How to check what ChatGPT, Gemini and Perplexity say about your company
By Miscited · Published 27 September 2026 · 11 min read
One question to ChatGPT, one reply, one screenshot forwarded to the team. That tells you what a single answer said on a single day. It does not tell you what your buyers are reading.
It is worth doing properly because buyers are asking. In a G2 survey of B2B software buyers and decision-makers run in March 2026 (n = 1,076), 51% said they now begin software research with an AI chatbot more often than with Google, up from 29% in April 2025. In the same survey, 64% said they get inaccurate recommendations from AI chatbots often or very often. That is one vendor's survey of one kind of buyer, so treat it as a reason to check rather than as a measure of your market.
Here is a method you can run by hand. It has seven steps, and the first two happen before you ask an assistant anything.
Before you ask anything
1. List your facts, with dates
A claim can only be wrong relative to something. Write down the facts a buyer might act on: prices and plans, limits, integrations, regions, certifications, ownership and leadership. Give each one the date it became true and, if it has stopped being true, the date it stopped. An old price is not a random error. It was right until a date, and without that date you cannot tell a stale answer from an invented one.
2. Write the questions buyers actually ask
Take them from search console, site search, sales calls, support tickets and lost-deal notes rather than from a brainstorm. Group them by intent: questions that name you, questions that compare you with a competitor, and category questions where you may or may not appear at all. Keep the groups apart in everything that follows. They have different answers and different base rates, and averaging them produces a number that describes none of them.
Asking the assistants
3. Ask each assistant, repeatedly
The same question can get a different answer each time it is asked, so one reply shows you what is possible, not what is typical. Ask each question several times on each assistant, in fresh sessions, and record the setup with every run: the assistant, the model where it is shown, whether web search was on, the country, the language and the date.
Then be honest about what the count supports. One wrong answer in two asks reads as 50%, and the 95% interval on that is 9% to 91%. To state a rate with a margin of roughly ten points you need about 100 runs per question group. By hand, use a few runs to find the claims worth worrying about, then measure those properly. The arithmetic is in how many prompts an AI visibility number actually needs.
4. Split every answer into claims
The unit of work is the claim, not the answer. An answer saying the Pro plan costs $49 a month and includes single sign-on holds two claims, and one can be right while the other is wrong. Write each claim as a short sentence that is either true or false, and note the question and the run it came from.
5. Record the pages each answer cites
When an answer shows sources, record the exact URL next to the claim it is attached to. Then open the page and check that it says what the answer says it says. A citation to your own site does not make a claim right: the page may be out of date, or it may not contain the claim at all. That second case is its own defect with its own fix, covered in how to fix a wrong AI answer about your company.
Scoring what came back
6. Compare each claim with your facts
Put every claim in one of four classes, using the dates from step 1.
| Class | What it means | Where the fix starts |
|---|---|---|
| Correct | Matches a fact in force on the day of the answer | No fix. Keep checking it. |
| Stale | Matches a fact whose end date has passed | The pages that still state the old fact |
| Contradicted | Conflicts with your facts and never matched any of them | Tracing the claim to wherever it came from |
| Unsupported | Your list neither confirms nor refutes it | Deciding whether it is true, then adding the fact |
The difference between stale and contradicted is the reason step 1 needs dates. A stale claim was true once, so look first for a page that still says it: an old pricing page, a launch post, a directory profile nobody updated. A contradicted claim has no such history, so finding its source takes longer. We look at where stale answers come from in why AI answers about your company go stale.
Keeping it true
7. Re-check on a schedule
Answers move when models update, when search indexes refresh and when your own facts change. Re-run the same questions on the same assistants on a fixed schedule, and again after a price change, a launch or a correction. Keep a few question groups you deliberately leave alone. Without them you cannot tell whether an answer changed because of your fix or because the model changed the same week.
The shortcut
Done by hand, this is slow, careful work. The free answer audit does a first pass for you. Submit your domain and you get a dated report: facts read from your own site, sampled answers from the assistants the audit is connected to, the pages those answers cited, and the limits of what was covered. It may not ask every assistant named above; the report lists the ones it did. The site facts are provisional until a person reviews them, and answers collected through an API can differ from what the consumer apps show. If the audit finds nothing worth fixing, it says so.
Sources
- New G2 Research: Half of B2B Software Buyers Now Start Their Research With AI Chatbots, G2 press release on PR Newswire, 15 April 2026. Online survey of 1,076 B2B software buyers and decision-makers, March 2026.
Questions people ask about this
How do I find out what ChatGPT says about my company?
Ask it the questions your buyers ask, several times each and in fresh sessions, and record every answer with its date, setup and cited pages. Then split the answers into single claims and compare each one with a list of your facts that records when each fact became true and, where it applies, when it stopped.
How many times should I ask an assistant the same question?
More often than feels necessary. One wrong answer in two asks reads as 50% with a 95% interval of 9% to 91%, which fits almost any true rate. To state a rate with a margin of roughly ten points, you need about 100 runs per question group. A handful of runs can show that a problem exists, but it cannot size it.
What is the difference between a stale and a contradicted AI answer?
A stale claim matches something that used to be true, such as a retired price or a former owner. A contradicted claim conflicts with your facts and was never true. The fixes differ: a stale claim usually leads back to an old page that still states the expired fact, while a contradicted claim has to be traced before it can be fixed.
How often should I re-check what AI assistants say about my company?
On a fixed schedule, and again after anything that changes your facts: a price change, a launch, a rename or a correction you made. Keep some question groups you deliberately leave alone as controls, so that when an answer moves you can tell whether your fix moved it or a model update did.