t erase the finding. They set the right claim: this is evidence that persona changes can affect AI recommendations, not proof that every buyer sees a wholly different market.

Huang’s recommendation follows from the same restraint: “Marketers should track AI visibility by audience, not just by platform.”

Your score needs a denominator

The first generation of AI-search tools made the work legible by counting mentions and citations. That was useful. The citation set can also shift quickly, so each persona needs a dated record rather than a one-time audit. It also created a false comfort, because the score aggregates the people whose answers matter with the people whose answers don’t.

A visibility report without an audience definition is like a paid-search report that combines enterprise buyers, students and job seekers into one conversion rate. The number still moves. It just can’t tell you whether the budget worked.

Keep the aggregate score if it helps show trend. Add a second view for the 3 buyer situations where a recommendation would change pipeline. Name the persona, use the exact question they would ask and log the sources the model uses.

The brand list is becoming audience-specific. The measurement should catch up. That makes AI-search scorecards more useful when they reveal the prompt set, and less useful when they pretend one aggregate score settles the question.