An AI assistant says something wrong about my business. How do I get it corrected?
You cannot edit the answer. You can change what the answer is built from. Work out whether the error came from a retrieved page, a stale third-party record, or the model's own memory, then correct the sources carrying it and publish an unambiguous statement of the correct fact where it can be found and quoted. Vendor feedback channels help for clear factual errors. None of this is instant and none of it is guaranteed.
First find where it came from
Ask the assistant for its sources, then actually open them. Three things happen. Sometimes a cited page really does say the wrong thing, which is the easy case. Sometimes the cited pages are fine and the error was introduced in summarizing them. Sometimes there are no citations at all, which points at training memory or at an aggregate impression formed from many weak sources.
Skipping this step is how companies end up rewriting a page that was never involved. It is also how a supplier's outdated product page, or a review site's decade-old category assignment, keeps producing the same wrong sentence for years while you edit your homepage.
Three origins, three remedies
The second case — accurate sources, inaccurate summary — is the one people handle worst, because there is nothing obviously broken to fix. It is almost always an ambiguity problem in your own writing, and it is addressed by rewriting the specific sentence that was misread rather than by touching listings or third-party records at all.
- A retrieved page is wrong. Correct the page if it is yours. If it is someone else's, request a correction; review platforms and directories usually have a process, and trade publications generally will fix a factual error.
- A stale record is wrong. Fix the origin listing, not just the copies, and make your own properties agree exactly.
- The model's memory is wrong. You cannot patch it. You can make the correct fact abundant and well-corroborated so that live retrieval overrides memory, and use the vendor's own feedback mechanism where one exists.
Write the correction so a machine can use it
A vague update does not travel. 'We've evolved our offerings' corrects nothing. 'As of 2026, the company no longer performs residential electrical work and does not hold a residential electrical license' is a complete, quotable, unambiguous statement — subject named, scope stated, dated.
Put it somewhere permanent rather than in a news post that scrolls away, and make sure it is in the served HTML rather than loaded by script after the page renders. A correction a fetcher never sees is not a correction.
What is genuinely outside your control
There is no service level agreement on AI answers. No vendor owes you a removal, a timeline, or an explanation. Caches persist, indexes lag, and models retrain on their own schedule. Anyone promising to make a specific assistant say a specific thing by a specific date is describing something they cannot deliver.
If the statement is seriously damaging rather than merely wrong — an allegation about safety, licensing or conduct — that moves beyond a marketing problem. Obligations and remedies differ by state and by circumstance, so involve your attorney early rather than treating it as a content task. Everything above still applies in parallel, because correcting the underlying sources is useful regardless of what else you do.
Topics: corrections · reputation · entities · accuracy
Have a version of this question about your own business?
The useful answer usually depends on which systems you run and how they're connected. That's a conversation, not a blog post.