ChatGPT Is Saying Something False About My Company. What Can I Actually Do?

There is no takedown form. No AI engine offers a mechanism for a business to have a specific false claim about it removed from generated answers, and the first defamation case to reach judgment went in the engine’s favor. What does work is changing the sources the engine reads, because a generated answer is a summary of retrieved material rather than a stored record you can edit.

The call usually arrives the same way. Somebody on the sales team asked ChatGPT about their own company out of curiosity, and the answer contained something that is not true. A regulatory matter that closed years ago described as ongoing. A product line that was discontinued. A founder’s name attached to a company they never worked for. Occasionally something worse.

The first instinct is to find the delete button. This article explains why that button does not exist, what the courts have said so far, and what actually changes an answer.

Why you cannot delete it

A search result is a pointer to a document that exists somewhere. Remove or outrank the document and the result changes. An AI answer is not a pointer. It is generated at the moment you ask, from two ingredients: what the model absorbed during training, and what it retrieved from the web during that conversation.

Neither ingredient has a per-company edit interface. Training data is baked into model weights and cannot be surgically corrected on request. Retrieval reaches out to live pages, so it changes when those pages change, which is the opening.

Every major engine offers a way to report a bad response. That feedback goes into general model improvement rather than into a case file about you, and nobody will write back to confirm your specific answer was fixed.

Where the law currently stands

Mark Walters, a radio host in Georgia, sued OpenAI after ChatGPT produced a legal summary stating he had been accused of embezzling funds from a nonprofit. He had not. The case was the first defamation claim over AI-generated output to reach a substantive ruling in the United States.

In May 2025 the court granted summary judgment to OpenAI. The reasoning mattered more than the outcome. The court held that a reasonable reader in that context would not have understood the output as a statement of fact, given the disclaimers presented and the circumstances of the exchange, and that the plaintiff had not established the fault standard required.

Other claims are moving through courts in the United States and elsewhere, and the position could change. As of now, planning your response around a lawsuit means planning around a route that has been tested once and did not open. Nothing here is legal advice, and if the content is genuinely defamatory you should speak to a lawyer in your jurisdiction as well as fixing the supply side.

The one question that decides your fix

Before doing anything, find out whether the false claim came from retrieval or from training memory. The two need opposite work, and doing the wrong one wastes a quarter.

Run this test. Ask the same question twice in ChatGPT, once in a session where it searches the web and once where it answers without searching. Then ask the same question in Perplexity, which retrieves for nearly every query and numbers its citations.

• Wrong only when it searches, or the citations point at a specific page. This is retrieval. Something published is producing the claim. Find that page and the fix is concrete.
• Wrong even without searching, and Perplexity gets it right. This is training memory. The claim came from material absorbed before the cutoff. Nothing you publish today removes it, but publishing a clearer current source gives retrieval something better to reach for, which is what the model uses when it does search.
• Wrong in both, on every engine, with no source anywhere. This is invention filling a gap. The engine had nothing solid and produced something plausible. The fix is to close the gap so there is a clear answer to find.

Free ChatGPT sessions currently run on a model whose training knowledge ends in October 2023, so a fair amount of what looks like a serious error is simply a model describing a company as it was.

What to do this week

Record the evidence before it changes

Answers are not stable. Ask again tomorrow and the wording moves. Before you do anything else, capture the full text of the answer, the engine, the exact prompt, the date, and whether search was on. If you skip this you will have no way to prove later that anything improved, because there is no history on the engine side. Our clients get this stored automatically in the client portal, and you can build the same thing in a spreadsheet.

Correct and date your own pages

This is the highest-yield step and the one most often skipped. Yext analyzed 6.8 million AI citations and found roughly 86 percent traced back to first-party pages and business listings. If a claim about you is stale, the correction has to exist somewhere unambiguous and current.

Write the correct fact plainly, on a page an engine can reach, with a visible date. Do not bury it in a PDF, a press release archive, or an image. If the matter is genuinely closed, say it is closed and say when.

Fix the machine-readable version of you

Engines resolve entities before they answer. If your structured data, your Wikidata record and your listings disagree about your name, your location or what you do, you are easier to confuse with somebody else. Merged identity is one of the most common causes of a genuinely false statement, and it is not a content problem.

Repair the third-party record

Review platforms carry weight out of proportion to their traffic. Profiles on Trustpilot, G2 and Capterra are roughly three times more likely to be cited than an equivalent page elsewhere. A thin or abandoned profile is a gap an engine fills from somewhere less friendly. Community threads matter too, and we cover how those get picked up in our note on Reddit and community citations.

Re-run the same question on a schedule

Monthly, same prompts, same engines, stored beside last month’s text. Perplexity tends to move first because retrieval is always in play. Anything sitting in training memory waits for a model release, and that is a fact about the calendar rather than a reflection of your effort.

What will not work

Arguing with the chatbot changes that conversation and nothing else. Publishing a rebuttal page that repeats the false claim in its heading hands the engine a fresh source that contains the claim. Buying reviews is visible to the platforms and to anyone reading them. And if the underlying complaint is accurate, no amount of source work outruns it for long, which is a harder conversation than most agencies in this category are willing to have.

Frequently asked questions

Can I contact OpenAI and ask them to remove it?

You can report the response through the interface. There is no route that reviews a specific claim about a named business and confirms a correction. Treat feedback as useful and slow rather than as a remedy.

How long until the answer changes?

If the cause is retrieval, weeks is realistic once the better source is published and crawled, and Perplexity often shows it first. If the cause is training memory, the honest answer is that it waits for a model release and nobody outside the lab knows the date.

Does this only affect ChatGPT?

No. Six engines answer questions about you and they disagree with each other constantly. An answer that is wrong on one engine and right on another is a retrieval problem. Wrong on all six is an entity problem. We track all six on our AI reputation management engagements.

Is this the same as SERM?

It shares the goal and almost none of the mechanics. Search reputation works on a page of results you can influence by ranking. There is no ranking here. Our search reputation management page covers the older discipline.

Where to start

Write down the exact question your buyer would ask before contacting you. Ask it in ChatGPT and Perplexity, in the language your market uses. Save both answers with today’s date. That single file is the beginning of every engagement we run, and you can produce it yourself in an afternoon.

If the answers are bad enough that you want them traced and fixed, that is what our AI reputation management service does, across six engines with monthly re-measurement.

Related guides

• GEO: Generative Engine Optimization – Your Guide to Ranking on ChatGPT
• AI Search Optimization in Egypt: Ranking on ChatGPT, Gemini, Perplexity & Copilot
• GEO vs SEO for Egyptian Businesses: What Actually Changes in 2026
• ChatGPT and Perplexity Optimization for Egyptian Brands: A Senior Playbook