Hallucination
A hallucination is a statement an AI assistant presents as fact that is not supported by its training data or by anything it retrieved. The sentence is fluent and confident; the thing it describes is not there.
Why a model invents
A language model produces the words most likely to follow, given everything before them. Asked a question whose answer it has not seen, it does not stop; it continues in the shape an answer would have. A plausible opening time, a plausible price, a plausible street. Retrieval changes this, because the model is then writing from pages it fetched a moment ago and can cite them, and a reader can check the citation against the page. It does not remove the failure. If the fetched pages disagree, or say nothing on the point, the model still has to write a sentence, and the sentence can still be wrong.
When the subject is a business
The business is the party with the least room to be misdescribed and the most control over the cause. An assistant asked about a company it can find little on will fill the gap; an assistant that finds three different phone numbers on the company’s own pages has three candidates and no way to choose. Across 9,483 small-business sites this site audited in September 2026, 43.1% were inconsistent about their own name, address or phone number across their three main pages, and 36.9% published more than one phone number. Each of those is an invitation to a confident wrong answer that the business wrote itself. The remedy is on the page: one identity, repeated exactly, in text and in schema, with the facts a customer asks about stated once where a machine can attribute them. Aggarwal et al. found in “GEO: Generative Engine Optimization” (KDD 2024) that what is on the source page measurably changes what the generated answer does with it. The audit figures are in what 9,483 small business websites serve to an AI crawler.
A common question
Why does an inconsistent phone number invite a wrong answer about a business?
An assistant that finds several different numbers across a company's own pages has no basis for choosing between them, and it still has to write a sentence. In the audit of 9,483 small-business sites, 43.1% were inconsistent about their name, address or phone number across their three main pages, and 36.9% published more than one number. The remedy sits on the page: one identity, repeated exactly in the text and in the schema, with the facts customers ask about stated once where a machine can attribute them.
Related terms
Sources
Source-side effects on generated answers: Aggarwal et al., “GEO: Generative Engine Optimization” (KDD 2024). The 43.1% and 36.9% figures are this site’s own; sample drawn per metro and trade, percentages describe the sample.