ChatGPT, Gemini or Perplexity has wrong information about your company: how to fix it

When a buyer asks an assistant about your company, the answer comes from one of two places, and often a mix of both. Search-backed answers retrieve pages at the time of the question and usually cite them: ChatGPT search, Perplexity, Google AI Overviews and AI Mode all work this way. Answers given without a search come from what the model learned in training, which can be months or years older than your current positioning.

The difference decides how you fix the problem. A wrong fact in a search-backed answer usually has a visible source you can correct, and the correction can show up once the page is recrawled. A wrong fact from a model's built-in knowledge changes only when the model is retrained or when the question triggers a search that finds better sources. That is why the work below focuses on sources, not on the assistants.

Typical B2B errors are ordinary: a retired product line, a former CEO, a pending acquisition that closed long ago, an old pricing model. Each one costs credibility with buyers checking you before the first call, as we describe in our article on how B2B buyers research vendors in ChatGPT.

Start with evidence. One screenshot will not tell you how often an error appears, where it came from or whether your fix worked. A small, repeatable audit will.

1. Write 20 to 30 prompts about your company itself: what you do, who you serve, where you are based, who leads the company, pricing model, integrations, security certifications, recent acquisitions and main competitors.
2. Add a few comparison prompts that name you next to two competitors.
3. Run each prompt in ChatGPT, Gemini, Perplexity, Microsoft Copilot and Google AI Overviews or AI Mode, in a logged-out or clean session where possible, and note the date.
4. Record the full answer, mark each factual claim as correct, outdated, wrong or missing, and copy every cited source URL.
5. Group the errors by fact, not by assistant: the same wrong founding year in three tools is one problem with one source.
6. Rank errors by commercial damage: what you sell, to whom, pricing, security and ownership go first.

Keep the prompt set fixed. Our guide to measuring share of answer covers visibility scoring; here, accuracy is the score.

For search-backed answers, the cited sources are your first lead. Open each one and look for the wrong fact in the text. When the assistant cites nothing, search the exact wrong phrase on Google and in Perplexity, and check the usual suspects below. The source is usually a page the company forgot it had, or a profile nobody has updated in years.

Your site is the one source you fully control, and third-party editors will check it before accepting any correction.

Update the About page so the basics are stated plainly in text, not only in images or a video: legal name, common name, what you sell, who you serve, headquarters and offices, founding year, leadership and ownership. Then publish a short company facts page that lists the same facts in one place with a last-updated date.

Add Organization structured data to the home page or to that one page describing the organization, which is where Google recommends it. Google lists properties such as sameAs for your official profiles, legalName, foundingDate and address, and notes that some properties can influence what appears in its results, including the knowledge panel. Google also advises that structured data should match the visible text on the page, so the markup and the facts page must say the same thing.

Then clean up what contradicts it: old service pages, retired product pages, outdated PDFs and blog posts that state old facts. Redirect, update or add a dated editor's note. Finally, consider an llms.txt file. It is a proposal, first published by Jeremy Howard in September 2024, for a Markdown file that gives language models a curated overview of a site. Google says you do not need AI text files or special markup to appear in its AI features, so treat llms.txt as a low-cost extra that points to your facts page, not as the fix.

Work through the sources your audit found, most cited first: directories, data providers, review sites, partner and marketplace listings, association pages, speaker bios and your own social profiles. Use the same wording as your facts page everywhere.

For old press and comparison listicles, send the publisher or author a short request naming the wrong fact and linking the page that proves the correct one. Many will add an update line.

Wikipedia and Wikidata have rules for people connected to the subject. Wikipedia's conflict of interest guideline strongly discourages employees and paid editors from editing their company's article directly; disclose the connection and propose changes on the talk page with an edit request, which volunteers may accept or decline. On Wikidata, the proposed verifiability policy says statements must be supported by a source, so attach a reference to an authoritative page, such as a closing announcement or registry record.

If Google shows a knowledge panel for your company, an authorized person can claim it by verifying through an official site or profile Google lists, then suggest changes through the panel's feedback option with links that support each change. Google notes that not every knowledge panel can be claimed.

Feedback flags an error to the vendor but does not edit a model or guarantee a change. We use it for high-damage errors after the sources are corrected, and log what was sent and when. The channels below are the ones we could verify in each vendor's own documentation on November 9, 2026.

• Google AI Overviews: click the thumbs down icon under the overview, choose Report a problem, pick a category, add details and submit. The three-dot menu at the top of each overview also offers feedback.
• Gemini: click Bad response under the answer, optionally select a reason and add written feedback, then submit. You need to be signed in.
• Perplexity: use the flag icon under the answer, or open a support ticket from account settings, and include the query URL, the error and the expected result. Misinformation and outdated information are listed issue types.
• OpenAI: the model behavior feedback form accepts examples where a model did not meet expectations, with either the ideal output or what an ideal answer would include. Submissions may be used for training and evaluation.
• Google knowledge panel: if you have verified, use Feedback below the panel and flag the specific fact, with supporting links.

For every report, include the source that proves the correct fact, ideally your facts page.

Some companies blocked AI crawlers when they first appeared and never revisited the decision. When that happens, search-backed assistants cannot read the corrected facts page, so they keep citing the stale directory instead. Vendors separate crawling for search from crawling for training.

Our default recommendation for B2B companies is to allow OAI-SearchBot, PerplexityBot and Googlebot, decide on GPTBot and Google-Extended as a separate content policy question, and check the firewall or CDN as well as robots.txt, because bot protection often blocks crawlers that robots.txt allows. Google warns that recrawling after a change can take from several days to several months, so request a recrawl of the facts page in Search Console once it is live.

OpenAI also adds utm_source=chatgpt.com to referral links from ChatGPT search, which lets you see in analytics when buyers land on your corrected pages from an answer.

Run the same prompt set every month and track two things: how many wrong facts still appear, and whether cited sources have shifted to your facts page and corrected profiles. Expect search-backed answers to change first; answers without a search can repeat old facts much longer. When an error persists, go back to Step 2: there is usually a second source you have not found yet.

Some things are outside your control:

• When a model is retrained, and what it remembers from older training data.
• Whether a given prompt triggers a search or is answered from memory.
• Accurate but unwelcome facts in well-sourced articles, which Wikipedia and Wikidata editors will keep.
• Opinions in reviews, forums and analyst notes, which you can answer but not remove.
• Variation between sessions: the same prompt can produce different answers on the same day.

What you control is the consistency of every source an assistant might read, which is also the base for our 30-day plan to get recommended by ChatGPT. If you want a structured starting point, our free AI search visibility checklist covers the site, source and crawler checks in this article.

Our generative engine optimization service runs this audit, the source corrections and the monthly re-test for B2B companies in English and French. If assistants are telling your buyers the wrong story, talk to our team.