AI Hallucination

For local businesses the common hallucinations are mundane and costly: wrong hours, wrong service area, a defunct phone number, or attributing a competitor's specialty to you. The model is not malicious — it is completing a pattern from thin or contradictory evidence.

The mechanism is worth understanding, because it explains which errors you can prevent. A language model generates the most plausible continuation of text. When the retrieval step returns clear, consistent facts, the most plausible continuation is the truth. When it returns three sources with three different phone numbers, or nothing at all about your Saturday hours, the model still produces a confident sentence — it just produces the most typical one. Silence in your evidence is not read as silence. It is filled in with the average of your category.

That is why the errors cluster where documentation is thinnest. Assistants rarely invent a business name from nothing; they routinely invent hours, pricing ranges, and service areas, because those are the facts businesses most often leave unstated or state only in an image, a PDF, or a booking widget a crawler cannot read.

The defence is boring consistency: one name, address, and phone everywhere; current data on your site and profiles; enough authoritative coverage that the retrieval step finds truth instead of guessing. State the facts you want repeated in plain text — hours, service radius, what you do and explicitly do not do — because an unstated fact is an invitation to guess.

The failure is also asymmetric, which is the argument for checking rather than assuming. You will hear about a wrong phone number when a customer finally reaches you. You will never hear about the answer that quietly sent someone else the job. Measuring what assistants currently claim about you is step one — the free report returns their actual answers, with the sources they leaned on.