Hallucination (AI)

A hallucination is what people call it when an AI system gives you an answer that sounds plausible and is delivered with total confidence, but isn't actually true, or isn't supported by any real source. For a business, this can mean a model stating the wrong phone number, inventing a service you don't offer, or citing a review that doesn't exist.

Grounding, where a model bases its answer on a live search or retrieved document instead of only its training data, reduces how often this happens, because the model has something real to check its answer against. It does not eliminate the problem. A grounded model can still misread a source, mash together facts from two different pages, or state something the source didn't actually say.

Be skeptical of any specific hallucination rate you see quoted for local business recommendations. Nobody has published a reliable, industry-wide measurement of how often AI systems hallucinate details about a specific local business, so a number like that is not something you should treat as fact, which is a little ironic given the subject.

The practical defense is the same one that helps with citations generally: keep the facts about your business (phone number, service area, pricing, licensing) consistent and current everywhere they appear, on your own site and on the directories and review platforms an AI system might pull from. The fewer conflicting or stale versions of your information exist, the less raw material there is for a model to get wrong.