What to do when ChatGPT or Gemini gets your brand wrong
By Ben Frampton · · 5 min read
Short answer: when an assistant gets your brand wrong, it's repeating something it learned in training or something it read on a page it retrieved. Find out which: check whether the assistant cites sources for the wrong claim, and search the web for the same wording. Then fix it at the source: update your own pages so the correct facts are clear and consistent, ask third-party sites to correct outdated information, and use the assistant's feedback tools. Expect retrieved errors to clear in weeks and training-data errors to linger until the next model.
Common ways assistants get brands wrong
- Out-of-date facts: old prices, retired plans, a previous name, former leadership.
- Discontinued products described as current.
- Wrong audience or category: "an enterprise tool" when you serve small businesses, or the reverse.
- Confusion with another company with a similar name.
- Missing features, or features that belong to a competitor.
- Invented details (often called hallucinations): a plausible-sounding fact with no source at all.
Why does it happen?
There are three causes, and each needs a different fix:
| Cause | How to recognise it | Main fix |
|---|---|---|
| Old training data | Wrong without citing a source; same error across different phrasings | Make correct facts widespread and wait for newer models |
| A wrong or outdated page | The answer cites a page, or the wording matches a page you can find | Correct the page, or ask its owner to |
| Ambiguity | Facts are mixed with another brand, or your own pages disagree with each other | Make your identity and facts unambiguous everywhere |
Invented details usually come from gaps: when there's little clear information, the model fills in something plausible. More clear, consistent information reduces them.
Step 1: find the error systematically
Don't rely on one screenshot. Ask the assistants several branded questions ("What is [brand]?", "How much does [brand] cost?", "Who is [brand] for?", "[brand] vs [competitor]"), several times each, and note:
- which claims are wrong,
- how often each error appears (one answer in twenty, or nearly every time),
- which assistants make it,
- which sources, if any, are cited.
An error that appears in one answer out of twenty is less urgent than one that appears in most. Tracking branded questions over time shows whether your fixes are working. See how to choose which prompts to track.
Step 2: fix your own pages first
Assistants and search engines should be able to find the correct facts on your site, stated plainly:
- An "about" section in plain words: what you do, who it's for, where you operate, when you were founded.
- Current pricing in text (not only in images or behind a calculator).
- A clear note on changes: "Formerly called X", "The Starter plan was replaced by Basic in March 2026". This helps models link old and new information correctly.
- No contradictions: old landing pages, PDFs and help articles with outdated facts should be updated, redirected or removed.
- Consistent organisation details in your structured data and social profiles.
Step 3: correct third-party sources
Search for the incorrect claim in quotation marks, and look at the sources the assistants cite. Then:
- Update the profiles you control: review platforms, directories, marketplaces, social profiles, Google Business Profile.
- Ask publishers to correct articles, politely, with evidence, and a short replacement sentence they can paste in.
- Update Wikidata or Wikipedia only following their rules on conflicts of interest: usually by suggesting changes on the talk page with reliable sources, not by editing yourself.
The sources assistants cite most for your category deserve attention first. See the sources AI assistants cite.
Step 4: report it to the assistant
Most assistants have a thumbs-down or "report" option on each answer. It won't fix things immediately, but it's free and feeds into providers' quality work. If an error is harmful (for example, a false claim about safety or legal trouble), use the provider's formal feedback or legal contact routes.
How long until the answer changes?
- Errors from retrieved pages can clear within days or weeks, once the corrected page is recrawled and ranks.
- Errors from training data last until a newer model replaces the current one. Assistants that search the web may override old training data sooner if they find clear, current pages.
Measure it rather than spot-checking: count how often the wrong claim appears across many answers before and after your fixes.
Frequently asked questions
Can I contact OpenAI or Google to correct what their AI says about my company?
You can use their feedback tools and, for serious issues, their legal or privacy contact routes. There's no general service to edit what a model says about a brand. Fixing the sources is more reliable.
Why does the assistant still show our old price after we changed it?
Either the model learned the old price in training, or a page it retrieves (yours or someone else's) still shows it. Search for the old price online and fix every copy you find.
Should I publish a page that lists common misconceptions about our brand?
A short, factual "facts about [brand]" or FAQ section can help, if it's written for people. Avoid pages addressed to AI systems; they don't have a proven effect. See do llms.txt and schema markup help?
Is a negative but accurate description an error?
No. If assistants report real criticism, for example that you're more expensive than alternatives, the fix is to address the issue or explain the trade-off clearly on your pages, not to try to hide it.