How to check whether ChatGPT knows your business

Check in a fresh chat, signed out or with memory turned off, and run each question twice: once with web search off to see what ChatGPT remembers, and once with search on to see what it can find. Those two answers are different measurements and mixing them is the main reason people get confusing results.

What ChatGPT does differently

It personalises. Saved memory, custom instructions and your own chat history all shape the answer. If you have spent six months discussing your company in that account, it will name your company, and that tells you nothing about what a buyer sees.

It also chooses whether to search. Without search, the answer reflects training data. With search, it reflects pages available now. A company can be invisible in one and present in the other, and the fix is different in each case.

And it varies between runs. The same question asked twice within an hour can produce a different list. So the unit of evidence is a pattern across readings, never a single screenshot.

What that means for whether you are named

Your test conditions are part of your result. A reading taken in your own logged in account, on a question containing your brand name, is not a reading at all. It is your own prompt coming back to you.

So record the conditions with every answer. Date, which mode, signed in or out, search on or off. Without those, two readings a month apart cannot be compared, and you will end up debating a difference that came from the settings.

The two checks, and why they are different

People run one test and get a confusing result. There are two tests, they answer different questions, and both are worth an hour.

The visibility check, blind. Buyer questions with no mention of your company. This answers whether you would be recommended to somebody who has never heard of you. This is the one that goes in the rate.

The knowledge check, named. Ask it to describe your company, your pricing, your coverage and who you serve. This answers whether the facts in circulation about you are correct. It must never be counted in the rate, because a question containing your name can only produce a polite answer.

Keep them on separate sheets with separate totals. The most common measurement mistake in this field is putting the second kind into the first kind's denominator, which produces a number that only goes up.

A worked example

We got a plain lesson about test conditions from our own audit. In September 2026, on our own site aiknowsus.com, Perplexity told us it had run a live search for each question in a batch.

Asked afterwards how many of those questions it had actually searched for, it withdrew the statement. It said it could not honestly substantiate the claim that it had searched for each one. In another batch it said its claim to have searched all five was not adequately supported.

So the mode column cannot be filled in from what the assistant says about itself. That is generated text, not a log. What can be trusted is what is on the screen: the citations, the links, and any marker in the interface showing that a search ran.

Apply that to the two checks on this page. Set the mode yourself, once with search off and once with search on, record what you set rather than what you were told, and keep the two as separate rows. A reading whose mode column came from the assistant's own account of itself is a reading you cannot compare with next month's.

What to change

• Use a fresh chat every time, and sign out or use a clean browser profile. Turn memory off if you stay signed in.
• Never put your company name in a visibility question. Ask what a buyer who does not know you would ask.
• Run each question twice, search off then search on, and log them as separate rows.
• Keep the question wording identical between months. A reworded question is a new question.
• Keep the named knowledge check on its own sheet, with its own total.

The five questions to ask in the named knowledge check

Kept on its own sheet, this check finds wrong facts fast. Ask these five, in this order, and paste the answers into the sheet rather than summarising them.

One, what does this company do and who does it serve. Compare the answer word for word with the one line description you publish.

Two, what does it cost. The most common wrong fact, and the most damaging, because a high wrong price sends buyers away silently.

Three, where does it operate and what does it cover. Cities, states, courts, categories, whatever coverage means in your category.

Four, who are its competitors. This tells you which category the assistant has filed you under, which is often not the one you would choose.

Five, what are its limitations or who is it not for. If the answer invents a weakness you do not have, some source is describing you wrongly and it is worth finding.

Repeat this set quarterly rather than monthly. It changes more slowly than the visibility reading and it produces a work list every time.

What this check cannot tell you

Four limits worth writing into the first report, before anybody over reads it.

It does not show what a specific buyer sees. Their history, their settings and their wording differ from yours. A signed out reading is a fair comparison over time, not a picture of one person's screen.

It does not prove absence. Not being named in six questions means you were not named in those six, on that day, in those conditions.

It cannot explain the result. The citation list gives you evidence to work from. The reason the model chose those names is not visible, and asking it produces generated text rather than a record.

It does not predict enquiries. Being named is not being chosen, and there is no reliable way to attribute a sale to an answer.

Common questions

Can I just ask ChatGPT whether it knows my company?

You can, and that is the knowledge check rather than the visibility check. It tells you whether the facts about you are right. It does not tell you whether you would be recommended to a stranger, which is the thing that changes revenue.

Does using the app without an account work?

It is a reasonable way to get a clean session. Keep the method identical every month, because a change of surface is a change you will mistake for a result.

Should I use temporary chat mode?

Yes if it is available to you, because it avoids history and memory affecting the answer. Note in the log which mode you used, so later readings are comparable.

How many questions for a first check?

Six to ten buyer questions is a usable first hour. Ten questions in two modes is twenty rows, which is enough to see where you stand without turning the check into a project.

The answer named us once and then not again. Which is true?

Neither on its own. Answers vary between runs. Read direction across three or more monthly readings, and treat any single answer as an example rather than a measurement.

Is the API a better way to check?

It is steadier and cheaper to repeat, and it can behave differently from the app your buyers use, especially about searching. Use it if you want volume, label the rows as API, and do not mix them with app rows in one total.

What to do first

One sheet, one row per answer, six columns: date, mode, question, named yes or no, companies named ahead of you, URLs cited. Nothing more complicated is needed for the first few months.

Run ten blind questions in two modes today, then diarise the same reading for next month. Judge by trend, not by any single check, because the same question can answer differently within the hour.

We sell a tool that automates exactly this, so the declaration is due. The manual version above is the same method, and for ten questions on two assistants it is an hour of work.