Triaging AI Hallucinations About Your Brand

A wrong AI answer about your brand has one of three likely causes, and each needs a different fix. Rewriting your homepage won't change an answer the model produced from memory. Emailing a vendor won't change a third-party page the engine keeps citing. Triage tells you which lever you actually hold.

It helps to think in four stages:

• Parametric: what the model learned before its knowledge cutoff. Nobody changes this quarter.
• Access: whether an AI crawler can fetch your pages at all. You control this.
• Retrieval: whether a page enters the model's context for a given question. You control this only partly.
• Generation: what the model says about what it retrieved. You don't control this.

Most wrong answers trace back to the first stage, to a bad source in the third, or to your own pages. If you're unsure whether you face an invented fact or simply an old one, read hallucination versus outdated answers first.

Run these steps for each wrong claim. One claim per row in a spreadsheet is enough.

1. Capture the claim. Copy the exact sentence, not a summary. Record the platform, the date, the plan, and whether you were logged in. OpenAI says ChatGPT may use saved memories when it rewrites a search query, so test logged out or with memory off where you can.
2. Re-ask it twice. Ask the same question two more times in fresh chats. A claim that appears once is noise to log. A claim that repeats is a pattern to fix.
3. Classify the claim. Mark it provably wrong, outdated or unverifiable (defined in the next section).
4. Grade severity. Critical, moderate or minor, based on what it does to a buyer's decision.
5. Check for citations. Note whether the answer showed sources. Open every cited URL and search the page for the claim.
6. Route to a cause. Use the routing table below: training data, a wrong cited source, or your own site.
7. Fix, log and retest. Record the fix and the date. Retest the same question with the same settings later.

If you don't have a list of questions yet, the free checker on this site runs up to three buyer questions and gives you a snapshot to start from. For a deeper repair plan once you know the cause, see the fix plan for AI getting your brand wrong.

These three labels lead to different fixes, so don't merge them.

Provably wrong

The claim contradicts a fact you can document. The AI says you charge a price you never charged, names the wrong founder, or says you were acquired when you weren't. You can point to evidence that it is false.

Outdated

The claim was true once. Your old pricing, a retired product name, a previous office. This is usually a training-data or stale-source problem, not invention.

Unverifiable

The AI asserts something you have never published, such as a headcount or an integration list. It may be wrong, but nobody can check it against a source. Factuality research often lumps this in with false claims: FActScore's "not supported" label covers claims the source contradicts or lacks evidence for. Split them yourself. A wrong claim needs a source corrected. An unverifiable claim needs the real fact published somewhere an engine can fetch.

Severity

• Critical: changes whether someone buys or trusts you. Says you shut down, were acquired, had a safety or legal problem, or quotes a wrong price.
• Moderate: wrong category, missing or invented feature, wrong integration, a competitor's detail attached to you.
• Minor: headcount off, founding year off by one, small wording errors.

Work critical items first. Wrong prices are common enough to have their own guide: ChatGPT getting your pricing wrong.

Citations are your best clue. If an answer showed sources, retrieval ran. If it showed none, the answer most likely came from training data, though you can't be certain.

Checks for the own-site route

• JavaScript. Vercel's crawler study found ChatGPT and Claude crawlers fetch JavaScript files but don't execute them. A price that only appears after scripts run may be invisible to them.
• Crawler permission. robots.txt rules are per bot. OAI-SearchBot controls ChatGPT search; GPTBot governs training only.
• Firewall and CDN. A firewall can block a crawler that robots.txt allows. The AI Crawler Checker requests a page as each AI crawler and shows which ones get turned away.
• Google eligibility. Google says a page must be indexed and eligible to show with a snippet to appear as a supporting link in AI Overviews or AI Mode.
• Contradictions. Your pricing page, footer and about page should state the same fact the same way.

State the fact as a specific sentence: the price, the number, the one-line definition. That gives an engine a sentence it can quote. It does not guarantee the engine will use it.

Record these fields for every wrong claim. Missing fields make the routing step guesswork.

• The question asked, word for word
• The wrong claim, copied exactly
• Platform and, if shown, the model or mode
• Date and time
• Logged in or out, memory on or off
• Whether web search ran (citations or a search indicator shown)
• Every cited URL, and whether each one contains the claim
• Classification: provably wrong, outdated or unverifiable
• Severity: critical, moderate or minor
• The correct fact and the URL where you publish it
• Likely cause: training data, third-party source, or your own site
• Fix taken and date
• Retest date and result

Keep screenshots. Answers change between runs, and a screenshot is the only proof of what a buyer may have seen.

A good triage result

• Each claim has one named cause and one owner.
• Critical claims are traced to a specific page, yours or someone else's, or explicitly marked as training data.
• Unverifiable claims have a matching fact now published on a page crawlers can read in raw HTML.
• Retests use the same question and settings as the original capture.
• On retest, search-enabled answers cite a page that states the right fact.

A bad triage result

• "ChatGPT is wrong about us" with no exact wording or date.
• One run treated as proof of a pattern.
• Every claim blamed on hallucination, with no citation check.
• Fixes copied from generic advice, such as adding schema or an llms.txt file, with no link to the actual cause.
• A single blended score used to judge progress, hiding which engine is still wrong.

If the AI says it knows nothing about you rather than something wrong, that is a different problem. See why ChatGPT says it has no information about your company.

Run the full flow on a fixed schedule and after specific triggers.

• Monthly: re-ask your core buyer questions on each engine and compare against last month's log.
• After any change to a public fact: prices, plans, product names, leadership, locations. Old answers will linger.
• After a model launch or default change: a new model can fix old errors or introduce new ones. As of October 2026, the vendors ship these often.
• After firewall, CDN or hosting changes: Cloudflare set training and agent crawlers to blocked by default on ad-displaying pages from 2026-09-15 for new domains and existing free-tier customers. A site can lose AI access without anyone editing it.
• After each fix: retest with the same settings. Expect search-enabled answers to move before training-data answers, which only change with a new model. Nobody can promise a date.

For Google's surfaces specifically, see wrong facts in AI Overviews.

Can I ask OpenAI or Google to correct a wrong answer directly? You can use feedback buttons in the apps, but there is no documented process that edits what a model says about a specific brand. The durable fixes are on the sources: your own pages and the third-party pages engines cite. Feedback is worth sending for critical errors. Treat it as a log entry, not a fix. More detail is in our answer on [asking AI companies to correct brand information](/questions/ask-ai-company-correct-brand-info).

Will adding schema markup stop AI from getting my facts wrong? The evidence doesn't support that. Ahrefs tracked 1,885 pages that added schema and found no major lift in AI citations on any platform, with AI Overviews citations falling 4.6%. Existing markup does no harm, but it isn't a lever for correcting answers. Putting the true fact in plain text on a page crawlers can fetch is the step with evidence behind it.

Why does the same question give different answers each time? Retrieval is per query and at the model's discretion. One run may search the web and cite sources; the next may answer from training data. Engines also rewrite your question into their own search queries, and logged-in context such as memory can change them. That's why the flow asks you to re-ask in fresh chats and record whether citations appeared.

Can someone else's page make AI say false things about my brand? Yes. Research on LLM search agents shows that manipulated content, once it enters the result set, can change what an agent endorses, and vulnerability varies by model. Hidden text and prompt injection are illegitimate tactics and a real reputational risk. If a cited third-party page contains instructions or claims aimed at AI, record it, ask the host to remove it, and keep evidence.

Is a citation proof the model used that page? No. Citation and use are different things. OpenAI's search tooling exposes the pages it retrieved separately from the ones it cited, and the retrieved set is usually larger. A model can also cite a page that doesn't contain the claim. Always open the cited URL and search it for the exact wording before you blame that source.