AI hallucination: definition and common causes for brands

A hallucination is a model asserting something its sources don't support. FActScore (Min et al., 2023), a standard factuality benchmark, marks a claim as not supported when the source contradicts it or lacks evidence for it. That line works well for brands. A wrong price and an invented integration are both hallucinations, even though one contradicts a fact and the other has no fact behind it.

A confident, fluent answer tells you nothing about which kind you're looking at. This entry is a tight definition. For causes and responses in depth, read the guide to AI hallucinations about brands.

Most brand errors trace back to one of four causes. Each one sits at a different stage of how an answer gets built.

• Stale parametric memory. With web search off, a model answers from training data frozen at a cutoff. OpenAI's model documentation lists a February 16, 2026 cutoff for the GPT-5.6 family and August 31, 2025 for the Instant path ChatGPT also uses. Anything you changed after the cutoff can't reach an answer given without retrieval. Nobody controls this stage this quarter. See how a hallucination differs from an outdated answer.
• Unreachable pages. With search on, the model can only use what its crawler could read. Vercel found that ChatGPT and Claude crawlers fetch JavaScript files but don't execute them, so client-rendered content may be invisible. A robots.txt block on a search bot, or a firewall returning 403, has the same effect. The model then fills the gap from memory or other sites.
• Contradictory sources. If your pricing page, an old press release and a review site state different numbers, the model has to pick one. Mistakes on your own website are the easiest of these to fix.
• Conflation. Brands with similar names, or close competitors in one category, can get their facts merged into one answer.

Retrieval is per query and often discretionary, as of October 2026. The same question can get a grounded answer one time and a memory-only answer the next.

Being named in an answer only helps if the answer is right. A buyer who reads a wrong price or a feature you don't offer gets a worse picture than one who never saw your name.

The useful news: wrong facts are specific and checkable. Start here:

1. Ask the same buyer question with web search on and off, and compare the answers.
2. Load your pricing and product pages with JavaScript disabled. Confirm the key facts appear in the raw HTML.
3. Check that robots.txt allows OAI-SearchBot, PerplexityBot and Claude-SearchBot, and that your CDN doesn't block them.
4. Compare each core fact (price, founding year, product names) across every page that states it.
5. Search third-party pages that describe you for outdated or wrong claims.

The free checker on this site shows what engines say for up to three buyer questions, which tells you where to start. For a full sequence, use the hallucination triage playbook.

• "Schema or an llms.txt file will stop it." Google says Search ignores llms.txt, and structured data isn't required for its generative AI features. No engine documents using either to correct facts.
• "A citation proves the claim." Research on how models cite (arXiv 2606.28358) shows citation output can be changed without changing the sources, so citation and use are separable. See why citations sometimes don't support the claim.
• "A tool can prevent it." Nobody outside the model operator controls generation. Tools can detect errors and point to fixable causes. Read whether tools can prevent hallucinations.
• "Hallucination always means invented." Many brand errors are outdated or merged facts, not fabrications.

• Outdated answer: a claim that was once true but no longer is.
• Grounding: an answer built from pages retrieved at answer time.
• Knowledge cutoff: the date after which a model's training data stops.
• Content injection: hidden instructions placed on pages to manipulate AI answers. It is illegitimate and a risk to brands described by third parties.

For a common case, see what to do when ChatGPT gets your pricing wrong.

Is an outdated AI answer the same as a hallucination? Not quite. An outdated answer repeats something that was true before the model's cutoff or before a page changed. A hallucination in the narrow sense has no source behind it. Both are unsupported by current evidence, so both count as wrong for a brand. The difference matters for the fix: outdated answers point to cutoffs and stale third-party pages, while invented claims often point to missing or unreadable facts.

Why does one AI platform get my facts right and another get them wrong? Each platform uses its own models, crawlers and indexes, and decides on its own when to search. A page one crawler reads cleanly may be blocked or unreadable for another. One engine may answer from memory while another retrieves your current page. Check each platform separately, and check each crawler's access separately. A single blended score hides which engine is wrong and why.

Can pages on other sites cause AI hallucinations about my brand? Yes. When engines cite other sites instead of yours, those sites shape the answer, including their errors. Some third-party pages also carry injected instructions meant to manipulate AI answers. A 2026 study (arXiv 2606.16821) measured that vulnerability in search agents using a mock search API. Scan pages that describe you for wrong facts and suspicious hidden text.

How often do AI models hallucinate about brands? No published study gives a reliable rate for commercial brands. Rates in the research depend on the benchmark, the model version, whether search was on and how claims were scored, so any single figure for brands should be treated with caution. What you can measure is your own case: which questions, on which platforms, return which wrong facts, and whether those errors repeat over time.