An AI assistant said something false about my company. What can I do?

First, reproduce it properly

One screenshot is not evidence of a systematic problem. Assistants sample, personalize by location, and in some products carry memory of your earlier conversations. Someone logged into an account that has discussed your company for months is not seeing what a stranger sees.

Test logged out, in a fresh session, with location set explicitly, several times, across more than one assistant. If the error appears once in ten runs, that is very different from appearing every time, and it deserves a very different response.

Determine whether it is retrieval or invention

Ask the assistant to show its sources. A cited claim points at a page you can go fix or get corrected - a stale directory, an old press release, a competitor comparison article, a review site summary.

An uncited claim generated from model memory is the harder case. There is nothing to correct upstream. What you can do is give the retrieval layer a better answer to find, so future responses ground themselves in something real instead of guessing.

Competitor comparison pages deserve special attention here. A page written by someone selling against you, phrased as a neutral comparison, is a legitimate source to a retrieval system. You cannot get it removed, but you can be the clearer, more specific source on the same question.

Publish the correction as an answer, not a press release

A page headlined with the question people actually ask, answering it plainly in the first paragraph, is far more retrievable than a news post announcing that a rumor is untrue. State the correct fact directly, in the vocabulary of the question, and repeat it consistently in your profiles and structured data.

This is exactly the format an answer hub produces, which is one of the less obvious reasons to run one - it gives you a fast, credible place to put a correction that systems can actually read.

Consistency is what makes a correction stick. One page saying the right thing against nine sources saying the wrong thing loses. The correction has to appear on your site, in your profiles, in your structured data and, where possible, in a third-party source, so the consensus moves rather than just your corner of it. That is the same reconciliation discipline behind listings work.

Know the limits, including the legal ones

Providers offer feedback mechanisms, and some have processes for factual disputes, but none guarantee a specific output. There is no takedown that reliably removes a generated sentence, because the sentence is regenerated each time.

If output is defamatory, discloses something it should not, or creates a regulatory problem for your business, that is a legal question. Requirements and remedies vary by state and are still developing. Talk to counsel rather than to a marketing vendor, and keep dated, reproducible records of what you observed.