Why it matters to a buyer
An assistant can state a wrong price, an outdated service list, or a service a business does not even offer, in the same confident tone it uses for anything true, and a buyer with no other presence in that model's memory has no way to correct it directly. This matters most for facts a model cannot easily verify live: pricing, guarantees and niche details are more prone to hallucination than well-documented, widely repeated facts, so a buyer with thin, inconsistent information online is at higher risk than one with clear, consistent public facts.
How it is measured or applied
Hallucinations about a specific brand are found by running a fixed set of buyer questions about the business across each major assistant on a schedule and recording any factual claim, with web search on and off where available, since grounded answers hallucinate less than answers pulled from training memory alone. Wrong claims are corrected at the source: publishing clear, consistent, dated facts on the site and on third-party profiles the model is likely to draw from, since there is no direct way to edit what a model already learned during training. Persistent, serious errors can be reported through each vendor's feedback channel.
Sources
- What are AI hallucinations?, Google Cloud
