Does ChatGPT recommend different vendors to different people?

ChatGPT can recommend different vendors to different people, and two separate mechanisms produce that result. One is personalisation: saved memories, custom instructions and location give the model context your colleague does not have. The other is sampling, which makes the same account get different answers to the same question minutes apart.

Sorting those two apart matters commercially, because they call for different responses. If your absence from an answer is personalisation, it tells you something about that one buyer's history. If it is sampling, it tells you nothing at all, and acting on it means rewriting pages on the strength of a coin toss.

The measured size of the second effect is larger than most teams assume. Schulte, Bleeker and Kaufmann, April 2026, queried eight prompts per campaign up to ten times in succession across ChatGPT, Gemini, Google AI Mode and Perplexity between 21 and 25 March 2026, comparing 3,409 pairs of runs inside a single 24 hour window, and found the cited sources overlapping at a mean Jaccard similarity of 0.32 to 0.43 depending on the campaign. By engine, the mean source overlap was about 0.23 for ChatGPT, 0.28 for Perplexity, 0.32 for Google AI Mode and 0.51 for Gemini.

The useful half of that study is what held steadier. Brand overlap ran higher than source overlap, at 0.33 to 0.48 within a day and 0.45 to 0.59 between consecutive days, which means the list of companies an assistant names is more stable than the list of pages it cites. For a vendor watching whether it gets recommended, the brand list is the signal and the source list is the noisier layer underneath it.

Does ChatGPT give the same answers to everyone?

ChatGPT does not give the same answer to everyone, and the differences come from settings a user controls rather than from a profile the model builds about your company. Saved memories, custom instructions, the chosen model and the user's location all feed the context the answer is generated from, and all of them can be reviewed, edited or switched off in the product's settings.

What those controls mean in practice is that an assistant answering a buying question is answering it for one person with one history. A procurement lead who has spent three months asking about compliance requirements is a different reader from a founder opening a fresh window, and the assistant has been told as much. Neither answer is the canonical one.

The part worth separating out is that none of this is targeting in the advertising sense. The model is not selecting vendors to show a particular buyer on the basis of a commercial relationship. It is carrying forward the stated preferences and constraints of the person in front of it, which is closer to a long conversation than to an ad auction.

For anyone measuring their own visibility, the operating rule that follows is simple and widely ignored: check in a clean session, signed out, with memory and custom instructions off, and keep the conditions identical from month to month. Your own logged-in account is the single worst place to judge whether buyers see you, because it has been trained by months of your own research into your own category. The mechanism behind the rest of the variation is covered in why AI engines give a different answer every time you ask.

What do the engines document about personalising answers?

Google documents its personalisation in the most detail of the four engines. Its help page for Personal Intelligence in AI Mode states that the feature "references previous searches and activity saved in your Search Services History", that personalisation is available to users who are 18 or older with history and personalised recommendations enabled, and that memory in AI Mode is "available in the US, in English", per Google's own help page for AI Mode personalisation, read 22 October 2026.

Two further details on that page matter for anyone doing measurement. Personalisation can be turned off through the personalised recommendations setting, which is what makes a controlled check possible at all. And Google states plainly that "AI Mode might not always get personalization right", with the user expected to correct it in a follow-up, so the personalised layer is a guess about context and not a reliable filter.

OpenAI documents a comparable set of controls for ChatGPT, built from saved memories and custom instructions that the user can inspect, delete or disable, with temporary chats available for sessions that should not be remembered. The company publishes those controls in its help centre, and the practical upshot for a vendor is the same as Google's: the person asking can change what the assistant knows about them, and you cannot see what any individual buyer has set.

What none of the engines publishes is a measurement of how much personalisation changes a vendor recommendation. That gap is the reason the next two sections exist, and it is also why claims that AI answers are now individually targeted should be treated as assertions, not findings. The honest statement is that personalisation exists, is documented, and has not been sized in public for commercial queries.

How much of the difference between two answers is personalisation?

Most of the difference between two answers is not personalisation. The clearest way to see that is to remove personalisation from the experiment entirely and watch how much variation is left, which is what the repeated-run studies do: identical prompts, no account history, answers that still disagree with each other.

Schulte, Bleeker and Kaufmann, April 2026, ran exactly that test. Across 45 to 46 days between 24 January and 20 March 2026, comparing 4,044 consecutive-day pairs across four engines, day-to-day overlap of cited sources averaged a Jaccard similarity of 0.34 to 0.42, which the authors read as roughly 35% of sources surviving from one day to the next. Rank order moved further, with a mean rank-biased overlap of 0.21 to 0.26.

So when a colleague reports a different vendor list from yours, the first explanation to check is the engine, not the colleague. A controlled comparison needs the same prompt, the same engine, the same day, several runs each, and both sessions clean. Anything less cannot separate the two causes, and most of the anecdotes that circulate in marketing teams are undiagnosed single runs.

This is also the reason a single screenshot is worthless as evidence in either direction, including when it flatters you. Our own category studies run five executions of every query for that reason, and in the process skid study no fabricator was named by all four engines in any contestable category. If you want to know how many prompts and runs it takes before a reading means something, how many prompts you need works through the arithmetic.

Does adding personal context change which vendors get named?

Adding context to a prompt does change which vendors get named, and the measured effect is to surface brands that never appear otherwise. Malthouse, Lee, Yang, Pal and Feng, September 2026, tested six commercial models with web search disabled across five consumer categories, taking 40 runs per model per category for plain category prompts and then 100 needs-based prompts, each sent to all six models twice.

The study's illustration of the difference is the pair of numbers to hold onto. Under category-only prompts, the study found that Craftsman and L.L.Bean never appeared at all, with a recommendation prevalence in the top five of 0%. Under needs-based prompts that stated a buyer's goals and constraints, Craftsman reached a prevalence of 35.4% and a mean reciprocal rank of .108, and L.L.Bean reached 5.4% and .025, each estimated over 240 lists. Diagnostic probes written to describe the brand's positioning directly pushed prevalence to 81.3% and 88.5%, and the authors call those probes intentionally unrealistic as consumer requests, so read them as a ceiling rather than as a tactic.

Three limits belong with those figures. The categories were consumer goods rather than B2B, web search was switched off so the models answered from their own knowledge, and the authors present the two-brand comparison as an illustration and not as the systematic result of the 100-prompt study. They describe their marketplace findings as exploratory associations only.

What it supports is narrower than "personalisation decides who wins" and more useful. The specificity of the question moves the answer: a buyer who states a constraint gets a different roster from one who asks a generic category question, because the constrained question matches different language. That is the same lever page work pulls, and it is why the pages that win are the ones written in the constrained wording a real buyer uses instead of in category terms.

How should you measure your AI visibility if answers differ by person?

Measure your AI visibility under personalisation by holding every condition you can hold still and then repeating the run. The approach is not impossible, it just rules out the casual check, and the four controls below are what separate a comparable series from a collection of screenshots.

What to control	How	Why it matters	What happens if you skip it
Account history	Signed out, memory and custom instructions off, temporary chat	Your own account has months of your category research in it	You measure your own reading habits and call it the market
Run count	Several runs of each prompt, on more than one day	Intra-day source overlap measured at 0.32 to 0.43 means one run is a sample	Noise gets reported as a win or a loss
Prompt wording	A fixed prompt set, written in buyers' words, changed only deliberately	Stated constraints change which brands surface at all	This month's number is not comparable with last month's
Engine and locale	Same engines, same country and language, recorded with each run	Google documents AI Mode memory as available in the US, in English	A locale difference reads as a visibility change

What you record matters as much as the controls. Log the brands named and the URLs cited as two separate series, because the brand list is the steadier of the two, and a page dropping out of the citations while your brand is still named is a different problem from your brand disappearing. If you would rather see that series built for your own category before deciding whether to run it in-house, our free visibility assessment produces the first two runs, and why two tools disagree about the same brand explains why vendors' dashboards rarely match.

What else do people ask about personalised AI answers?

The five questions below come up about personalised answers when a marketing team compares what they see in an assistant with what a colleague or a customer reports seeing.