Hallucination
Updated 2026-09-15
A hallucination is an AI-generated statement presented confidently but unsupported by any retrieved source or by fact. In brand terms it usually appears as an invented feature, an outdated price, or a capability attributed to the wrong product.
How it shows up in brand answers
Three patterns account for most brand hallucinations. Memory answers describe a product as it was in the training data, which is often a year stale. Ambiguous names cause attribute transfer between similarly named products. And thin documentation invites the model to fill gaps with category defaults — if your page never states what you do not do, the model may assume you do. Measurement means reading sampled answers, not only counting mentions.
Illustrative case (example, not a customer measurement)
Five of thirty sampled answers describe a product as having a free tier it has never had. Four of the five were ungrounded memory answers; the fifth cited a two-year-old review. The correction is a current, explicit pricing statement plus an outreach to the stale source — not a change to the prompt set.
Common mistakes
- Counting a hallucinated mention as visibility. A wrong description can be worse than absence.
- Fixing only your own page when the error is being carried by a third-party source engines prefer.
- Assuming a correction propagates immediately — engines re-crawl and re-summarise on their own schedule.
Frequently asked questions
- Can hallucinations about our brand be prevented?
- Not eliminated, but they can be starved: current, explicit, easily retrievable statements about what the product is and is not leave less room for the model to guess, and grounded answers hallucinate markedly less than memory answers.
- Should hallucinations be tracked as a metric?
- They should be read rather than scored. Sampling answers and recording factual errors as a list gives you something actionable; a "hallucination rate" invites arguing about the denominator.
