A prospect asks ChatGPT about your pricing. The answer comes back with a plan you discontinued eighteen months ago. A journalist uses Perplexity to research your company before a call. The description it returns is half accurate, half a blend of you and a competitor with a similar name. A buyer asks Claude whether your product integrates with a platform they use. It says no. The integration has existed for two years.
These are not edge cases. Hallucination rates on brand-specific information run between 20% and 76% depending on the engine and query type.
Why AI Gets Your Brand Wrong
Three failure modes produce most brand hallucinations.
Conflicting sources. If your website says one thing, an old press release from 2023 says another, and a third-party review site scraped from that press release says a third, the model tries to resolve the conflict by picking the most plausible version. It often picks the wrong one. Pricing, feature lists, and product names are the most common casualties.
Missing entity signals. When AI engines build a picture of your brand, they need clear, consistent entity declarations: your official name, product category, audience, key people, and how you relate to other known brands. If these are absent from your schema, missing from your main pages, or inconsistent across your site, the engine infers what it can and invents what it cannot.
Training data lag. Every AI engine has a knowledge cutoff from its last training cycle. A product launch from six months ago may simply not exist in the model's world yet.
The Detection Audit
A real detection pass covers the prompt types where brand-specific inaccuracies are most damaging:
- Branded definition prompts. "What is [Your Brand]?" "What does [Your Brand] do?" These test whether the model has an accurate baseline picture.
- Product and feature prompts. "What are the features of [Product Name]?" "Does [Your Brand] have [specific feature]?" "What does [Product Name] cost?" These are the prompts where outdated information causes direct commercial damage.
- Comparison prompts. "[Your Brand] vs [Competitor]." These test whether the model correctly distinguishes you from competitors.
- Category recommendation prompts. "Best tools for [category your brand serves]." These test whether you appear in recommendation answers and in what context.
Run each prompt type across at least three engines: ChatGPT, Perplexity, and Claude. Document exactly what each engine says and which sources it cites. Save outputs in a hallucination register with the date, prompt, engine, inaccuracy type, and commercial impact severity.
Why Updating Your Website Alone Does Not Fix It
The instinct when you find a brand hallucination is to update your website. That is necessary but not sufficient. Training data does not update in real time for closed models. The incorrect information will continue appearing in model memory until the next training cycle incorporates the corrected version, which can take months. In the meantime, the problem persists in live retrieval if the sources feeding that retrieval still contain the old information.
The more important intervention is fixing the third-party source layer. If Crunchbase still shows your old product description, if a comparison article from 2023 still lists discontinued pricing tiers, those sources are feeding the incorrect answer. The engine is not making things up: it is accurately reflecting what the most-cited sources say, and those sources happen to be wrong.
The Fix Sequence
- Owned content first. Your homepage, About page, pricing page, product feature pages, and FAQ page each need to state your core brand facts in a format a retrieval system can extract cleanly. Add or update Organisation schema with current information and current sameAs links.
- High-authority third-party profiles. Check and correct Crunchbase, LinkedIn, G2, Capterra, Clutch. Pay attention to product category fields, pricing information, and key personnel.
- Historical content. Search for your brand name combined with terms from the hallucinated answer. Find the page making the incorrect association. Contact the site owner or redirect and update your own page.
- Authoritative new content. For persistent hallucinations, publish a structured page specifically designed to answer the question the AI is getting wrong. Within 14 days in documented cases, AI engines updated their answers after a new, clearly structured page was indexed.
Track your target prompts weekly throughout the correction process. The NotionCue Prompt Tracker surfaces changes in how your brand is described across engines on each tracked prompt, so you can confirm when a hallucination has resolved rather than guessing.