Brand hallucination
A brand hallucination is when an AI assistant presents false information about a company as fact, such as wrong pricing or a claim that the company has shut down, with nothing in the answer to signal uncertainty.
Brand hallucinations tend to be concrete and checkable: a price quoted from a plan you retired in 2024, or a confident claim that a feature exists when it never shipped. In our July 2026 study, 27 of 30 brands were materially misdescribed by an AI assistant answering from memory, and 13 of 30 were still misdescribed when the assistant had live web search. Because the answer arrives in fluent, assured prose, buyers rarely question it.
The mechanics are mundane. Model training data has a cutoff, so an assistant answering from memory may describe your 2024 pricing as current. Gaps get filled by pattern completion: if most tools in your category ship an API, the model tends to assume yours does too. Live web search only narrows the problem, because retrieval often lands on stale third-party pages rather than your own site, such as an old G2 listing or a comparison post written before your last pricing change.
To detect a brand hallucination, ask the assistants what a prospect would ask: what the product costs, or whether it integrates with Salesforce. Compare each factual claim in the answer against ground truth, and repeat on a schedule, because answers drift as models update and retrieval indexes change. Discoverable runs these checks: the free AI visibility checker needs no login for the snapshot, and paid plans add weekly automatic re-checks. Claude checks are live today, with other engines rolling out.
Correction runs through the sources engines read, since you cannot edit a model's weights directly. Fix the surfaces retrieval pulls from: an accurate pricing page marked up with schema.org Offer data, plus corrected listings on G2 and Crunchbase. If your company has a Wikipedia article, correct that too; it feeds both model training data and live retrieval. Memory-based errors take longer to clear, because the fix only lands when the model is next trained on fresh crawls, which is a reason to publish accurate, crawlable facts early rather than after the damage shows up.