A prospect says ChatGPT told them you closed an office, charge a price you dropped two years ago, or are owned by a competitor. You tried to correct it in the chat and the next answer repeated the mistake.
That is normal, and it is fixable, but not where most people look. The chat window is the last place the error lives. The error lives in a page, a profile or a pattern of pages that ChatGPT read. This guide shows how to find it in five minutes, fix it at the source, and prove the fix worked. For the wider picture of getting named at all, start with the AI visibility hub.
01Why ChatGPT gets your company wrong
ChatGPT gets company facts wrong for three reasons: it is repeating an old or wrong page, it found too little to say and guessed, or it confused you with another company. MaxAEO, summarising OpenAI research, puts the guessing down to training that rewards an answer over an admission of not knowing. Which of the three applies decides the fix, so name it first.
One vendor, MaxAEO, published a count of 412 confirmed brand-fact errors across 63 B2B companies and found that 61% traced to retrieval (the live search step), 24% to training data and 15% to entity confusion. The vendor does not publish its method, so treat the split as a rough guide, not a benchmark. It still matches what we see: most wrong answers come from a specific page that can be found and changed.
| Cause | What it looks like | Where the fix lives |
|---|---|---|
| Retrieval error | The answer cites a page, and that page is old, wrong or written by someone else | The cited page, then every page that repeats the fact |
| Training-data error | No citation. The fact was true years ago, such as a former address or an old founder | Your own pages and profiles, then wait for the model to catch up |
| Entity confusion | The answer mixes you with a company that has a similar name or the same industry | Your canonical facts page, schema and profile wording so you are clearly a separate entity |
Cause shares and categories from MaxAEO, ChatGPT Wrong Company Information (vendor data, method not published)
02Run the five-minute diagnosis
Run the diagnosis before you edit anything. Edits made before you know the cause are the most common reason a fix does not hold.
Write down the wrong claim exactly
Copy the sentence ChatGPT produced and the prompt you used. One claim per row: price, location, owner, service, founder, size.
Ask the same question twice: search on, search off
With web search on, ChatGPT shows sources. With it off, it answers from memory. A wrong answer with search on points at a page. A wrong answer with search off points at older training data.
Repeat each version three to five times
Answers vary between runs. A fact that is wrong once and right four times is a different problem from a fact that is wrong every time.
Open every cited source
Find the sentence that holds the wrong fact. If the source is your own site, you found the bug. If it is a directory, a review site or an old article, that is where you start.
Ask a differently worded question
Try the question the way a buyer would phrase it. If the answer changes, the problem is thin coverage, not a bad source.
03Match the error to its source
Where a wrong fact comes from tells you who has to change it. Your own site is the fastest to fix. Third-party pages need a request. Memory-only errors need patience and a strong, consistent signal on the open web.
| Wrong fact | Check first | Then check |
|---|---|---|
| Old price or discontinued service | Your pricing page and any cached or archived copy | Comparison sites and review listings |
| Wrong city or closed office | Google Business Profile and your contact page | Directories and press that quote your old address |
| Wrong owner or founder | Your about page and LinkedIn company page | Wikidata, Crunchbase-style profiles and news |
| Competitor named as parent or partner | Third-party roundups that list both of you | Your own about page, which may never say who you are not |
| Wrong size or customer count | Old press releases and conference bios | Your homepage claims and job listings |
04Fix the source, not the chatbot
Fix in this order: the page ChatGPT cited, then your canonical facts page, then third-party profiles, then recrawling. Going in this order means each step can be checked against the answer before you spend time on the next.
Correct the cited page
If it is yours, change the sentence and keep the old fact out of nearby copy, since a stray mention in a sidebar or an old blog post can undo the fix.
Publish one canonical facts page
A plain page with who you are, what you sell, where you operate, who runs you, and the current prices or price range. Use the exact wording you want repeated.
Align every profile to one sentence
Use the same one-line description on your homepage, LinkedIn page, Google Business Profile, app listings and review profiles. Matching wording is how an AI recognises that these are all you.
Ask owners of third-party pages to change them
Send a short note with the old fact, the correct fact and a link to your canonical page. Directories and publishers usually respond to a specific request faster than to a general one.
Get the pages recrawled
Request indexing in Google Search Console and Bing Webmaster Tools, and make sure your sitemap lists the changed pages. Google says recrawling can take days to months, so ask early.
Use in-chat feedback as a supporting signal
Thumbs-down feedback does not rewrite the answer. MaxAEO describes it as a supporting signal only, and OpenAI has told a complainant it cannot correct the data, so do not rely on it.
05Check that AI crawlers can read the fix
A corrected page only helps if the right crawler can fetch it. OpenAI documents three bots with different jobs, and blocking the wrong one removes you from ChatGPT search without helping anything else.
| Bot | What it does | If you block it |
|---|---|---|
| OAI-SearchBot | Surfaces sites in ChatGPT search results | You stay out of ChatGPT search answers, apart from possible navigational links |
| GPTBot | Collects content that can be used to train OpenAI models | You signal that your content should not be used for training. Search is unaffected |
| ChatGPT-User | Visits a page when a person asks ChatGPT about it | OpenAI says robots.txt rules may not apply, because a person started the visit |
| PerplexityBot | Surfaces and links sites in Perplexity results | You can drop out of Perplexity results |
| Perplexity-User | Fetches a page when a person asks a question | Perplexity says it generally ignores robots.txt for these user-requested fetches |
OpenAI, Overview of OpenAI crawlers; Perplexity, Perplexity crawlers
Both companies say robots.txt changes take effect within about 24 hours. If your site sits behind a firewall or a bot-protection setting, check that it is not turning these bots away, since the file can say yes while the firewall says no. If you want a plain-text summary file as well, our llms.txt generator makes one, but Google says no special AI file is required, and we treat it as optional, cheap and unproven.
06A canonical facts page you can copy
A canonical facts page is one URL that states the facts about your company in short, self-contained sentences, so every other page and profile can point to it. It is also your entity home, the page that tells machines which company you are. Our guide to what an entity home is goes deeper.
What your canonical facts page should state
0 of 8 done07What you cannot fix, and what the law says
You cannot make ChatGPT forget a fact on request. The privacy group noyb filed a complaint in Europe after ChatGPT gave a wrong birth date for a public figure. According to noyb, OpenAI said it could not correct the data and could only filter or block information for certain prompts, such as the person's name, which would block everything about that person. OpenAI told noyb that factual accuracy in large language models remains an area of active research.
That case concerns personal data under the GDPR, the European data protection law, which gives people rights to accuracy, rectification and access. A company's pricing or office location is not personal data in the same way, so the legal route is narrower. For a company, the practical route remains the one above: change what the model reads, and re-test.
Fixing it at the source suits you if
- The wrong fact is one a buyer would act on, such as price, location or ownership
- Search-on answers cite a page you can identify
- You can change your own site and profiles this week
Expect less if
- The error only appears with search off and no source is cited, since memory-only errors change slowly
- The wrong fact sits on a page you do not control and the owner does not reply
- You are mistaken for a company with the same name and a far bigger web footprint, which needs entity work first
08How to know it is fixed
Treat a fix as done only when the correct fact appears in most runs of the same prompt over about a week, with search on and with it off. A single right answer proves nothing, because answers vary between runs. Keep a plain log so you can show what changed and when you changed it.
| Prompt | Mode | Runs | Correct | Cited source | Action taken |
|---|---|---|---|---|---|
| Where is [company] based? | Search on | 5 | 2 of 5 | Old directory | Asked directory to change |
| Where is [company] based? | Search off | 5 | 1 of 5 | Nothing yet | Facts page published |
| What does [company] charge? | Search on | 5 | 5 of 5 | Our pricing page | None needed |
MaxAEO reports median times to a corrected answer by engine, from 9 days for Perplexity to 33 days for Gemini, and says Perplexity fixes on high-authority sites took 3 to 4 days. Those are one vendor's figures and their method is unpublished, so use them only to set expectations: change quickly, then wait and re-test rather than rewriting the page every day.
Once your facts hold, keep watching. A monthly check of 15 to 25 prompts across ChatGPT, Perplexity, Gemini and Claude is enough for most companies, and our guide on how to measure AI visibility shows how to set it up. If ChatGPT never names you at all, you have a different problem, covered in brand not showing in ChatGPT.
09When the wrong answer keeps coming back
If the answer stays wrong after the cited page and your profiles agree, look for a copy you did not know existed: a scraped version, a cached comparison page, or a competitor's roundup. Search for the exact wrong sentence in quotes, then search for your own old wording. Repeat copies are why a fix looks like it worked and then reverses.
If the confusion is with a similarly named company, the fix is identity work: a clear entity home, matching profile wording, and links between your own properties. Our entity SEO page and the guide to the Google Knowledge Graph cover how machines decide that two names are one company. If you want this run for you, ChatGPT SEO services and the AI visibility audit start with exactly this diagnosis, at $2,000 a month for ongoing work.
What most companies do today
- Correct ChatGPT inside the chat window and assume it learned
- Edit the homepage only, leaving older pages and profiles with the old fact
- Ask once, see a correct answer, and stop checking
- Block every AI bot to be safe, which also removes them from search answers
- Wait for the model to catch up without changing any source
What we build
- A diagnosis that separates retrieval, training-data and entity errors
- One canonical facts page, with every profile repeating its first sentence
- A request log for third-party pages that hold the old fact
- Crawler checks for OAI-SearchBot, GPTBot, ChatGPT-User and PerplexityBot
- A repeated prompt panel across engines, with a log you can show your team

