An AI model told a prospect that you do not offer the exact service you specialize in. Another one recommended a competitor and described your pricing wrong. If you have seen this happen, you know the strange powerlessness of it: there is no support line for ChatGPT answers. There is, however, a repeatable process, and this article walks through it.
How can I tell what AI is saying about my brand?
Ask the models the questions your customers ask, in a private browser window, and record the answers. Cover your brand name, your category ("best [category] in [country]") and your reputation ("is [brand] reliable"). Run each question on ChatGPT, Gemini, Perplexity and Copilot at minimum, because they answer from different sources and get different things wrong.
For a one-off audit the manual route works. For anything structural you want daily tracking with stored responses, because errors appear and disappear between model updates, and you cannot fix what you saw once and cannot reproduce. Trackbase stores every response across 11 models, which makes the "when did this start" question answerable.
What types of misinformation show up in AI answers?
Four patterns cover almost everything we see in tracked responses:
Outdated facts. Old pricing, discontinued products, a previous positioning. The model learned a snapshot and never got the update.
Fabricated details. Features you never built, offices you never opened, integrations that do not exist. Models fill gaps with plausible guesses.
Misattribution. Your achievements assigned to a competitor, or their weaknesses assigned to you. Common when brand names are similar.
Entity confusion. The model merges you with another company that shares your name. This one is the most damaging, because every downstream fact inherits the confusion.
Where do AI tools get information about my brand?
From three layers. Training data: the web as it existed when the model was trained, which can be a year or more old. Retrieval: live sources the model consults at answer time, such as review sites, news, comparison articles, Reddit and Wikipedia. And structured signals: your own site, schema markup and business profiles.
The practical consequence: most misinformation does not come from your website. It comes from what third parties wrote about you, or failed to write. That is where the fixing happens.
How do I fix incorrect information about my brand in AI answers?
Work from the source backwards. The sequence that works:
Find the source of the error. Perplexity and Google AI Overview cite their sources directly. For other models, search the exact wrong claim; it usually traces to a specific page. The source view in Trackbase ranks the domains models cite about you, which shortens this step considerably.
Correct it at the source. Outdated directory entry: update it. Wrong fact in a comparison article: email the author, most respond. Stale review page: push satisfied customers to leave current reviews.
Publish the correct answer yourself. A clear page on your own domain that states the fact plainly ("[Brand] pricing starts at X", "[Brand] is a [category] company founded in [year]") gives retrieval-based models something better to quote.
Strengthen your entity. Consistent name, description and schema markup across your site, LinkedIn, Crunchbase and directories. This is the fix for confusion cases.
Report where reporting exists. ChatGPT and Gemini both have feedback mechanisms on answers. Low individual impact, but free and occasionally effective.
How long does it take for corrections to appear?
Weeks to months, and it varies sharply per model. Retrieval-based answers (Perplexity, AI Overview) can pick up a corrected source within days. Training-data answers move only when the model is updated. Plan for a quarter, measure weekly, and resist the urge to declare victory after one good answer, since answers vary between runs.
How do I know if the answers are improving?
Track the same prompts daily and watch three numbers: the share of responses containing the wrong claim, your sentiment score, and your presence rate. In our own tracking data, sentiment classification across 535 recent citations for one brand split into 314 positive, 121 neutral and 100 negative. That negative fifth is where the correction work lives, and watching it shrink week over week is how you know the process works.
FAQ
Can I make ChatGPT delete a wrong claim about my brand?
Not directly. You can report the answer, correct the sources it learned from, and publish better information. Removal on request only realistically exists for legal categories such as defamation or privacy violations.
Should I respond to every single error?
No. Prioritize by damage: errors on money prompts (category recommendations) and reputation prompts first. A wrong founding year in a rarely asked answer can wait.
Does correcting my own website help at all?
Yes, for retrieval-based models that read your site at answer time, and as the canonical source other writers copy from. It is necessary, just not sufficient.
Why does one model get it right and another get it wrong?
Different training cutoffs and different retrieval sources. This is also why tracking one model gives you a false sense of security in either direction.
What if a competitor is spreading the misinformation?
Document it, correct the record on neutral third-party surfaces, and outrank the claim with better sources. Direct disputes rarely help; better evidence usually does.
