Skip to content

ResourcesAEO Handbook

How to find out whether AI assistants describe your company inaccurately

Ask an assistant about your company and it answers with claims, some of them wrong. Logging each claim with its source shows which, and where the fix belongs.

Updated Jack F, Co-founder

SentimentAEObrand accuracyAI answersMicrosoft CopilotPerplexityBrand Facts

Part of the Answyn AEO Handbook.

Now readingThe short answer

Ask Copilot, Perplexity and the other assistants your buyers use the questions those buyers ask, branded and unbranded, several times each, because answers change between runs. Log every claim about your company with the page it cited. Decide whether each claim is false, out of date, a third-party error or accurate but unflattering, then fix the source before you recheck. Marking a claim wrong in a tool does not change the answer.

How do you find out whether AI assistants get your company wrong?

You find out by asking the assistants your buyers use the questions those buyers ask, writing down every claim each answer makes about your company with the page it came from, and checking each claim against your own records. Do it on each assistant separately, and more than once, because the same question can get a different answer tomorrow.

The worry usually arrives as one question, but it is two. An answer can be wrong about you: an old price, or a feature you never had. Or it can be right about you and unflattering: a fair caveat, or a rival preferred on one criterion. The first needs a correction at the source. The second is a positioning question, and there is nothing to correct.

A sentiment score only sees the second, so a wrong claim that flatters you scores as positive. The method below keeps accuracy and tone in separate columns.

We run the same checks on Answyn. The figures on this page come from our own tracking of our category: 25 questions, asked of ChatGPT, Google AI Overviews and Claude in the UK between 6 October and 4 November 2026, 1,575 answers in all. None of the 63 answers to this page's own question named Answyn. The wrong claims in the worked example come from Answyn's Brand Facts for the same period.

Build a claim ledger

A claim ledger is a running list of the claims AI answers make about your company, each with its source, whether it is correct, and the fix. A spreadsheet is enough to start.

  1. Write down the questions buyers ask, in their words, and split them into branded and unbranded. Branded questions name you ("is [your company] any good for a small team?") and show what an assistant says when asked directly. Unbranded questions do not ("which tools do this for a small team?") and show whether you come up at all, and how you are framed when you do.
  2. Ask each question on each assistant separately, because they search differently and cite different sources. Then ask again on other days.
  3. Record every factual claim an answer makes about you, in the answer's own words, with the link it cited.

Answers move between runs. A 2026 study of repeated runs on ChatGPT, Gemini, Google AI Mode and Perplexity found that the sources cited by two runs of the same prompt on the same day overlapped by only 32 to 43% on average. It suggested at least seven runs per prompt per day for monitoring brand visibility, a figure drawn from one dataset rather than a rule (Schulte, Bleeker and Kaufmann, arXiv 2604.07585). In our own tracking, ChatGPT searched the web for all 750 of its answers, and across 30 runs of one question it used between 26 and 95 different search wordings, depending on the question. One answer is a sample of one.

The claims worth recording are the ones a buyer would act on:

  • Prices, plans and what each plan includes
  • Who can buy, in which markets, and how access works
  • Features you have, and features the answer says you lack
  • How you compare with named rivals on price, coverage or support
  • "Best for" labels, which tell a buyer whether you are meant for them
  • Caveats and warnings, such as limits, gaps or complaints the answer repeats

Give each row the question, the assistant, the date, the claim, the cited page, whether the claim is correct, what kind of problem it is, the fix, and the date you rechecked. The last three columns are what turn a list of grievances into a plan of work.

Classify each claim, then trace where it came from

Classify each claim before you fix anything, because the six kinds of problem need different fixes, and one needs no correction at all.

KindWhat it looks likeWhere the fix usually is
FalseThe claim has never been true: a plan, feature or offer you do not haveOn the page the claim came from, plus one plain statement of the truth on your own site
Out of dateIt was true once: last year's price, a retired plan, an old product nameWherever the old version still lives, starting with your own pages
Ambiguous on your own siteYour own pages support two readings, or contradict each otherYour own site: one clear statement, in one place
Third-party errorA review site, directory, roundup or rival's comparison page states it wronglyWith that publisher: a correction request, or a profile you can claim and edit
Accurate but negativeTrue, and framed as a drawbackNo correction. Say plainly who you are for, and compare fairly where a rival is preferred
Missing contextTrue as far as it goes, but without the condition that changes its meaningPut the condition next to the claim on your own page

Then trace each claim. Open the pages the answer cited and find the sentence that says it. If a cited page says it, that page is the source, whoever owns it. If no cited page says it, the assistant has combined sources or filled a gap on its own. There is then no page to correct, and the fix is to make the true version easy to find on pages the assistant already reads.

Tracing can change the fix completely: a claim that looks like a rival's spin can turn out to come from an old help article on your own site.

What our own claim ledger shows

Two claims the engines made about Answyn between 6 October and 2 November 2026, taken from Answyn's own Brand Facts.

ChatGPT said both of the Answyn plans it described "advertise a seven-day free trial" in 6 answers between 10 October and 2 November 2026. Answyn has no self-serve sign-up or free trial, so the claim is wrong today: access is by invitation, starting with a 30-minute call. Our pages already state the true version, so the fix is wherever ChatGPT found the claim, and the pages cited in those six answers are where the trace starts.

Google AI Overviews said on 6 October 2026 that "industry roundups" rate Answyn the best-value AEO platform. We know of no roundup that does, or any other source that supports it. That makes the claim false and unsupported. There is no page to ask anyone to correct, so in the ledger the row is marked incorrect and rechecked, and that is all the fix there is.

Both claims flatter us. A tone score could count both answers as positive, and both would mislead a buyer: one sends them looking for a trial they will not find, the other quotes a verdict we cannot find anywhere. That is why accuracy and tone sit in separate columns.

Fix the source of truth first, then recheck

Fix in order: your own site, then the pages that carry the claim, then recheck. Asking a publisher to correct something your own pages still say wastes the request.

On your own site, give each fact one plain statement on the page that owns it: prices and access on the pricing or enquiry page, features on the product page, markets wherever you say who you serve. Put it in the page text rather than in an image or a widget that only appears after JavaScript runs, because AI retrieval crawlers largely do not run JavaScript. Then remove or date the old version everywhere you control, old blog posts and help articles included.

For a third-party error, write to the publisher with the correct fact and a link to where you state it, and claim any directory or review profile you can edit yourself. When the cited sources are roundups, reviews and threads you do not own, that is where the claim lives, and how brands earn citations in AI answers covers why so much of what assistants cite sits off your site.

Then recheck: the same questions, on the same assistants, over the following weeks, with each result added to the ledger. Assistants notice changed pages at different speeds, and only some give you any way to tell them a page has changed; our guide to content freshness and AI citations goes through what each engine reads. Judge a fix on the run of answers, not on the next one.

In Answyn, this is what Brand Facts and Factcheck are for. Brand Facts, on the Sentiment tab of AEO Hub, lists the claims the engines made about your brand. You mark each one correct or incorrect and add a short correction where it is wrong, and reviewed facts become the approved source of truth for later analysis. They do not rewrite past answers or change what an assistant says. Factcheck, which is in beta, rechecks one question over time, such as the question behind our trial claim, on ChatGPT or Google AI Overviews in one market. It saves each answer with its sources and the claims about the brand, and you judge whether the claim has gone.

What Copilot, Perplexity, ChatGPT and AI Overviews show you

The method is the same everywhere. What differs is how much each assistant shows you about its sources.

Microsoft says Copilot Chat turns your prompt into a short search query and sends it to Bing, and when it uses web search, a Sources button under the response shows the exact query and the sources it used. On work and school accounts the organisation can switch web search off, and answers then come without web results, so there is no source list to trace (Microsoft Support). If your buyers use Copilot at work, ask with web search on and with it off. Microsoft's support pages on Copilot, the web search page among them, are also what ChatGPT kept citing on this subject: across our three sentiment questions, it cited support.microsoft.com 57 times.

Perplexity's help centre says every answer carries numbered citations that link to the original sources. Record the numbered source next to each claim in your ledger, rather than one source list for the whole answer.

ChatGPT searched the web for every one of the 750 answers we collected, usually with two or three searches per answer in the runs we read, and the wording changed from run to run. For our category its most-cited sources were vendors' own product, pricing and feature pages, roundups, and official documentation. A rival's comparison page that describes you is a third-party source like any other, and it belongs in the ledger.

Google AI Overviews leans on video and social posts. Asked the question this page answers, it cited youtube.com in all 30 of its answers, and across our three sentiment questions it cited youtube.com 98 times and trysight.ai 54 times. When the source of a claim is a video or a post, the fix is a correction request to its creator, or a clearer video or post of your own, rather than a page edit.

Know which of these a tool covers before you rely on it. Answyn collects Gemini, ChatGPT, Google AI Overviews, and Google AI Mode every day, and can add Microsoft Copilot, Perplexity, and Claude by agreement. Each plan includes the engines that business needs. If your plan does not include Copilot or Perplexity, run the ledger for them by hand. It works the same way.

If you want the collecting done for you, Answyn does it on the engines each plan includes. Access is by invitation, and every plan is quoted for the business using it. Enquire about access, or read how Answyn pricing works.

Questions people ask

Can I check what Copilot or Perplexity say about my company without a tool?
Yes. Ask each one your buyers' questions by hand, record every claim and its source in a spreadsheet, and ask the same questions again over several weeks. Copilot Chat shows the query it sent to Bing and the sources it used under a Sources button, and Perplexity numbers its citations, so tracing a claim takes a click. What a spreadsheet makes slow is repetition, and repetition is what separates a one-off from a pattern.
Does Answyn track Microsoft Copilot and Perplexity?
Not by default. Answyn collects Gemini, ChatGPT, Google AI Overviews, and Google AI Mode every day, and can add Microsoft Copilot, Perplexity, and Claude by agreement. Each plan includes the engines that business needs.
Can I stop ChatGPT saying something wrong about my company?
Not directly. Answyn does not change how ChatGPT answers, and you cannot edit an answer yourself. What you can change is what it reads: correct your own pages first, ask the publishers of the pages it cites to correct theirs, then recheck the same question over the following weeks to see whether the answer moves.
What if an AI answer about my company is accurate but negative?
Then there is nothing to correct, and it is a positioning question instead. Check whether the answer reflects a real trade-off, such as a market you do not serve, and if it does, say plainly on your own pages who you are for and who you are not for. If the negative framing comes from a comparison with a rival, a fair comparison page of your own gives the assistant something other than the rival's page to read.
Why does the same assistant say different things about my company each time?
Because each answer is built from fresh web searches, and the searches vary. In Answyn's own tracking, between 6 October and 4 November 2026, ChatGPT searched the web for every answer, and across 30 runs of one question it used between 26 and 95 different search wordings, depending on the question. Different searches find different pages, and different pages make different claims.

Sources

Answyn's own figures on this page come from Answyn's tracking of its own category, 6 October to 4 November 2026: UK market, ChatGPT, Google AI Overviews and Claude, 25 tracked questions, 1,575 answers. The two wrong claims in the worked example come from Answyn's Brand Facts for the same period.

Run the numbers on your own brand.

Answyn puts the questions your buyers actually ask to every major engine, daily, and records who gets named and who gets cited.