AI hallucination

Where the wrong answer comes from

An AI assistant does not look your business up the way a person does. It builds an answer out of patterns in everything it has read, and it fills small gaps with wording that fits. Most of the time the gap filling is harmless. Sometimes it invents a fact.

Say a customer asks an assistant whether a tree service handles Sunday emergencies. If nothing the assistant has read settles the question, it may answer anyway, and it may answer yes, because yes is the answer that fits the pattern of what it has read. The sentence reads like a quote from the owner. It never came from the owner.

A hallucination is not the same as an out of date fact. Out of date means the system found something that was true last year. A hallucination means the detail was never true anywhere, on any page, for any business.

Gaps are what feed it: two different phone numbers on two directories, an address you moved out of three years ago, a services page that never quite says what you do. Every gap is an invitation for the system to fill in the blank on its own.

Two flavors matter to an owner. The first is an invented detail about your company, such as a service you dropped years ago. The second is an invented general claim that a customer arrives already believing, such as a typical price for the kind of work you do. The second one is worse in person, because you end up arguing at the door with a number nobody at your shop ever quoted.

A confident wrong answer costs you the call

The customer has no way to tell an invented answer from a checked one. They act on it. They drive to a closed door, or they cross you off because a chatbot told them you do not cover their town. You never see that lead, and nothing bounces back to tell you it happened.

You cannot patch a model, and no platform owes your business a correction. What you can control is the supply of facts: state your hours, your service area, your services and your phone number plainly on your own site, keep those details identical everywhere else you are listed, and write them as sentences instead of burying them in an image or a PDF. That is the practical defense, and it is what AI search optimization work is built around.

Direct beats clever here. A sentence that says the shop answers emergency calls seven days a week in three named counties is hard to get wrong. A headline that says always here when you need us gives a machine nothing to repeat, so it repeats a guess instead.

It also helps to keep the answers people get straight from you under your own control. A scripted chat concierge we built for a Marco Island boat tour company answers guest questions across roughly 500 pages using owner approved answers only, so nothing gets improvised on the spot.

Terms that sit next to this one

Grounding is the countermeasure: tying an answer to a source the system just retrieved instead of letting it recall from memory. Retrieval augmented generation is the usual way that gets done. When an answer names where it came from, that is an AI citation, and a cited claim is far easier to check than a bare paragraph inside AI Overviews. None of these remove the risk. They shrink the space where a guess can hide.