AI Hallucinations About Brands: Why Models Make Up Company Facts and How to Respond

An AI hallucination is a claim a model states with confidence that no source supports. About brands, it shows up as a wrong price, a made-up headcount, an invented founding year, a funding round that never happened, or a product feature you retired years ago.

Not every wrong answer is a hallucination in the strict sense. Some are outdated: the model learned a true fact that has since changed. Some are conflation: the model blends your company with another one that has a similar name, which is brand conflation. Some come from a real page that is simply wrong.

The distinction matters because each cause has a different fix. A fabrication needs a clear fact to replace the gap. An outdated answer needs a current fact the engine can fetch. A wrong source page needs correcting at the source.

Four stages decide what an AI says about you. Errors can enter at each one.

Parametric knowledge (training)

With no web access, a model answers from what it learned before its knowledge cutoff. OpenAI's model documentation, checked in August 2026, lists a February 16, 2026 cutoff for the GPT-5.6 family and August 31, 2025 for the Instant path used in ChatGPT. So "what ChatGPT knows" is not one value. Anything you changed after a cutoff is invisible to an answer given without search. When the model has weak or no data about a niche company, it still produces fluent text, and fluent text about a headcount or a founding year can be invented. Our page on why AI makes up details like headcount covers that pattern.

Access

If an AI crawler can't fetch your page, the engine can't use it. Vercel's crawler study found that ChatGPT and Claude crawlers fetch JavaScript files but "don't execute them. They can't read client-side rendered content." A price that loads only after JavaScript runs may not exist for those engines.

Retrieval

As of October 2026, retrieval is discretionary on every major platform. ChatGPT decides per query whether to search. Claude.ai search is a user toggle with no published default. So the same question can get a grounded answer one time and a training-only answer the next.

Generation

Even with sources in hand, the model chooses what to use and cite. Contradictory sources (your pricing page says one thing, your footer another) give it two answers to pick from. Third-party pages that mention you can also be wrong, or deliberately manipulated.

Academic factuality metrics share one rule: they divide errors by the number of things checked. None of them simply adds up mistakes.

• FActScore (Min et al., EMNLP 2023) splits a response into atomic facts and scores the share supported by a knowledge source. Its authors state that "every atomic fact has an equal weight of importance."
• SAFE (Wei et al., NeurIPS 2024) adds a recall term. The authors argue that precision alone "does not penalize systems that abstain from answering too often or generate text with fewer facts."
• HaluEval 2.0 reports two rates: hallucinated statements over all statements, and responses with at least one hallucination over all responses.

Two lessons carry over to brand monitoring. First, count errors against the number of answers or claims you checked. Ten errors in 200 answers is a different situation from ten errors in 20. Second, a pure average can hide one serious error. SAFE exists partly because a rate alone lets a single bad answer dilute away as the sample grows. A claim that you were acquired by a competitor deserves attention even if 199 other answers are clean.

One caveat: weighting errors by severity is not standard benchmark practice. If a tool ranks a wrong price above a wrong headcount, that's a product judgment, which is reasonable but not research-backed.

FActScore's "not supported" label covers two different things: a source contradicts the claim, or no source has evidence for it. For a brand, splitting those apart tells you what to do next.

A specific fabricated number is usually worse than a vague true statement, because buyers act on numbers. Prioritize provably wrong prices, ownership and product claims first. The hallucination vs outdated answer comparison walks through diagnosing which one you have.

Be plain about the boundary. It saves wasted effort.

In your control

• Whether your own site states each key fact once, consistently, everywhere it appears.
• Whether those facts are in the HTML before JavaScript runs.
• Whether robots.txt allows each retrieval crawler, and whether your CDN or firewall lets them through.
• Whether you've removed accidental noindex, nosnippet or data-nosnippet rules from pages carrying facts.
• Whether you've asked directories, review sites and partners to correct wrong listings.

Not in your control

• What a model learned before its cutoff. Nobody can edit that this quarter.
• Whether an engine searches for a given question.
• Which retrieved pages the model uses or cites.

One point on crawlers trips people up. OpenAI's crawler page says GPTBot governs training and OAI-SearchBot governs ChatGPT search, and "each setting is independent of the others." Blocking GPTBot does not remove you from ChatGPT search, and allowing it does not get you cited. The same separation exists for other vendors, such as ClaudeBot versus Claude-SearchBot.

Work from ground truth outward. Each step depends on the one before.

1. Write down the facts. List the claims buyers rely on: pricing, plans, founding year, headquarters, ownership, product names, integrations. A brand fact sheet template gives you the structure. This is the reference every answer gets checked against.
2. Check what the engines say. Ask each platform the same buyer questions and record every claim. Note whether the answer showed sources, since an answer without them may come from training data alone. The free checker on this site runs up to three buyer questions as a snapshot.
3. Classify each error. Use the four types above. Count errors against total answers, and flag any single critical error separately.
4. Fix your own site first. Hunt for the same fact stated differently on the pricing page, about page, footer and old blog posts. Mistakes on your own website are the most fixable source of AI errors.
5. Make facts fetchable. Fetch key pages with JavaScript disabled and confirm the facts appear in the raw HTML. If they don't, server-render them.
6. Confirm crawler access. Check robots.txt per bot, then check that your firewall doesn't return 403 to crawlers robots.txt allows. The AI Crawler Checker requests a page as each AI crawler and shows which ones get turned away.
7. Correct third-party sources. When engines cite other sites instead of yours, being named accurately on those sites is associated with appearing in the answer.
8. Re-check on a schedule. Grounded answers can change once fixed pages are recrawled. Training-only errors persist until the next model, which is why correcting facts past a knowledge cutoff is a longer game.

• Adding schema to stop hallucinations. Google says structured data isn't required for generative AI search. Ahrefs' matched test of 1,885 pages found AI Overviews citations fell 4.6% after schema was added. Existing markup does no harm, but it isn't a fix.
• Publishing an llms.txt. Google says Search ignores it, and no other engine documents reading it.
• Inventing credentials. In Otterly's 2026 experiment, the AI-generated site with a fictional expert author box was dropped from Google after about 14 days, versus about 90 for the site with no author. One site per arm proves little, but it is no evidence that fake authority helps.
• Mass-publishing pages to drown out errors. That same experiment ended with both sites deindexed by Google, and Google's AI surfaces stopped citing them.
• Hidden text or prompt injection. Content aimed only at AI crawlers exists, and research shows it can sway some search agents once retrieved. It's illegitimate, it can be detected, and it puts your reputation at risk.
• Trusting one blended score. Each engine has a different cutoff, retrieval behavior and source mix. Measure them separately.
• Counting errors instead of rates. More questions will surface more errors. Judge the share, not the total.