If an AI assistant says something false about your company, I would preserve the exact answer before changing the website. The useful response begins with a specific claim and the evidence that contradicts it, not a general campaign to make the business more visible.

You can correct sources you control and request corrections elsewhere. You cannot assume that either action directly edits every answer a model may generate.

Capture the incident while the context is available

Save the prompt, the complete answer and the sentence that is wrong. Include the visible citations, language and any location context. Record the platform and mode as displayed, without guessing which underlying system produced the response.

This context helps distinguish different problems. An assistant might describe a former service as current, confuse two businesses, misread a valid source or make an unsupported statement. Each calls for a different investigation.

I would also note whether the answer occurred in a fresh conversation or after several questions. Earlier messages can change what the assistant is being asked to infer. A screenshot of the false sentence alone may be enough to show concern, but it is rarely enough to reconstruct the task.

Keep the original record even if a second attempt produces a correct answer. Variation is part of what needs to be understood. One favorable response does not make the earlier incident imaginary, and one unfavorable response does not establish how often customers encounter it.

Establish the correct statement independently

Before asking where the assistant went wrong, agree on what it should have said. The relevant person inside the company should confirm the service, location, leadership role or other fact in dispute.

Separate a factual error from wording the marketing team dislikes. A description can be unflattering, incomplete or differently phrased without being false. The correction process is stronger when the disputed claim can be expressed precisely.

For a hypothetical consultancy, “serves only local retailers” is a factual restriction that may contradict its actual offering. “One of several consultancies serving retailers” may simply omit the company's preferred positioning. Those are different editorial concerns.

I would prepare a concise corrected statement and its supporting public evidence. If the only support is an internal assertion that has never been published, the first task may be to make accurate information available in a suitable place.

Find out what the cited sources actually say

Open each relevant citation and locate the passage that could support the disputed sentence. Check whether it refers to the same company and whether it describes a current state or a past event.

OpenAI's search guidance notes that search responses may provide citations. Those citations are a useful starting point, but their presence should not substitute for reading the pages. I would treat them as evidence to examine, not a complete explanation of how the answer was produced.

A source may be correct while the answer is wrong. For example, a report about a former executive appointment can be accurately written and still be misrepresented as the person's current role. Rewriting that historical report is not necessarily the right remedy.

If the cited page contains the same mistake, record the matching statement. If it does not, keep that distinction visible. “The answer cites this page” is a narrower and more defensible conclusion than “this page caused the error.”

Handle answers without identifiable sources differently

Do not assume every answer used live web retrieval. If no source trail is available, it may not be possible to identify where a particular statement came from.

I would still search for obvious public contradictions: old directories, unclear company descriptions or another business with the same name. Finding them creates useful correction candidates, but does not prove they generated the answer.

Asking the assistant to explain its mistake can produce leads. It does not create independent evidence of provenance. A confident account of its own reasoning should not become the basis for accusing a publisher of supplying the false information.

The investigation can therefore end with a partial conclusion: the claim is wrong, the available public record has specific weaknesses, and the exact source of the generated statement remains unresolved. That is enough to guide useful work without inventing certainty.

Correct the relevant page, not every page

Publish a clear statement where a reader would expect to find it. A service-scope error belongs on the service page; a leadership error belongs in an appropriate current biography; entity confusion may require a clearer company identity and location.

I would use plain, visible text and align existing metadata and structured data with it. The correction should help a customer understand the company even if no assistant ever quotes it.

Avoid creating a collection of repetitive pages that restate the same fact for different AI products. That can introduce maintenance problems and make the site less coherent. One reliable explanation, supported by appropriate pages, is a more manageable public record.

For third-party errors, send the publisher the exact passage, corrected fact and supporting evidence. Record the request and response. The company controls the quality of that request, not the editor's decision or the timing of subsequent processing.

Keep access checks and feedback in their proper place

If the corrected material cannot be accessed, the technical team should investigate relevant crawler rules and infrastructure restrictions. Perplexity's crawler documentation, for example, distinguishes search discovery from user-triggered visits. Different access paths should not be treated as one universal AI setting.

For Google's AI search features, its documentation describes indexing and snippet eligibility as requirements for supporting pages. Eligibility does not guarantee selection. Opening access is therefore a technical action, not an answer-correction promise.

Use an available feedback mechanism when appropriate, with the false statement and evidence. I would record feedback as submitted, rather than marking the incident resolved. Interface options can vary, so a platform-specific procedure should be checked when someone carries it out.

No access setting or feedback submission should be represented internally as a way to rewrite a model's knowledge on demand.

Retest without coaching the desired result

Repeat the original question in a fresh conversation under a comparable context, then try a small set of realistic variants. A prompt that supplies the correction and asks for confirmation tests a different task from the customer's original question.

Save both correct and incorrect responses. Distinguish source corrected, page accessible, feedback submitted and answer improved. These are separate outcomes, and the distinction prevents premature closure.

I would prioritize continued work according to the consequence of the error and how consistently it is observed. The goal is an accurate, accessible public record and a documented response to remaining mistakes. That is work a company can manage, even when it cannot control every sentence an assistant produces.