How to Report Incorrect Information to ChatGPT, Google, Perplexity and Other AI Products

Two labels mark evidence boundaries on this page. Documented means the statement it sits beside is directly supported by the linked provider documentation, quoted, and applies to that provider alone. Recommendation means a workflow or record that CoreAEX prescribes and no provider documents. Untagged text is ordinary explanation. Where a provider's documentation says nothing about what happens after a report, this page says the documentation does not address it. It does not fill the gap with an inference about how the provider works internally.

Every provider whose documentation we read for this page documents a way to tell it that a response was wrong. None of that documentation describes a business-facing process for correcting a factual claim about a company. The routes that exist are of three kinds: per-response feedback, policy or illegal-content reporting, and privacy or legal removal scoped to individuals. OpenAI is the only provider reviewed here that describes a specific potential action tied to reported domains or content: conditional review, and possible source-level mitigation intended to reduce future reliance on unreliable sources. The others describe feedback as contributing to broader product or safety improvement - Google says it helps improve AI Overviews, Microsoft says it reviews feedback to provide a safe experience, Perplexity says a report is taken to improve response quality - without connecting any of it to correcting the reported claim or a future answer.

Report it where a provider documents an applicable response-feedback route - after you have captured the incident and corrected the published reference you control, because both make the report worth reading. Then treat the submission as recorded and requested, not fixed: the documentation reviewed here carries no outcome commitment for an ordinary company fact. Read the next section before you write anything, because which family a route belongs to sets what it is built to receive. Reporting to a provider is one step among several, and where it fits in the full capture-to-re-test workflow is set out on its own page.

Three Families of Mechanism, and What Each Acts On

The routes providers document fall into three families, and they are documented to receive three different kinds of thing. What each family receives is documented; what any of them changes afterwards mostly is not. Sending the right report to the wrong family is the most common wasted effort here.

For an ordinary wrong price that raises no policy, legal or personal-data issue, use the provider's factual-response feedback route where one is documented. Recommendation The policy and privacy routes reviewed here are not documented for ordinary company product facts. A form may still technically accept the submission - that is not the same as the route being documented to handle it, and a report filed outside a route's stated purpose is one you cannot chase.

None of the three is a correction process, and that is a statement about a defined search rather than about the industry. We read, in full: OpenAI's reporting-content, ChatGPT-search and personal-data-removal pages; Google's AI Overviews feedback, knowledge-panel feedback, AI-features and Search generative AI control pages; Perplexity's report-an-answer page; Microsoft's Bing concern-reporting page and Copilot privacy FAQ; and Anthropic's incorrect-responses and privacy-rights pages. Every documented route we found was one of the three families above. None was a business-fact correction process. That does not establish that no provider ever acts on such a report - only that none of these pages documents a route for it.

Provider by Provider: The Route, and What It Does Not Say

Each row below is that provider's documentation and applies to that provider alone. There is no common process, no common scope and no common commitment. The "what it does not say" column records either the absence of a sentence or a limit the documentation states in its own words - never an inference about how a provider works internally. Documented

Two of those rows deserve reading twice, for opposite reasons.

OpenAI's is the only sentence in this set that describes any downstream effect, and it is the easiest to overstate. Read it clause by clause: reported content may be reviewed; the team may apply filters or other mitigations; the effect described is to help prevent reliance on unreliable sources; and it concerns future responses. It is doubly conditional, it operates on a source rather than on a fact, and it says nothing about the response you reported. A report about your pricing page is not, on this documentation, a request to correct a price - it is a report that a domain may be unreliable, which is a different thing to ask for and worth being deliberate about.

Bing's is the only one that states its own ceiling out loud. Most of the pages here are silent about outcomes; this one says plainly that reporting will not necessarily result in removal. Silence and a stated limit are not the same thing, and neither is a commitment.

Two of these pages display only a relative update date rather than an absolute one, so we cite them to the date we read them and record the relative display as a property of the page. A relative date cannot be checked against later, which is a reason to open the page yourself rather than to trust this table's freshness.

Privacy and Legal Removal, Which Is Scoped to People

Two providers document a removal route, and both are about personal data rather than product facts. Documented

OpenAI documents a personal-data removal route where Requests are assessed on a case-by-case basis and involve balancing an individual's privacy rights against other important considerations, such as freedom of expression and the public interest and is explicit about the boundary of what removal achieves: Removing your personal data from ChatGPT prevents it from appearing in ChatGPT responses but does not remove it from external websites or search engines That is stated as prevention at the level of the output.

Anthropic documents a privacy-rights route relating to its training data, displayed as last updated July 8, 2026 and scoped to personal data. Its description of what such a request can achieve is careful: We use techniques that can reduce the likelihood of data from our training data being reproduced by Claude and Models do not store text like a database, and they do not have access to or pull from the original training data once the models have been trained It makes no mention of correcting factual information about a company.

Worth reading those two together, because they are the reason this page does not describe a "remove it from the model" route. One provider documents preventing personal data from appearing in responses; the other documents techniques that reduce the likelihood of reproduction and says plainly that a trained model does not hold the training text to begin with. Neither describes editing model parameters, and neither route is documented for ordinary facts about a company, such as its price, features or integrations. Filing through one of these routes is therefore not a documented shortcut to correcting a company fact.

The privacy and legal-removal routes located for this review are documented as applying to personal data about individuals, not ordinary product facts about companies. Questions involving defamation, regulated disclosures or other legal obligations require advice from the company's counsel.

Five Things That Get Called "Correction"

Most confusion about what a report achieves comes from five separate mechanisms sharing one word. They act on different objects and different people own them.

Exclusion is a sixth thing and it is not on that list, because it is not a correction at all. A control that removes your content from a feature replaces a wrong answer with no contribution from you - not with a right answer. It has its own section below.

The practical consequence is that reporting sits in the fourth row. It does not edit your pages and it does not directly request an index update. Beyond OpenAI's conditional source-level mitigation statement, the documentation reviewed here does not explain whether or how a report affects later retrieval, later responses or model behaviour - and that silence is recorded here as silence rather than filled in either direction. Correcting your own estate and correcting third-party sources are the first row, and they are the rows you can act on directly.

Google's Exclusion Control: A Documented Non-correction

What to Put in a Report

Start from what the route is documented to receive, then write to that. Recommendation Perplexity is the only provider in this set that publishes what a report should contain - the query URL, and a description of the error and the expected result - and that shape is a reasonable default elsewhere, on the understanding that no other provider has asked for it.

• The exact prompt, and the response, captured. Most of these forms are attached to a specific response, so file the report while you still have it in front of you. Capturing the answer properly takes a few minutes and is worth doing before you click anything.
• The single claim you are disputing, quoted. One claim per report. A report listing six problems is a research task for whoever opens it.
• The correct value, and the URL that publishes it. If your own published source does not yet state it, fix that first - you are otherwise asking a provider to prefer an assertion over a page.
• The date the correct value took effect, which distinguishes a wrong claim from an outdated one.
• Nothing about rankings, competitors or recommendations. That is a different subject with different evidence, and mixing it in makes a factual report look like a commercial complaint.

What you attach depends on what you have established, not on what you suspect. Tracing what a claim can be connected to sets out which observations are demonstrable and which are inference - and a report that asserts a source as the origin of a claim, where the evidence only shows the source was displayed, is weaker than one that says what was actually observed.

Recording What You Filed

Keep your own incident record, because provider acknowledgements - where they are documented at all - do not amount to a record of the claim, the evidence, the submission and what you observed afterwards. Recommendation Google documents a confirmation email with a resolution update for verified knowledge-panel feedback - a non-AI surface. For the other ordinary factual-response routes reviewed here, the cited pages do not document an acknowledgement commitment. Several of these routes run through support tickets and email, which may well generate their own receipts; what is absent is a documented commitment, not necessarily a reply.

Keep the last row separate from the rest, and resist the sentence that joins them. A report filed on Tuesday and a changed answer seen on Friday are two dated observations. What a correction panel can establish sets out why that pairing supports a temporal association and not an attribution - and the same limit applies to a report as to an edit.

What to Expect, and Why That Is the Honest Answer

None of the documentation we read commits to an outcome or a timeframe for a report about a factual claim, and only one of these routes commits to an acknowledgement - Google's, for verified knowledge-panel feedback, on a surface that is not an AI answer. One states that a report will not necessarily result in removal. OpenAI describes a conditional, source-level mitigation affecting future responses - the only specific potential action tied to reported content in the set. The rest describe feedback as contributing to product or safety improvement, and stop there.

That is worth stating flatly to whoever asked you to "get it fixed." The report is a real action with a real record and no documented commitment behind it. Presented that way it is easy to justify; presented as a fix, it produces a follow-up question in three weeks that has no good answer.

What is actually within reach is the source layer. Correcting your own pages and the third-party sources that describe you is work you control, on evidence you can produce, for reasons that hold regardless of what any AI system does - and the order in which to do it is the practical next step. Whether any of it changes an answer is a separate question, observed rather than assumed, and how long that observation window should run is covered on its own page. Both sit inside the wider workflow this page is one step of.

Sources

Sources: Provider documentation, quoted as read on September 1, 2026, each statement applying to the named provider and surface alone. Google, Send feedback about AI Overviews (no date displayed) - the thumbs-down and Report a problem route, and that feedback helps improve AI Overviews for everyone. Google, Submit feedback on content about you (no date displayed) - the verified-user review turnaround and the policy on manual creation or deletion. Google, Search generative AI control (page notes an August 31, 2026 rollout-completion date) - what the control excludes, that it is not a ranking or inclusion signal, that it does not affect AI training, and the 1-2 day exclusion timing; that figure is exclusion latency for this control at this provider and is not used here as any other kind of timing. Google, Refresh Outdated Content tool (no date displayed) - scoped to pages the requester does not own, where content no longer exists or is significantly different; not a route for correcting your own pages, and not described as one here. OpenAI, Reporting content in ChatGPT and OpenAI platforms - the reporting route and the Model Quality team sentence, quoted in full; this page displays only a relative update date, so it is cited to the review date. OpenAI, ChatGPT search - that search results and citations can be incomplete, outdated or incorrect; relative update date only. OpenAI, Right to be forgotten and personal data removal from ChatGPT - case-by-case assessment, and that removal does not remove information from external websites or search engines; relative update date only. Perplexity, How can I report incorrect or inaccurate answers? (Last updated: July 16, 2026) - the flag, ticket and email routes, what a report should include, the named issue examples, and that a report is taken to improve the quality of responses. Anthropic, Claude is providing incorrect or misleading responses (Last Updated: March 16, 2026) - the thumbs-down and email routes and the single-source-of-truth statement. Anthropic, Privacy rights requests relating to Anthropic's training data (July 8, 2026) - scoped to personal data; the reduce-the-likelihood-of-reproduction sentence and the statement that models do not store text like a database, both read twice at the record. Neither this page nor OpenAI's describes editing model parameters, and this article makes no such claim. Microsoft, Privacy FAQ for Microsoft Copilot (08/31/2026) - the three feedback routes and the review statement. Microsoft, How to report a concern or contact Bing (10/21/2025) - the concern-reporting route and that reporting will not necessarily result in removal; this page does not mention Copilot or AI-generated answers. Not verified, and therefore not described. Bing Webmaster Tools' content-removal documentation could not be retrieved on the review date; that route is absent from this page rather than summarised from memory or from a third-party account. Scoped absence. Across the provider pages listed above, read in full on the review date, every documented route was per-response feedback, policy or illegal-content reporting, or privacy and legal removal scoped to individuals. We located no business-facing process for correcting a factual claim about a company. That is a result about the pages we read, not a claim that no provider ever acts on such a report. What is not claimed. Nothing here states that feedback retrains a model, that reporting corrects the reported response, that any provider's route changes a future answer, or that a documented figure at one provider describes any event at another. Equally, nothing here states that a report has no downstream effect: beyond OpenAI's conditional source-level mitigation statement, the documentation reviewed is silent on whether or how a report bears on later retrieval, later responses or model behaviour, and silence is recorded here as silence in both directions. No route reviewed here is described as acting on model parameters, because none of the cited documentation describes that.

---