GLOSSARY
Hallucination
- Hallucination
- A hallucination is an AI answer stating something false with confidence — a wrong address, a product you don't sell, a fee structure you don't have. For businesses it's a reputation surface no one reviews. You can't correct the model directly, but you can fix the sources it grounds on, which is where corrections actually happen.
Hallucinations happen when a model generates from memory where its knowledge is thin, stale or contradictory — and business facts are exactly that kind of knowledge. Old addresses survive in training data after a move; similarly named companies blur together; a discontinued service keeps being offered in answers. The result reads as authoritative as everything else the engine says, and no notification tells you it's happening.
The practical response has three parts. First, discover: run the prompts your customers actually ask — including questions about your own business by name — across engines on a schedule, and log what comes back. Second, correct the inputs: hallucinations about businesses usually trace to weak or conflicting sources, so a strong entity footprint (consistent directories, current structured data, clear first-party fact pages) gives grounding something better to anchor on. Third, keep the record: dated, verbatim logs of inaccurate answers — for regulated businesses especially, that documentation belongs in your compliance file.
Example: an engine tells a prospect a financial firm charges commissions when it's fee-only. The trace leads to an outdated third-party profile; the profile gets corrected, the firm's fee page gets an extractable plain-language answer, and subsequent retrievals state it correctly. That loop — monitor, trace, fix the source — is a standing part of our monitoring service, because the answers change whether you're watching or not.