TL;DR
We asked ChatGPT, Claude, and Gemini the same 15 buying questions, sometimes with a one-sentence description of the person asking and sometimes without, and collected 675 answers.
When the description said something about the buyer's situation, such as their budget, job, company size, or home, the brands in the answer changed a lot. For B2B accounting software, the top three brands named overlapped with the anonymous answer's top three by only about a third, compared with about two thirds between repeat runs of the same question.
A description that gave only age and gender changed nothing beyond normal run-to-run variation. In our two categories, energy and accounting software, the product doesn't depend on who you are.
That fits with Profound's recent persona study of clothing, furniture, and credit cards, where gender, age, and income do change what people need. Taken together, the two studies suggest demographics change answers when the product itself depends on them, while the buyer's situation changed them in every category we tested.
Asking the identical question twice already returns a noticeably different list of brands, so a single AI answer is a sample rather than a verdict.
For anyone working on answer-engine optimisation (AEO), the lesson is to measure visibility for the buyers you actually sell to, and to make it obvious in your content which situations you're the right choice for.
What we learned
There isn't one AI answer for your category
Most AI visibility tracking asks questions as an anonymous person in a given country. It's a sensible default, but it isn't how people use these tools. Someone asking ChatGPT for accounting software has usually told it something about themselves by then, in the question itself, earlier in the chat, or in the memory and custom instructions the app keeps.
That context turned out to matter a great deal. We added one sentence about the person asking, the sort of thing people write in ChatGPT's custom instructions, and watched which brands came back.
For the finance director of a 200-person manufacturer with several legal entities. Microsoft Dynamics 365 Business Central went from appearing in 3% of answers to 71%, and Oracle NetSuite from 7% to 61%. FreeAgent, named in 90% of the anonymous answers, dropped to 44%. Zoho fell from 58% to 11%.
The founder of a two-person startup got a different list again. Sage dropped from 78% of answers to 38%, while NatWest and Mettle business accounts, which come with free bookkeeping software, turned up more often. For a self-employed plumber who wanted something simple on their phone, Xero fell from 88% to 56%. The answers also got shorter, with about seven brands each instead of 11, so every place on the list was worth more.
A brand can look strong in AI search overall and still be missing for the buyer it cares about most. FreeAgent's anonymous numbers are excellent, yet it vanishes from more than half the answers a finance director gets. That might be perfectly fine, since FreeAgent isn't built for multi-entity manufacturers. You only find out by checking.
Engines respond to needs, and to demographics only when the product depends on them
One of our test descriptions said nothing except "I'm a 45-year-old woman". It made no measurable difference in either category, on any of the three engines. Those answers were as similar to the anonymous ones as two anonymous answers are to each other.
A busy parent choosing an energy supplier didn't move the answers either. Being a parent is a big part of someone's life, but this description gave the engine nothing that bears on which tariff to pick. The descriptions that worked all carried a need the decision turns on: a tight budget, solar panels on the roof, audit controls across several companies, or very little confidence with technology.
Clothing differs by gender and age, and credit cards and furniture differ by budget, so in those categories a demographic detail is also a need. An energy tariff or a bookkeeping app doesn't depend on whether the buyer is a man or a woman, and there age and gender did nothing. Income, meanwhile, is a budget by another name, and budget moved the answers in our test too.
The rule we'd follow is to include a demographic detail only when your product genuinely varies by it, and otherwise to describe the buyer's situation. If you sell clothing, gender matters. If you sell payroll software, company size matters and gender almost certainly doesn't. Building personas out of demographic boxes alone, in a category where they don't apply, will mostly show you noise.
In consumer categories, the sources move more than the brands
UK energy behaved differently from B2B software. Octopus Energy appeared in 83–89% of answers whoever was asking, and the other suppliers named barely shifted.
What changed was the advice around them. The student on a tight budget was sent to comparison sites far more often: MoneySavingExpert went from 21% of answers to 38%, and Uswitch from 22% to 33%. The retiree who cared about good service heard about Which? in 71% of answers, up from 48%, and saw Trustpilot and Citizens Advice more often too.
For a consumer brand, that changes where the work goes. Price-sensitive buyers are being pointed at comparison sites, while service-minded buyers are pointed at reviews and consumer advice. Being well represented in those places is part of being recommended to those people.
Each engine listens to a different degree
Claude adapted its answers the most, Gemini a little less, and ChatGPT the least. ChatGPT ignored the energy student altogether, yet it still changed course clearly for every one of the B2B buyers. The same buyer can look quite different from one engine to the next, which is easy to miss in a single blended score.
A single answer proves very little
Before comparing buyers, we measured how much the answers vary when nothing changes at all. Ask the same engine the same question three times and only 58–69% of the brands named overlap from one run to the next. For the top three brands, the overlap ranged from 46% to 71%.
That's the bar every claim about AI visibility has to clear. A screenshot of one ChatGPT answer, flattering or not, is a single throw of the dice. Trends built from repeated runs are the only dependable read, and that applies just as much to comparing buyers.
What this means for your AEO work
Track the buyers you sell to. If your revenue depends on mid-market finance teams, find out what AI recommends to a finance director, as well as what it tells a stranger.
Say plainly who you're for. The engines match on situation, so pages that spell out the situations you're the right choice for give them something to match against. "Built for groups with several legal entities" tells an engine far more than "powerful and flexible".
Look at where your buyers are being sent. In consumer categories especially, the comparison, review, and advice sites vary by buyer. Check which ones appear for your audience, and make sure you're well represented on them.
Use demographics only where your product depends on them. Age and gender added nothing for energy tariffs or accounting software. If your products genuinely differ by who's buying, as clothing does, include them; otherwise put the effort into describing situations.
Judge trends rather than single answers. Run the same questions repeatedly and watch the share of answers that name you over several weeks.
The research we did
We set out to answer one question: does describing the person asking change which brands AI engines name, and is that change bigger than the ordinary variation between runs?
The questions. We used 15 real buying questions from prompt sets we already track: seven about UK energy suppliers and eight about accounting software for UK small businesses. We left out any question that already named an audience, such as "for sole traders", so that the description was the only thing telling the engine who was asking.
The buyers. Each question ran five ways. One version had no description, which is how AI visibility tracking normally works. The other four added a short first-person description. For energy, those were:
a 21-year-old student renting a room in a shared flat, with money tight
a 68-year-old retiree with a detached house and solar panels, who values good service over the lowest price
a 38-year-old nurse with two young children, who wants predictable bills
"I'm a 45-year-old woman", with nothing else
For accounting software, they were:
the 29-year-old founder of a two-person startup who wants something cheap and quick to set up
the finance director of a 200-person manufacturer with several legal entities
a 52-year-old self-employed plumber who isn't confident with technology
the same demographic-only line as for energy
The engines. We used ChatGPT, Claude, and Gemini through their APIs, with web search on and the location set to the UK. The description went in as an "About me" line in the instructions sent alongside each question. Google's AI Overviews and AI Mode weren't part of the test, because there's no way to tell them who is searching beyond location and language, unless you are Google themselves.
Two choices set this test apart from other persona studies we've read. Every question also ran with no description at all, so we could compare each buyer with the anonymous answer most tracking tools report, as well as with the ordinary variation between runs. And the description went in as background, the way saved custom instructions reach the model, rather than inside the question, so the question itself stayed exactly as a buyer would type it.
The repeats. Every combination ran three times, which came to 675 answers (15 questions, three engines, five versions, three runs). The repeats are what let us measure the noise.
The measurement. A language model read each answer and listed the brands it named, in order. We then checked that every brand it listed really does appear in the answer text. For each question and engine, we compared two things:
how much the brand lists overlapped between repeat runs of the same version
how much they overlapped between the anonymous version and each described buyer
If a description does nothing, those two numbers should match. If it changes the answer, the second should be lower. To check the gaps weren't luck, we shuffled the labels thousands of times. All five situational descriptions passed comfortably. The busy parent and both demographic-only lines did not.
The limits. This was a focused test. It covered two UK categories in English, collected on a single day, with three runs per combination and one wording for each buyer. We tested the engines through their APIs rather than the consumer apps, which keep their own memory of each user and may carry far more context than one sentence does. The pattern was clear in both categories, but expect the size of the effect to vary by category and market.