Why does ChatGPT say wrong things about my company? How I fix it

Why does ChatGPT get facts about a company wrong?

Three causes, in the order I see them:

1. Thin sources. If the web has little about your company, the model has little to go on. It fills the gap with what a company like yours usually says.
2. Stale sources. An old directory listing, an old pricing page or a press article from three years ago still reads as true.
3. Sources that disagree. Your site says one thing, your LinkedIn page another, and a listing a third. The model picks one, or blends them.

When search is on, the tool reads live pages and often shows its sources. When it is off, it answers from training data, which can be older than your last change.

What does the data actually show?

Seer Interactive published a study on 27 August 2026. They ran 1,562 branded prompts across six platforms (Google AI Mode, AI Overviews, ChatGPT, Claude, Gemini and Perplexity) between 1 May and 9 June 2026, and scored 28,123 responses.

Read that carefully. Wrong answers are real, but when a tool answers it is mostly right. The larger gap is that about a third of the time it says nothing, and on comparisons it says nothing four times out of five. For a small B2B brand, being absent when a buyer compares options is the more common problem.

That does not make wrong answers harmless. One invented guarantee in front of the wrong buyer is enough. It means your fix has two jobs: stop the wrong, and fill the silence.

How do I find out what AI says about my company?

Step 1: Ask ten questions and write the answers down

Use the tools your buyers use, logged out of your own account if you can, with search on. Ten to start:

1. What does {Company} do?
2. Who are {Company}'s customers?
3. Where is {Company} based and who founded it?
4. How much does {Company} cost? (Tests whether it invents prices.)
5. Does {Company} offer {a service you do not offer}?
6. What is {Company} known for?
7. What are alternatives to {Company}?
8. Is {Company} better than {competitor}?
9. Who should hire {Company}?
10. What do customers say about {Company}?

Log each one: date, tool, the answer, and a score of accurate, outdated, wrong or not mentioned. A sheet with a column per tool is enough. The full version of this is a 30-prompt panel, which I will publish as a template.

Step 2: Find where the wrong answer came from

With search on, look at the sources the tool shows. Wrong facts often trace to a single old page. If there are no sources, search for the wrong sentence yourself.

Step 3: Match the problem to the fix

Step 4: Publish one clear page of facts

One page on your own site (About or FAQ) that states, in plain sentences in the first lines, what you do, for whom, where, since when, and what you do not do. Short sentences, dates visible, one claim per line. This is the page the tool is most likely to quote.

Step 5: Make the profiles agree

Use the same name, the same one-line description and the same facts on your site, LinkedIn, Google Business Profile and the directories you appear in. When sources agree, the model stops blending them.

Step 6: Test again

Rerun the same ten questions after four to eight weeks, then monthly. Jasper suggests rechecking affected prompts within 60 days of a correction, and monthly in fast-moving categories. If a fact changes, update the source first and the test second.

What does not work

• Asking the AI to remember the correction. A correction in your own chat fixes your chat, not what a buyer sees.
• Relying on llms.txt. Server logs of about 137,000 domains showed that 97% of llms.txt files get no AI requests (Ahrefs, May 2026). I keep one on my own site because it costs nothing, and I would not sell it to a client as a fix.
• Counting on the model to be fair. It answers from what exists. Mentions on other sites are the strongest signal I know of: in Ahrefs' study of 75,000 brands, brand mentions on other sites correlated most strongly with appearing in AI Overviews.

Mistakes I see

1. Testing once. One run proves nothing. Answers change. Test monthly with the same prompts.
2. Fixing the symptom. Correcting the answer in a chat instead of the page it came from.
3. A website that never says what it does not offer. The model fills that silence with a guess.
4. Only checking ChatGPT. Buyers use several tools. Check at least ChatGPT, Gemini, Perplexity and Google AI Mode.

FAQ

Why does ChatGPT make things up? Because it predicts likely text. When it lacks reliable information about a company, the likely text is a confident guess.

Can I ask OpenAI to correct what ChatGPT says about my company? I do not know a dependable way to get a specific answer corrected. What works is changing the sources it reads.

Does llms.txt fix it? No evidence that it does. See above.

How often should I test? Monthly, with the same prompts, and again within 60 days of any correction.

Is it the same in Claude and Perplexity? Each tool has different sources and behaviour. In the Seer data, Claude answered fewer prompts than the others, and when it did answer it was accurate 98% of the time. Test each one.

One next step

Run the free AI-Ready Check: a score, three fixes, and a look at what ChatGPT says about your company.

By Vedran Obradović, Founder & Branding Specialist at Vizura Studio® in Zagreb. 250+ brand projects over 8+ years, most for B2B tech companies. Updated 21 October 2026.

Sources

• Seer Interactive, AI brand accuracy study (27 August 2026): https://www.seerinteractive.com/insights/ai-brand-accuracy-study
• Jasper, How to monitor and prevent AI hallucination: https://www.jasper.ai/blog/how-to-monitor-and-prevent-ai-hallucination
• Ahrefs, 75,000 brands and AI Overview mentions: https://ahrefs.com/blog/ai-overview-brand-correlation/
• llms.txt adoption vs usage (Ahrefs logs via PPC Land): https://ppc.land/llms-txt-adoption-rises-8-8x-but-97-of-files-get-zero-ai-requests/