peham

CEO notes · Wrong AI answers12 min read

Fix wrong AI answers.

When ChatGPT, Gemini, Perplexity or a Google AI Overview says something false about your company, you usually can't edit the answer. You can fix what the answer is built from. Capture the exact prompt and answer, check how widespread it is, correct the sources the engines trust, use the official feedback paths, and re-run the same prompt until the claim changes.

A 73-second walkthrough of the method in this article. Captions are burned in; a text transcript is below.
Transcript

When ChatGPT, Gemini, Perplexity, or a Google AI Overview says something false about your company, you usually cannot edit the answer. Here is the four-step check I use to fix what the answer is built from.

Step one: capture the exact prompt and the exact sentence the engine used. A screenshot is fine. A copy of the text is better, because I will compare the wording later.

Step two: I check how widespread it is. I run the same prompt in the other engines buyers use. Then I classify the problem: is the claim invented, out of date, or are you simply missing? The fix is different in each case.

Step three: I fix the sources I control, the About page, pricing, and Organization structured data, so engines have a correct page to trust. Then I use official feedback paths, like ChatGPT's content report, where they exist. There is no edit this answer button.

Step four: I re-run the same prompt over the following weeks, and I compare the wording of the claim. One clean answer proves nothing. The full playbook is in the article on my site. I'm Thomas Peham, CEO and co-founder of OtterlyAI.

Key takeaways

  • There is no "edit this answer" button. The fix is almost always in the sources, not in the engine.
  • Wrong, missing and out of date are different problems with different fixes. Diagnose before you act.
  • Your own pages come first: a plain-facts About page, a pricing page in readable text, and Organization structured data that matches both.
  • Use the feedback paths, but don't rely on them. ChatGPT and Google both document one; neither promises a fix.
  • Prove it with the same prompts. Re-run a fixed prompt set across engines and compare four-week windows.

The short answer: what to do in the first 48 hours

In the first 48 hours, document the wrong answer, measure how far it has spread, and trace it to a source you can correct. Rewriting your website in a panic comes later, if at all.

  1. Capture the evidence. Save the exact prompt, the engine, the date, the full answer and every cited link. A screenshot is fine; a copy of the text is better, because you'll compare wording later.
  2. Check how widespread it is. Run the same prompt, and two or three close variants, in the other engines your buyers use. Run each more than once: AI answers vary between runs, so one bad answer can be noise and five bad answers is a pattern.
  3. Classify the problem. Is the claim invented, outdated, or are you simply missing? The next section shows why this decides the fix.
  4. Find the source. Open the citations. The false claim often traces back to something real: an old pricing page, a directory listing, a press release, or a different company with a similar name.
  5. Fix what you control, then report. Correct your own pages and profiles first, then use the engine's feedback option (covered below).
  6. Re-check the same prompts on a schedule, and don't declare victory after one good answer.

Wrong, missing and out of date are three different problems

An AI answer can fail your company in three distinct ways, and each one needs a different fix. Treating all of them as "the AI is lying about us" wastes the first week.

ProblemWhat it looks likeTypical causeFirst fix
Wrong (hallucinated or mixed up)A false founder, a service you don't offer, a feature that doesn't exist, or another company's details attached to your nameA gap in clear, public facts about you, or a similarly named entity the engine confuses you withState the facts plainly on your own site and in Organization markup; correct third-party profiles; report the answer
Out of dateLast year's pricing, a retired product, a former CEO, an old office addressAn outdated page that still ranks or is still cited: your own old URL, a directory, an old press releaseUpdate or redirect the old page; ask the third-party site to update; make the current fact easy to find
MissingThe answer to "best tools for X" or "how do I solve Y" doesn't mention you at allNot enough content or third-party coverage for that questionThis is a visibility problem, not an accuracy problem. Treat it with content and earned coverage

The "mixed up" case is more common than it sounds. In one thread in Google's own Search help community, a business owner reported that an AI Overview attached a different company's services, service areas and pricing to their business. In another thread, a company said the Overview named a person who had never worked there.

The "missing" row is the one I see confused most often. If you're absent from category answers, read my guides on Generative Engine Optimization and the 4-pillar framework for AI search visibility instead. This article is about answers that mention you and get the facts wrong.

How I monitor for wrong AI answers

I catch wrong answers by running a small, fixed set of prompts about our own company across engines on a schedule, and reading the full answers, not just a score. Most false claims sit in exactly the questions a buyer asks right before a decision.

The prompt set is short and boring on purpose:

  • "What is [company]?" and "What does [company] do?"
  • "How much does [company] cost?"
  • "Who is [company] for?" and "Is [company] good for [use case]?"
  • "Who founded [company]?" and "Where is [company] based?"
  • "What are the alternatives to [company]?" (to see how you're described next to others)

Three rules make this work:

  • Same prompts, same engines, every time. If the prompts change, you can't tell whether the answer moved or your test did.
  • Store the full answer. When a claim disappears or appears, you want to read the sentence that replaced it.
  • Report brand prompts separately. Prompts that contain your name inflate visibility numbers. I keep them out of the headline metric in my monthly AI search report and use them as an accuracy check instead.

You can do this by hand with a spreadsheet for a pilot: prompt, engine, date, answer, cited URLs, correct or not. It becomes tedious once you add several engines and run it weekly. For general brand mention tracking, my 6-step guide to tracking brand mentions in AI search covers the setup.

I'll be transparent about why this article exists. In our own OtterlyAI tracking (US, 17 October to 15 November 2026), OtterlyAI did not appear in a single tracked answer to the prompt "how to stop AI from giving wrong info about my company." For a company that monitors AI answers for a living, that's a gap worth closing.

Correct the sources the engines already trust

Correct your own pages first, then the third-party pages the engines cite. AI search features retrieve and summarise existing web content; Google describes its AI Overviews and AI Mode as rooted in its core Search ranking systems, retrieving pages from its index to ground answers. If the pages are wrong, the answers will be too.

Your own site

  • A plain-facts About page. Company name, what you do in one sentence, who it's for, founding year, founders, headquarters, current leadership. No taglines in place of facts.
  • A pricing page in readable text. Assistants read pricing pages live. On otterly.ai, the pricing page received 24,080 of 67,238 AI agent visits in 30 days, more than a third of the total, as I described in my AI crawlers analysis. If prices sit in an image, behind a form or only in JavaScript, the engine will fill the gap from somewhere else.
  • Retire or redirect stale pages. Old pricing pages, sunset product pages and outdated "about" copy keep getting cited as long as they're live and indexable.
  • Organization structured data. Google says Organization markup helps it understand your organization's administrative details and disambiguate it from other organizations. Add name, url, logo, description, foundingDate and sameAs links to your official profiles, on your homepage or About page.
  • Keep markup and visible text identical. Google's guidance is explicit that structured data should match the visible content on the page. Markup that says one thing while the page says another is noise, not a fix.

A caveat on schema: Google also says structured data isn't required for its generative AI features and there is no special markup to add. Treat it as a way to state facts unambiguously, not as a switch that corrects answers.

Check access before you blame the engine

Engines can only use pages they can reach. OpenAI says that to be eligible for ChatGPT search results, a site must allow OAI-SearchBot and let its traffic through the host or CDN. Google requires a page to be indexed and eligible for a snippet to appear as a supporting link in AI Overviews. Blocking AI bots does not protect your brand; it removes your own version of the facts from the answer. I covered which bots to allow in AI Crawlers.

Third-party sources

  • Directories, review sites and marketplaces. Update your profiles and ask for corrections where you can't edit directly.
  • Google Business Profile, if you have one. Google lists keeping Business Profile information up to date among its AI features best practices.
  • Old press coverage. Ask the publisher to update or add a correction note when an article states a fact that is plainly wrong.
  • Wikipedia. Don't edit your own company's article. Wikipedia's conflict of interest guideline asks affected people to propose changes on the article's talk page instead.

What you can and cannot submit to the engines

You can send feedback about a wrong answer to ChatGPT and Google, but there is no way to edit the answer yourself and no promise of a fix. I use the feedback paths every time, and I never treat them as the plan.

EngineWhat you can doWhat it does not do
ChatGPTUse the thumbs-down under the answer and describe the issue. For content and sources, OpenAI has a content reporting form; it says reported domains may be reviewed by its Model Quality team, which may apply filters so ChatGPT avoids relying on unreliable sources.Correct a specific fact on request, or tell you when anything changed
Google AI OverviewsSelect the thumbs-down at the bottom of the Overview, then Report a problem and describe the issue, per Google Search Help. Volunteers in Google's community threads also suggest choosing "Not factually correct" and explaining what's wrong.Let you edit or remove the Overview. Google says AI Overviews are a core Search feature and can't be turned off
Other assistantsCheck each assistant's help docs for its in-product feedback option and use it with the exact prompt and claim.Guarantee a correction

OpenAI itself is clear about the underlying limitation: its help page Does ChatGPT tell the truth? says the model can produce incorrect or misleading outputs, sometimes confidently, and lists incorrect facts and fabricated citations as examples. Feedback helps the engines improve. Corrected sources are what change the next answer.

If a false claim is damaging rather than just inaccurate, talk to a lawyer. That's outside what any GEO process can fix.

How I know the fix worked

A fix worked when the same prompts, on the same engines, stop producing the false claim across a full four-week window. One clean answer on a Tuesday proves nothing, because answers vary from run to run.

  • Re-run the exact prompts you captured on day one, plus the variants, on every engine where the claim appeared.
  • Compare the wording of the claim, not just whether you were mentioned. A wrong founding year being replaced by the correct one, in the engine's own words, is the result you want.
  • Watch the citations. When your corrected page starts being cited for that prompt, the engine is using your version.
  • Give it time. Google notes that recrawling can take anywhere from several days to several months, depending on the page. Request a recrawl of the corrected page in Search Console to speed it up.
  • Check Search Console. Google points site owners to the Generative AI performance report in Search Console to see how content performs in its generative AI features.
  • Annotate the change. Log the date you fixed each source, so a later shift has an explanation. I use the same habit in my monthly AI search report.

Disclosure, and how I do this in OtterlyAI

I'm the CEO and co-founder of OtterlyAI, so weigh this accordingly. We use our own product for exactly this check. The OtterlyAI AI Prompt Tracker runs the same prompts across ChatGPT, Perplexity, Google AI Overviews, Google AI Mode, Gemini, Copilot and Claude on a daily or weekly schedule, stores every full answer, and scores sentiment per prompt, so when a claim about you changes you can read the answer that replaced it. Plans start at $29/month.

The short version: capture the answer, classify the problem, fix the source, report it, and re-check the same prompts for a month. If you're new to monitoring, start with how to track brand mentions in AI search, then build the monthly view with the AI search report I want as a CEO.

FAQs

How do I stop AI from giving wrong information about my company?

You usually cannot edit an AI answer directly. Capture the exact prompt and answer, check whether other engines repeat it, find the source the claim comes from, and correct that source: your About and pricing pages, your Organization structured data, and third-party profiles. Then send feedback through the engine's own reporting option and re-run the same prompt over the following weeks.

How do I report a wrong ChatGPT or Google AI Overview answer about my brand?

In ChatGPT, use the thumbs-down under the answer and describe the error; OpenAI also has a content reporting form, and says reported domains may be reviewed by its Model Quality team. In a Google AI Overview, select the thumbs-down, then Report a problem, and describe the incorrect claim. Neither path guarantees a change or a timeline, so fix the underlying sources as well.

How is an AI hallucination about a brand different from low AI search visibility?

A hallucination is an answer that mentions your brand with a false fact, such as a wrong founder, price or service. Low visibility is an answer that leaves your brand out. The first is an accuracy problem you fix by correcting sources; the second is a visibility problem you fix with content and third-party coverage.

How often should I check whether AI engines still have the facts wrong?

Run a small, fixed set of brand-fact prompts on a schedule, weekly at minimum and daily while you are actively correcting something. Compare four-week windows rather than single days, because AI answers vary from run to run.

Does adding Organization schema fix incorrect AI answers?

Not on its own. Google says Organization structured data helps it understand and disambiguate your organization, but also that structured data is not required for its generative AI features. Use it to state your facts clearly, keep it identical to the visible text on the page, and pair it with corrected page copy and third-party profiles.

What is the fastest way to monitor AI hallucinations about my company?

Track a fixed list of prompts about your company (what it is, what it costs, who it is for) across the engines your buyers use, and store the full answers so you can see when a claim changes. You can do this by hand for a pilot; a monitoring tool automates the runs across engines.

Keep
reading.

All writing →
OtterlyAI

See how your brand shows up in AI search.

Track mentions, citations and share of voice across ChatGPT, Perplexity, Gemini, Google AI Overviews, AI Mode and Copilot.

Try OtterlyAI ↗

Talks · advisory · press

Want Thomas on your stage or podcast?

Get in touch →