ChatGPT says wrong things about my company. How do I correct it?

First, find where the wrong fact comes from

ChatGPT gets facts about a company from two places: the pages it retrieves when Search is on, and what the model learned in training. You can fix the first directly and the second only indirectly, so work out which one you’re facing.

Record the question, the exact wrong sentence, whether Search was on, which sources were shown, and the date. Ask logged out, in a private window. Then match the symptom:

We’ve seen this ourselves. On 15 Oct 2026, Brave’s AI answer described Katman as a “measurement service”, our positioning from early September. Its source was an old page of ours, dated 4 Sep 2026 in Brave’s index, which by then redirected to our About page (own check). The fact wasn’t invented; the index was stale.

Fix it at the source

1. Your own pages first. Search the repo and the live site for the old fact (grep -ri "old price" src/). Update it everywhere: page text, llms.txt, JSON-LD, Open Graph descriptions, PDFs. Old URLs that hold stale facts get a 301 to the current page or a 404.
2. Tell the search engines. Send changed URLs through IndexNow, request indexing in Search Console, and submit key pages to Brave. Bing says IndexNow helps AI answers reference the current version of a page (Bing Webmaster Blog, 10 Feb 2026, read 15 Oct 2026).
3. Third-party profiles. Update every listing you control. For articles, send the author a short correction with one link to your source page. Keep it factual and polite.
4. Name clashes. Use your distinctive name in full and add a “we’re not that company” line; the details are under cause 4 of Why doesn’t ChatGPT recommend my brand?.
5. Model memory. You can’t edit a model. OpenAI’s privacy policy warns that output may not be factually accurate and offers correction or removal requests for information about you as a person through privacy.openai.com (OpenAI privacy policy, updated 6 Feb 2026, read 15 Oct 2026). That route covers individuals; for company facts, the fix is current sources that Search finds and newer models train on.

Make the right fact unmistakable

Errors creep back when one fact lives in five places. Keep your company facts (name, what you do, prices, plan limits, location, what you’re not) in one file in your repo, and have pages, llms.txt and JSON-LD read from it. We learned this the hard way: our own llms.txt once showed a different price from the page it summarized.

An example Organization block, for an invented company, built from that file:

{
  "@context": "https://schema.org",
  "@type": "Organization",
  "name": "Acme Invoices",
  "alternateName": ["acmeinvoices.com"],
  "url": "https://acmeinvoices.com",
  "description": "Acme Invoices matches supplier invoices to purchase orders for small finance teams.",
  "disambiguatingDescription": "Not related to Acme Corp or Acme Payments.",
  "foundingDate": "2024",
  "sameAs": ["https://www.linkedin.com/company/acme-invoices"]
}

Every value in it must also be visible on the page. Then make the facts quotable: one sentence per fact, with the subject named, as described on How to get cited by ChatGPT. A fact stated once, clearly, on your own page, gives an answer with Search on something better to cite than an old directory entry.

Retest and keep a log

1. Ask the same question, logged out, with Search on, 14 days after the fix.
2. Save the answer, the sources and the date next to the original error.
3. Ask once more without Search to see what the model itself says.
4. Repeat monthly for the facts that matter most: price, what you do, who you are.

Answers vary from run to run, so one correct answer doesn’t close the case. The test protocol is on How to check if ChatGPT recommends you, and llms.txt is explained on What is llms.txt?.

Where Katman fits, and what it can’t do

Katman MCP runs inside your AI coding tool. katman_research keeps your one-sentence description, facts and not-X note in one file. katman_generate writes llms.txt and Organization JSON-LD from that file, so they can’t drift from your pages, and katman_lint_content flags a page whose description doesn’t match it. katman_audit_site reads the JSON-LD and llms.txt on your live site, and katman_manual_check logs the wrong answer you pasted so the retest has a baseline. Katman Pro is $99 a month for up to 3 sites, with fix prompts and page briefs uncapped. Or have us do it: Growth writes 30 pages a month for you. Plans are on the pricing page.

It can’t change what a model has memorized, remove third-party content or give legal advice, and it doesn’t edit your site by itself. Nobody can promise when an assistant will update its answer.

Written by the Katman team, Rnify Limited. OpenAI and Bing pages checked 15 October 2026; our own check as dated.