ChatGPT says something wrong about your company. Where did it come from, and how do you fix it?

A prospect forwards you a ChatGPT answer. It quotes a price you retired last year, lists a feature you never built, or puts your company in the wrong country. Your first instinct is to complain to the engine. Our view: that is the third thing to do, not the first. Most wrong answers repeat something the engine found, and the page it found is usually one you can change or get changed.

Start with the source, not the engine

The fix starts with evidence, not with a feedback button. An AI answer is assembled from what the engine retrieves at the moment of the question and what the model absorbed in training. If the wrong fact sits on a page the engine still reads, a report changes little: the next retrieval finds the same page and repeats the same claim.

So we work in four moves. Capture the error so you can prove it and test for it later. Trace it to the page or profile it repeats. Fix the fact there, and state the correct version plainly on your own site. Then report it through the engine's own route and recheck on a schedule, keeping a log.

The order matters because each step feeds the next. A good capture tells you which engine and mode to trace. A traced source tells you who has to make the change. A fixed source gives the report something to point at.

The correction playbook, in order

1. Capture

   Save the prompt, engine, mode, country, language, date, the full text and a screenshot.

2. Trace

   Open every cited link and search for the exact wrong phrase on the open web.

3. Classify

   Name the error type: stale fact, wrong entity, invented detail or a rival's claim.

4. Fix at source

   Change the page, profile or listing the engine reads, and state the fact on your own site.

5. Report

   Use the engine's documented feedback route and link the corrected page.

6. Recheck

   Run the same prompts on a fixed schedule and log each result with its date.

Step 1: capture the error properly

A capture is a record you could hand to a colleague who has never seen the answer. Write down the exact prompt and the engine and mode: ChatGPT with search on or off, Google AI Overviews or AI Mode. Add the country, language and date, and whether you were signed in. Save the full answer text and a screenshot, with any cited links.

Then run the same prompt a few more times, and from a second country if you sell in more than one. AI answers vary between runs, so one bad answer may be a rare draw rather than the engine's settled view. Three runs that repeat the same wrong price tell you much more than one.

Label captures the way we label ours: "Captured {date}, {engine}, {country}", with the raw capture stored. Anything you rewrite or simplify for a deck is "Illustrative". It sounds fussy. It is also the only way to show later that the answer changed after your fix, and not because you asked a different question.

Step 2: trace it to a source

Most wrong answers have an address. Start with the links the engine cites, then search the exact wrong phrase in quotes. Then check the places engines lean on for company facts: your own site, business directories, review platforms, marketplaces, partner pages and old press releases.

Your own old pages

Look at your own site first, because engines read it heavily. In Profound's analysis of 11.84 billion citations across eight AI engines, collected from 16 April to 16 July 2026, about 57% of citations went to brand websites.¹ Profound sells AI visibility software, so read it as a vendor study, but the direction is plain: your pages are a large share of what engines quote.

That cuts both ways. An archived pricing page can still be crawlable. So can a PDF datasheet from two versions ago, a help article about a retired plan or a landing page for a discontinued product. Search your own domain for the wrong figure or name. Check subdomains, documentation sites and the files in your sitemap, not only the pages in your navigation.

Directories, profiles and reviews

Next, check where other sites describe you. Business directories, software marketplaces, review platforms, partner listings and data aggregators copy each other, so one stale entry can spread to a dozen pages. Profound's study of inaccurate AI claims, published in August 2026, reported that pricing claims were wrong more often than other claim types.² The same study found brands' own content among the sources cited in inaccurate claims.

Price is the classic case. You change it on your site, but a marketplace listing, a comparison article and a reseller page keep the old one for a year. A rival's comparison page is another common source: it may describe your product as it was when they wrote it, or simply get it wrong. For retailers, our playbook for e-commerce brands runs the same check across feeds, marketplaces and reviews.

Model memory with no citation

Sometimes no cited page and no search result contains the error. Then it probably comes from what the model learned in training, or from blending two companies with similar names. You cannot edit training data. You can make the correct fact consistent and easy to find across the sources engines trust, and recheck after the engine ships a model update.

Error sources by fix route

Effort to fix →

Model memory or a rival's pageCorrection requests with evidence; consistent facts everywhere.

Retired product on many old pagesMark each page retired and link the replacement.

Directory or marketplace listingClaim the profile, update it, then report.

Old price on your own siteRedirect or update it this week. Start here.

Your control over the source →

Step 3: fix it where the engine reads

Fix the fact where the engine found it, then make the correct version the easiest thing to find. On your own site, that means one clear page per fact that buyers ask about: pricing, locations served, product names, leadership, what you do not offer. Put the answer in the first sentence of the relevant section, in plain text rather than an image, with the date it was last updated.

Retire old pages properly. Redirect a dead pricing page to the live one rather than leaving it orphaned. Remove or update outdated PDFs. If a page has to stay for history, say at the top that it is out of date and link the current version. Markup alone will not carry a correction, and our review of the evidence on schema and llms.txt explains why.

Off your site, the owner of each listing has to change it. Claim your directory and marketplace profiles, update them, and ask publishers to correct factual errors with a link to your source page. Our view: a polite, specific correction request with evidence works better with editors than any other approach, and it leaves a record.

Illustrative: our working classification, not a sourced dataset.

Step 4: report it, recheck it and keep a log

Report the error only after the source is fixed, so the report can point at something true. Then measure whether the answer changes, on a schedule, with a written log. Feedback alone is a request, not a correction.

Each engine's feedback route

Every major engine documents a way to flag a bad answer. We read each one on its own help page and list the routes below, as written there. Where the help page describes a written field, use it: name the wrong claim, give the correct fact in one sentence and link the page that states it.

Help pages checked 6 October 2026. Routes change; open the linked page before you report.

What none of them promise

None of these help pages describes a brand-correction form, and none states how long a review takes (checked 6 October 2026). OpenAI's page comes closest to describing what happens next. It says reported domains and other content "may be reviewed by OpenAI's Model Quality team, which may apply filters or other mitigations to help prevent ChatGPT from relying on unreliable sources in future responses."³ Note the two uses of "may".

One vendor, MaxAEO, has published median correction times from its own tracking of business-to-business companies; no engine publishes one, and we have not verified the vendor's figures.¹⁰ MaxAEO also published an error taxonomy, and its piece is worth reading. Other guides circulate correction ranges with no stated source, and we do not repeat them.

Our view: anyone who promises you a correction date is guessing. What you can promise your own team is a schedule for checking and an honest log of what changed.

A recheck schedule you can keep

Pick the prompts that surfaced the error, plus two or three close variants. Run them on the same engines, modes and countries each time. Record the date, the answer, whether the wrong fact appeared and which sources were cited. If the cited source changes before the answer does, that is progress worth logging.

Keep the schedule modest enough to survive a busy month. If you would rather not run the loop yourself, it is part of our AI Visibility work. There, corrections sit in one register with everything else engines say about you.

A recheck log for one error

1. Day 0

   Capture, trace, fix the source and report the answer.

2. Week 1

   Rerun the prompt set; note whether the cited source has changed.

3. Weeks 2 to 4

   Rerun weekly; log answer, sources and date for each engine.

4. Monthly after

   Keep the prompts in the monthly set; recheck after each model update.

When the error is about a person

When the wrong fact is about a named person, a different set of rights applies, and it belongs to that person rather than to the company. This matters for founders, partners and named staff whose biographies engines repeat.

What GDPR rectification covers

Article 16 of the GDPR gives a data subject "the right to obtain from the controller without undue delay the rectification of inaccurate personal data concerning him or her."¹¹ A data subject is a living individual. So a founder can use it for a wrong claim about their own career; a company cannot use it for a wrong price. OpenAI's help pages say individuals in certain jurisdictions can make data subject rights requests through its Privacy Portal.¹²

Which route fits which error

Facts about the company

• Prices, products, locations, features
• Fix at the source, then feedback routes
• No personal-data right applies

Facts about a named person

• Career, role, conduct, biography
• Fix at the source as well
• The person may use their data rights

Two complaints still pending

Whether rectification works against a language model is still being tested. The privacy group noyb filed a complaint against OpenAI with the Austrian authority on 29 April 2024. OpenAI had said it could not correct a person's data, and noyb lists that case as pending.¹³ It filed a second complaint with Norway's Datatilsynet on 20 March 2025 over a fabricated criminal history. noyb lists that case as pending too, with the Irish Data Protection Commission as lead authority.¹⁴ This is not legal advice; a person in this position should speak to a lawyer.

What buyers do with a wrong answer

Many buyers check what AI tells them, which gives a wrong answer two costs: the doubt it plants and the time a buyer spends checking it. In TrustRadius's survey of 1,862 technology buyers in January 2026, 94% of buyers who used AI to research a purchase said they still fact-checked what it told them.¹⁵ Where do they check? Often on your site, a review platform or with your sales team.

Buyers are also wary of the channel itself. Gartner surveyed 645 B2B buyers in August and September 2025, so the data is now more than a year old and the geography was not stated. In that survey, 51% said they were more likely to meet misleading information from generative AI, against 49% from a sales rep.¹⁶

Our view: the cost of a wrong answer is rarely a lost deal you can point to. It is a sales call that starts with "I read that you…", and a buyer who now trusts your pricing page a little less than before. Those buyers often come back later by typing your name, which is why proving AI influence on pipeline takes more than a referral report.

How often do AI assistants get facts wrong?

The best large public test is about news, not brands, and it is dated. The EBU and the BBC collected answers from four free AI assistants in late May and early June 2025. Journalists from 22 public service media organisations, in 18 countries and 14 languages, reviewed 2,709 responses. In that study, 45% of responses had at least one significant issue.¹⁷

That figure is more than a year old and measures news questions, so it says nothing direct about how often engines get a company's prices or products wrong. We have not seen a public, independent study of brand-fact errors with a published method. Vendors report their own rates, and we leave them out until the method is open. What the news study does show is that errors were common enough across several assistants to be worth checking for, rather than assuming they are rare.

What not to do

Do not try to drown out a wrong answer with volume. Fake reviews, sock-puppet forum posts and paid placements dressed as independent opinion can get a company dropped from an engine's trusted sources. They also turn a fact problem into a trust problem.

When the problem is absence, not error

Sometimes the complaint is not that the engine is wrong about you, but that it does not mention you at all. That is a visibility problem rather than a correction problem, and it needs a different method: a fixed prompt set, repeated runs and a share-of-answer figure with an honest margin. Our guide on how many prompts and runs an AI visibility score needs covers that method in detail. For firms where reputation is the product, our professional services playbook shows how the two problems overlap.

The short version

A wrong AI answer is usually a symptom of a page somewhere. Capture it so you can test it, trace it to that page, fix the fact there and on your own site, then report it and recheck on a schedule. Treat the feedback button as the last step, not the first, and keep the log, because the log is the only evidence that your fix worked.