Does everyone get the same answer when they ask AI about a business?

Two people can ask for the same kind of business on the same day and be given different names. That is not a fault, and neither of them did it wrong. Three separate things move the answer, and only one of them has anything to do with the business being named.

Why does the answer move when nothing about the business has changed?

Because the answer is written from scratch every time, and chance goes into the writing. A lab test shows how little it takes. Researchers put one question to one AI 1,000 times, with the setting that adds variety turned right down. They still got 80 different answers back (Source: Thinking Machines Lab, 2025).

That was a controlled test, with everything else held still on purpose. Out in the open, with real people on their own accounts, it is looser still. In the study behind the answer above, 600 volunteers put 12 questions through three AI tools 2,961 times between them. Two answers to the same question matched on the same set of brands less than 1 time in 100, and matched on the same set in the same order less than 1 time in 1,000 (Source: SparkToro, 2026).

Does who is asking change which businesses come up?

Yes, and by a lot. One study made up 320 customers on paper, differing in income, age, sex, and whether they lived nearby. It asked three AI tools where to eat in 304 neighborhoods across five cities. Every made up customer was offered the same real places. The picks were not the same (Source: Chen, Weng, and Moro, 2026).

The customers on the highest incomes were sent to places with about 38.3% fewer real visits, and about 106.0% more spent per purchase, than the places offered to the lowest incomes. Customers written as living nearby got cheaper, busier places. Customers written as visiting were pushed towards pricier ones (Source: Chen, Weng, and Moro, 2026).

Read that for what it is. It was a test about places to eat, run with made up customers on three AI tools, and not a reading of any real business. The part that carries over has nothing to do with food. The person doing the asking shaped the list before any business in it got a look in, and a plumber, a clinic, or a shop gets picked out of a list the same way.

Does Claude or Gemini remember the person asking?

Yes, when that person is logged in. Anthropic, the company that makes Claude, says in its help pages that Claude "can also remember context from your chats and carry it into new conversations". It says memory "is on by default for Free, Pro, and Max plans", which are its free plan and its two paid personal plans (Source: Anthropic documentation, 2026).

Google describes the same kind of thing for Gemini. Its help page says an answer can be shaped by "the memory of your past Gemini chats", by what a person has done in the Google apps they choose to connect, and by saved notes on how they want answers written. Those settings need a personal Google account, and they are switched off while somebody is logged in to "a work, school, or supervised account" (Source: Google documentation, 2026).

What each company keeps, and for how long, is a setting that moves. The detail is not really the point. The point is that a logged in regular and a stranger in a private window, the kind that forgets you the moment you close it, are not asking the same question, even when they type the same words.

What does one screenshot actually prove?

Very little by itself. A customer sends a picture of an answer naming three competitors, and it feels final. Ask that same question again and the list probably will not match the picture.

It cuts both ways, which is the part that costs businesses money. One answer naming the business is not proof of being named, and one answer leaving it out is not proof of being missed. Either way it is a single draw from something that comes out differently nearly every time. A rate over many asks is a fact. One answer is a story with a picture attached.

What can nobody tell a business owner?

Nobody can show what one particular customer saw. That answer was shaped by chance, by whatever that person had typed into the same AI before, and by where they were sitting. It cannot be made to happen again, by any tool, at any price.

So three sales lines are worth hearing twice. That a check reveals exactly what your customers are being told. That a business holds a position, or a rank, inside an AI answer. And that a move since last month proves the work is paying off. The published research behind this page says the opposite of all three: answers "can vary across runs, prompts, and time, making one-off observations unreliable" (Source: Schulte, Bleeker, and Kaufmann, 2026).

How can a business owner check this properly?

By asking many times instead of once, and by writing down what comes back each time. Many asks turn a yes or no into a rate, such as named in 4 answers out of 12. A rate can be set against next month. One answer cannot.

A few habits make all the difference, and none of them cost anything. Ask the way a customer would, naming your trade and your town, rather than naming your business. Log out first, or use a private window, so the answer is not shaped by one person's own history. Put that same question to each of ChatGPT, Perplexity, Gemini, and Claude, several times each, rather than trusting whichever one happens to be open. Write the date beside every count, because the companies keep changing the AI underneath the answers. And keep the questions word for word, so next month compares like with like.

Where does that leave a business owner?

Different people get different answers, so one answer is never a measurement. We ask the questions your buyers really use, many times each, on ChatGPT, Perplexity, Gemini, and Claude. We start logged out. Then we count how often the business is named. The date and the full question list go next to the count. Gemini is asked without letting it search the web, so it answers from what it already knows and names no websites. That note goes next to its number too. The first check is free.