When an AI answer states something false about your company

An engine says your company was founded in a city you have never had an office in, or that a product was discontinued, or that pricing starts at a figure you have never charged. The instinct is to look for the place to file a correction. Some of those places exist. None of them works the way a correction to a database works.

Two failure modes, and the difference decides the fix

Retrieval error. The engine found a real document and repeated it. The document is outdated, or it is about a different company with a similar name, or its author got the fact wrong. The claim exists somewhere in text, which makes it the tractable case: documents can be edited, and the fix is source-side.

Generation error. The model produced a plausible detail that appears in none of the sources it retrieved. Language models compose fluent text, and where the retrieved material is thin they fill the gap with something statistically likely. There is nothing to correct at the source, because there is no source. The fix is indirect: make the correct fact abundant and easy to retrieve, so there is less room to fill in.

Confusing the two is expensive. Hunting for a source that does not exist wastes weeks, and treating a real bad source as a model quirk leaves a wrong page live.

How to tell which one you have

1. Ask the engine for its sources. Most surfaces show citations, and most will list them if asked directly.
2. Open each cited page and search it for the claim. If the false statement, or a version of it, appears there, you have a retrieval error and the URL to fix.
3. If it appears in none of the cited pages, search the open web for it before concluding the model invented it.
4. Repeat across engines, phrasings and, where relevant, languages before concluding anything. Generated answers differ between identical runs, so a claim seen once may not reproduce, and one clean run does not mean it is gone. The discipline is the one set out in measuring AI search visibility honestly.

The remedies, in order

Correct the upstream source where one exists. Start with your own pages, because you control them and they are usually the culprit: superseded pricing pages, old press releases, stale about text, documentation for a version you no longer ship. Then contact third-party publishers, state the correct fact with evidence, and ask for a dated update. A corrected source removes the input instead of arguing with the output.

Publish an unambiguous statement of the correct fact. One canonical page, plain sentences, the subject named rather than referred to as "it", so the passage can be lifted out and stay true. How to write one is in the grounding page.

Disambiguate from same-name entities. A large share of wrong facts are correct facts about a different organization. Consistent naming and clear statements of what you are and are not reduce the collision, as described in entities and how engines identify your brand.

Keep the correction fresh. Recency affects citation: a study of 7,683 pages and 47,097 citations found Gemini cited content updated within the past year 78 percent of the time, ChatGPT 73 percent and Perplexity 65 percent. More in freshness.

The reporting channels that exist

Google documents in-product feedback for AI Overviews. Its help page describes thumbs up and thumbs down icons at the bottom of each overview, a "Share more feedback" or "Report a problem" option and a text field for detail. The same page states that AI Overviews can and will make mistakes. It is a feedback form: no case number, no published timeline, no reply.

Google's legal removal process is a separate route, for content that is unlawful rather than merely wrong. Its defamation guidance states that a request can be made by the person or business being defamed, or by a legal representative, and requires the exact URLs and an explanation of why the statements are false and damaging. If approved, Google restricts access to the content on the Google service for the country where it is considered defamatory.

OpenAI documents a privacy route, not a business-facts route. Its help article on personal data removal from ChatGPT describes a "Remove my personal data from ChatGPT responses" request through the OpenAI privacy portal, available where information about a person is inaccurate or no longer appropriate. OpenAI states that requests are assessed case by case, that it "cannot independently investigate disputed facts", and that they are not always approved. Nothing there covers a company correcting a claim about the company.

Anthropic documents a correction right in its privacy policy. The policy states that Anthropic "cannot guarantee the factual accuracy of Outputs", that a correction request for factually inaccurate personal data in outputs receives "a reasonable effort", and that because of "the technical complexity of our large language models, it may not always be possible". As with OpenAI, the right covers personal data, not statements about a company.

Perplexity documents a general inaccuracy report. Its help article on reporting incorrect or inaccurate answers describes a flag icon below the answer, a support ticket, or an email to [email protected], and asks for the URL of the query and a description of the error. Misinformation and outdated information are among the listed issues.

These are feedback channels. None is a correction guarantee, none publishes a response time, and none commits to changing a specific future answer. File the report, because it is cheap, then proceed with the source-side work as though no reply is coming.

When the statement is unlawful

Defamatory or otherwise unlawful statements sit on a separate track from product feedback. Each company operates a legal or removal process for them, and local law may provide remedies no company process offers. Those routes carry evidentiary requirements and consequences, and the decision belongs with counsel, not with SEO.

What you can and cannot change

You cannot patch a model's weights, and you cannot edit the answer text. No vendor can, and nobody sells access to the step where the answer is composed. What you can change is what the model retrieves and what the public record says: corrected sources, a canonical statement of the fact, a clearly distinguished entity, and recent dates on all of it. Then measure. Run the same prompts on a fixed schedule, count how often the false claim still appears, and compare month over month. A falling rate is the realistic outcome.