How to find out whether AI mentions your business

Ask the assistants the questions your buyers ask, never naming your own company, and record who gets named and which pages get cited. That is the whole method. There is no dashboard inside ChatGPT or Gemini that will tell you, so the reading has to be made rather than looked up.

What is actually happening underneath

Every check is a sample, and it is a sample of two things at once. You sampled the questions, because you did not ask every question a buyer could ask. And you sampled the answers, because the same question asked twice can come back different.

That is why one question answered once is not a measurement. It is an anecdote. Asking the same set of questions repeatedly and watching the pattern is a measurement.

Two settings quietly change what you see. Your chat history and any saved memory make the assistant answer you differently from how it answers a stranger. And whether the assistant searched the web for this answer changes the sources completely. So do the check in a fresh chat, signed out if you can, and note whether citations appeared.

The five things to fix before your first reading

A reading taken without these is not wrong so much as uninterpretable, and people usually discover that after three months of collecting rows.

One, the question set, written down in a file. Not in somebody's head and not retyped each month. The set is the denominator, so it is part of the number.

Two, no brand questions in that set. A question naming your company produces a polite answer about your company and measures nothing. Keep those in a separate sheet for a separate purpose.

Three, a clean session. Fresh chat, signed out or with memory off, or a separate browser profile you only use for this.

Four, at least two assistants. They answer the same question differently on the same day. One app is one opinion, not your visibility.

Five, a mode note. Record whether the answer carried citations. A cited answer and an uncited one are two different measurements and blending them produces a number nobody can explain later.

What it means for a business

You need a record, not a screenshot. One sheet, one row per answer, with these columns:

• Date, and which assistant.
• The question, copied exactly as you typed it.
• Whether the answer carried citations.
• Were you named, yes or no.
• If named, where in the list you appeared.
• Every company named ahead of you.
• Every URL cited.
• The exact sentence written about you, if there was one.

Keep the wording of the questions identical between readings. Change a word and you have started a new measurement without meaning to. If you must change a question, add it as a new one and keep the old one running.

The three columns people skip are the three that pay. The competitors named ahead of you tell you who the assistant thinks the category is. The cited URLs tell you which sources it trusts, which is the shortest route to changing the answer. And the sentence about you, compared with the description you publish, tells you whether it is reading your version of you or somebody else's.

A worked example

We ran this method on ourselves first. In September 2026 we asked Perplexity six questions about our own category, with no brand named in any of them. aiknowsus.com is us.

Asked afterwards to audit its own answers, it reported that it had not cited or recommended aiknowsus.com in any of the six. Not named far down a list, not described without a link: absent. There was no position for us to hold.

Written honestly, that reading is "named in 0 of 6 answers, from 6 category questions on 1 assistant, no brand questions in the set". Note what is not written: a percentage. Zero out of six is a count anybody can audit, and a set that size cannot carry a rate.

A reading like that looks like bad news and it is the cheapest kind of finding there is. It says the work has not started, it costs an hour, and it leaves a dated row that next month's reading can be compared against. We keep ours for exactly that reason.

What this method cannot tell you

It gives you a reading of your own question set on the apps you chose, on the day you asked. It does not give you a share of anything, because nobody knows how often real buyers ask these questions or in what words.

Four other limits worth writing in the report itself. A small set cannot produce a percentage, and a doubled percentage on a small set often means no change at all. You are seeing answers, not the logic behind them, so every explanation you write is inference. Signed out readings do not tell you what a logged in buyer with history sees. And none of it predicts revenue, because being named is not the same as being chosen.

Say these in the report. A reading presented with its limits survives the meeting. One presented as a certainty gets discarded the first time it moves for no reason.

Common questions

Is there a setting in ChatGPT or Gemini that shows my brand mentions?

No. There is no publisher console or brand dashboard inside these apps. Anything that reports your mentions is asking questions and reading answers, whether that is you in a spreadsheet or a tool doing it at scale.

How many questions do I need?

Enough to cover every kind of buying question you sell into, which usually means five to ten questions for each zone you serve. Ten questions is a real start. Five questions is five answers, not a rate. Our own tracked sets begin at 100 questions for this reason.

Should I use the free app or the API?

Start with the app, because that is what your buyer uses. The API is steadier and cheaper to repeat, and it can behave differently from the app, especially about whether it searches. If you use the API, say so in the report and do not compare those rows with app rows.

Can I just ask the assistant whether it knows my company?

You can, and it is a different test. That tells you whether it can describe you, which is worth knowing. It does not tell you whether it would recommend you to somebody who has not heard of you, which is the thing that changes revenue.

What if I am named but described wrongly?

Treat that as more urgent than absence. A wrong price or a wrong segment actively routes buyers away while sounding helpful. Fix the source that carries the old fact, correct your own pages in place, and re-check monthly.

How long does one reading take by hand?

Ten questions on two assistants is about an hour, including writing the rows. The cost is not the hour. It is remembering to do it every month, which is why a fixed date in the calendar matters more than the tooling.

What to do first

Ten questions, two assistants, one sheet, today. Write the questions as a buyer who has not decided yet would type them, including the qualifiers a buyer would include, such as the city, the company size or the budget.

Do not put your brand name in any of the ten. Keep a small, clearly separate set that does name you, for checking whether the assistants have your pricing right or have confused you with a similarly named company. Useful, and never counted in the same number.

Then diarise the same reading for the same date next month. A modest baseline you actually repeat beats a perfect one you take once.