Free AI Hallucination Check for Your Brand

A hallucination check answers one narrow question: when an AI assistant describes your company, which statements match the facts? It does not tell you whether you are visible or how often you are named. It tells you what is wrong and where the error probably came from.

That second part matters most. The same wrong price can come from two places. It can come from the model's training data, which is frozen at a cutoff date. It can also come from a page the assistant fetched while answering. The fixes are different, so the check has to separate them. Our guide to hallucinated versus outdated AI answers covers that diagnosis in more depth.

You can't call an answer wrong without a reference. Write the list before you open any chatbot.

Keep it short and specific. Ten to fifteen facts is enough. Good candidates:

• Current prices and plan names
• Founding year, headquarters city and ownership
• Product names, including any you have retired
• Key integrations and certifications
• One-line description of what you do and who it serves

For each fact, record three things: the exact wording, the URL where you publish it, and the date you last confirmed it. Then give each fact a severity. A wrong owner or a wrong price is critical. A wrong headcount is minor. This weighting is a practical judgement, not a standard: the FActScore benchmark gives every fact equal weight.

Finally, check your own site for contradictions. If your pricing page, footer and about page state the same fact differently, an assistant may repeat any of them. Fix those first.

1. Confirm your facts are in the raw HTML. Many AI crawlers download JavaScript but do not run it. Vercel found that ChatGPT and Claude crawlers "can't read client-side rendered content". Fetch the page without a browser and search for the fact:

curl -s https://example.com/pricing | grep -i "per month"

If the price only shows up in a browser, crawlers that skip JavaScript never see it.

2. Check crawler access. Open yoursite.com/robots.txt. Look for rules on OAI-SearchBot (ChatGPT search), PerplexityBot and Claude-SearchBot. OpenAI states its bot settings are independent, so blocking GPTBot only affects training.

3. Use a clean session. Log out, or turn memory off. OpenAI says ChatGPT may use saved memories when rewriting a search query, which makes results hard to repeat.

4. Ask each question twice. Once with web search off, once with it on. Use plain buyer questions: "What does [brand] cost?", "Who owns [brand]?", "What does [brand] do?"

5. Record everything. Save the answer text, the date, the surface (Gemini app, Google AI Overviews and AI Mode are different products), and every cited URL.

6. Score each claim against your list using the categories below.

Sort every factual claim in each answer into one group.

Keep "wrong" and "unverifiable" apart. An assistant claiming you have 200 staff when you never published a headcount is not proven false. It is a gap in your list. Academic benchmarks blur this line: FActScore's "not supported" label covers both contradicted and unevidenced claims. For your purposes, only contradictions are errors.

Then diagnose each wrong claim:

• Wrong with search off, and the fact changed after the model's cutoff: probably outdated training data. As of October 2026, OpenAI lists a February 16, 2026 cutoff for its GPT-5.6 models and August 31, 2025 for the Instant path ChatGPT uses.
• Wrong with search on, with citations: a retrieval error. Open the cited pages. One of them usually states the wrong fact.
• Wrong with search on, no citations: you can't tell which. Note it and retest.

The hallucination triage playbook walks through ranking these by severity.

Retrieval errors are the ones you can act on. If the cited page is yours, correct it and make sure the fact is in the server-rendered HTML. A specific fact stated plainly on a fetchable page gives the engine a sentence it can quote. If the cited page belongs to someone else, such as an old review or directory listing, ask that site to update it. When engines cite other sites instead of yours, those pages are the route into the answer.

Access problems look like missing or stale facts. Fix robots.txt for the search bots, then confirm your CDN or firewall doesn't return a 403 to them. A crawler that robots.txt allows can still be blocked at the network layer.

Outdated training data is the hard case. Nothing you publish today reaches an answer given with search off. The fix is making sure the correct fact is reachable for answers that do search, and waiting for a later model. You can also report errors to vendors; see whether OpenAI or Google will correct your company's information.

No fix guarantees the next answer is right. You control what engines can read. You don't control what they say.

A hand check is a sample. Keep these limits in mind:

• Answers vary. The same question can produce different answers minutes apart.
• Retrieval is discretionary. ChatGPT decides per query whether to search. Without visible citations, you don't know which kind of answer you got.
• Cited is not the same as used. OpenAI exposes more retrieved sources than it cites, so the citation list is incomplete.
• Google has several surfaces. Google says AI Overviews and AI Mode "may use different models and techniques". Check them separately; see checking AI Overview accuracy.
• No tool sees real user prompts. Your questions stand in for what buyers type.

Prices are the most common critical error. The free pricing accuracy check focuses on that one fact type.

Running six questions across several assistants by hand takes time, and keeping the records consistent is the tedious part. The free checker on this site runs up to three buyer questions, yours or ones written from your homepage, and scores the answers in one pass. It is a snapshot, not ongoing tracking.

For a check you can repeat from a standalone page, the AI Visibility Checker runs the same kind of test. Either way, keep your verified fact list. It is what turns an answer into a finding. If ChatGPT is the main problem, the guide on ChatGPT having wrong information about your business covers platform-specific steps.