Introduction

AI engines describe your brand to buyers every day, and they do not always get it right. An answer may quote last year’s pricing, list a feature you retired, mix you up with a company that has a similar name, or say you do not support something you launched months ago. This playbook shows how to find wrong AI claims about your brand, trace each one to the pages that feed it, and correct those sources so engines have better material to work from.

Who it is for: brand, product marketing, PR and support teams who hear “but ChatGPT said…” from prospects and customers, and want a repeatable way to deal with it.

One thing to set out up front: you cannot edit an AI answer. What you can change is what engines read, and how consistently the facts appear across your own site and the third-party pages they trust. That is slower than editing a page, but it is the part you control, and it is the part you can measure.

The challenge

Wrong claims are hard to see. They sit inside long answers, they vary by engine and by prompt, and they can change from one daily run to the next. Most teams only learn about one when a customer repeats it on a sales call, and by then nobody knows which engine said it, when, or why.

The cause is rarely a single page. An outdated claim can come from an old pricing page you forgot to redirect, a review written two years ago, a forum thread, a directory listing nobody updates, or a competitor’s comparison table. When several of those sources agree, one correction on your own site may not be enough to change the answer.

Automated sentiment can also mislead. An answer can read as neutral while it states something false, and a correct answer can be labeled negative because it mentions a limitation. You need to read the words, not just the label.

The approach

Work in four passes: detect, trace, correct and measure. aeoora covers each pass with features that sit next to your visibility data, so the routine fits into the place you already check every week.

  • Detect: the Mentions tab shows every tracked answer that names you, with the text around your brand, the sentiment, your position and the date. The Perception view shows the words and ideas engines tie to your brand.
  • Trace: Citations lists the domains and URLs cited in those answers, split into your own pages, competitor pages and third-party sites.
  • Correct: the Fix wrong claims skill in the Chat agent drafts a correction plan from your data, and the Content Editor helps you update the pages.
  • Measure: send the issue to the Action Center, where completed actions are checked at 7, 14 and 30 days.

Keep each correction small and specific. “Our starter plan now includes X” is a fix you can ship and check. “Engines misunderstand our positioning” is not, so break it down until each item names one claim and one correct fact.

Step by step

Work through these steps once to clear the backlog, then repeat steps 3 to 10 weekly while corrections are open.

  1. Write your fact list: list the facts that most often go wrong, such as plan names and prices, what each plan includes, supported integrations, regions you serve, your company name and history, and the things you do not do. Save the list in Chat’s Memory page so the correct facts are on hand whenever Chat drafts a fix.
  2. Track the prompts that surface claims: add a small set of prompts that name your brand, such as “how much does [brand] cost” or “does [brand] integrate with [tool]”. This is one of the few cases where naming yourself in a prompt makes sense. Put them in their own topic so you can read them apart from your discovery prompts.
  3. Read the Answers view: open Visibility, go to Mentions and switch to Answers. Filter by topic and engine, then read the context text for each row. Note every claim that contradicts your fact list, with the prompt, engine and date. Export CSV, up to 5,000 rows per file, if you prefer to work through them in a sheet.
  4. Check Perception: look at the words engines associate with you. An old product name, a retired feature, or a label like “expensive” or “enterprise only” that no longer fits points you to the claims worth fixing first.
  5. Correct wrong sentiment labels: if a mention is labeled negative but reads fine, change the label in the Answers view. This keeps net sentiment honest, but remember that it changes your report, not the answer itself.
  6. Trace the sources: open the prompt to see which pages the engine cited for that answer. In Citations, filter by prompt and engine, then switch between Brand and Third-Party. If no page is cited, the claim likely comes from what the model already knows, which tends to change more slowly.
  7. Run the Fix wrong claims skill: in Chat, type / in an empty box, pick Fix wrong claims and give it the claim and the correct fact. Review the plan it drafts: which of your pages to update, the wording to use and which outside sources to contact. Chat shows what it will do before it does anything.
  8. Fix your own pages first: update the source pages in the Content Editor. State the correct fact in plain words near the top of the page, add a short FAQ entry for it, and update or redirect outdated pages instead of leaving them live. Check that your Organization schema matches the facts. Send the page to your CMS; IndexNow runs on publish.
  9. Ask third parties to correct: for each outside page that repeats the claim, send a short, polite note to the owner with the correct fact and a link to your updated page. Update your own profiles on review sites, directories and marketplaces yourself, since those are usually the fastest wins.
  10. Track it in the Action Center: send the mention to the Action Center so it becomes an action with tasks and a check method. When you close it, History measures the result at 7, 14 and 30 days.

Tip: Feedback buttons inside an assistant may help over time, but you cannot rely on them, and they do not tell you whether other people still see the claim. Changing the sources is the part you can measure.

Prioritize the claims that cost you

Not every inaccuracy deserves the same effort. Start with claims about price, availability, security and integrations, because those stop deals. Then fix claims that appear on several engines or several prompts, since they usually point to a source many engines share. Leave one-off oddities on a single engine until the bigger items are done.

What to measure

Corrections take time to show, so measure the same way each week and look at the trend rather than a single run.

  • How many tracked answers still repeat the wrong claim, counted from the Answers view or your CSV export.
  • Net sentiment and the Perception terms for the affected topic.
  • Whether your updated page is now cited for the prompts where the claim appeared.
  • Third-party corrections requested and confirmed.
  • Result checks at 7, 14 and 30 days for each closed action, read as Improved, Mixed, No meaningful change or Declined.
  • Time from spotting a claim to shipping the fix.

Engines that search the web live usually pick up a corrected page sooner than answers drawn from model knowledge. If a claim keeps returning on one engine with no citation, keep your pages consistent and check again after a few weeks, rather than rewriting the same page again and again.

Next steps

Read the guides on mentions and sentiment, brand perception and citations, and see how Chat skills and Memory work. The free Organization schema generator helps you state the basic facts about your company in markup, and our post on brand mentions vs citations explains why the two need different fixes. Compare plans on pricing, or start a 2-day free trial and run your first claims check this week.

More playbooks