AI recommendation

An AI recommendation is an answer in which an assistant names a specific business, product or provider as a suitable choice for the person asking. It is stronger than a citation, which credits a page for a fact, and stronger than a mention, which only notes that the business exists.

From a question to a name

The person asks something with a decision in it: who to call, where to go, which one to buy. The assistant retrieves pages, reads them, and answers with a short list of names and a reason for each. When this site put one such question per business to Claude with web retrieval for 9,240 small businesses in September 2026, 14.9% of the businesses were named; among the 5,928 answers that gave a differentiated list rather than a hedge, 17.7%. Directional, one assistant, one question, one day. The part worth dwelling on is who else was in those lists: of every name returned, 99.4% were businesses and 0.6% were directories or aggregators. A results page for the same questions is largely aggregators. The account is in what 9,483 small business websites serve to an AI crawler.

What the assistant has to establish first

Three things, and it establishes them from the site and from what others say about the site. That the business exists and is one business: 43.1% of the 9,483 audited sites were inconsistent about name, address or phone across their own three main pages, 36.9% published more than one phone number, and 34.1% had no postal address a machine could parse. That someone vouches for it: 64.7% showed no review signal a crawler could read, and 8.9% carried rating markup. And that the site answers the question being asked, which 61.5% did not attempt in any question-and-answer form. An assistant deciding whether to name a business is choosing among candidates it can describe; a business it cannot pin down is not a candidate. Aggarwal et al., “GEO: Generative Engine Optimization” (KDD 2024), is the controlled evidence that the source side of this can be moved.

A common question

Why would an assistant skip a business that clearly does the work being asked about?

Because an assistant can only name candidates it is able to describe. Before naming one it has to establish that the business exists as a single business, that someone vouches for it, and that the site addresses the question being asked. That is where the audited sites fell down: 43.1% were inconsistent about name, address or phone across their three main pages, 64.7% showed no review signal a crawler could read, and 34.1% had no postal address a machine could parse. A business that cannot be pinned down is not a candidate at all.

Sources

Mechanism: Aggarwal et al., “GEO: Generative Engine Optimization” (KDD 2024). The probe and audit figures are this site’s own, computed from its audit records; sample drawn per metro and trade, not from a random frame, and the probe is a Claude number.