How do AI assistants choose businesses?

The four steps

Bay Area AEO explains the process in four steps, and the same four words appear in its query fan-out explainer: understand, fan out, retrieve, answer. They describe how retrieval-augmented assistants work generally. The model reads the request, plans a set of searches, reads what comes back, and writes from what it has just read.

Understand. The assistant works out what is being asked: the kind of business, the place, and every condition attached. A question like “My water heater is leaking. Is there a plumber in Hayward who can come today?” carries a service (water heater repair), a city (Hayward) and a constraint (today).

Fan out. It then turns that one question into several narrower searches. Plumbers in Hayward. Emergency or same-day plumbing near Hayward. Water heater repair reviews. Licensed plumbers in Alameda County. Each sub-search can reach a different kind of source.

Retrieve. The assistant reads what those searches return: business listings, the businesses' own service pages and review text, plus directories and local articles. This is the step where a business either supplies clear facts or leaves the assistant guessing.

Answer. Finally it combines the results. The names that survive are the ones the sources support from several directions at once. A plumber whose listing shows emergency hours, whose site has a water heater page, whose reviews mention same-day visits and whose license number matches state records fits every branch of the fan-out.

Which sources each platform leans on

The platforms share the four-step shape and differ in where they look. Google's Gemini and AI Overviews draw on Google's index and Google Business Profile data. Microsoft Copilot draws heavily on Bing's index and Bing Places. Perplexity searches the web for every question and shows numbered citations beside its answer, so a reader can see the pages it used. ChatGPT and Claude, when they search the web, compose their answers from the pages they retrieve. Voice assistants such as Siri draw on the same listing data the text assistants read.

That spread is why Bay Area AEO covers the listings and directories each platform relies on, along with the content, and why a business that tends only its Google profile can appear in one assistant's answer and be missing from another's.

What earns a mention

AI assistants recommend the businesses they understand. In Bay Area AEO's words, those are the ones with consistent listings, clear websites, specific reviews and mentions in sources AI trusts. Size matters less than legibility: a small practice with precise, matching facts on every platform can be easier to recommend than a larger one whose details conflict.

The five signals Bay Area AEO builds map onto the fan-out branches. Consistent listings let the assistant confirm place and hours. An answer-ready website answers the service branch. Structured data states services, service area and hours in a form machines read without inference. Specific reviews supply the trust branch with the customer's own words. Trusted mentions in local press and guides, or in directories, give outside confirmation. The entry on the five AEO signals treats each one.

Live information moves quickly

Because the retrieve step happens at the moment of asking, much of what an assistant reads is current. Change the hours on a Google Business Profile, correct a phone number on Bing Places or add a missing service category, and the next search that reaches that listing reads the corrected version. Bay Area AEO states this directly: listing and data fixes can start showing up in AI answers within a few weeks, because many assistants pull live information. Content and outside mentions are slower, since a page has to be written and indexed and a guide has to choose to mention the business. That difference in speed shapes the order in which the work is usually done.

Why answers vary from one asking to the next

The same question asked twice, or asked on two platforms, can produce different names. Each assistant plans its own sub-searches, reaches its own sources and writes a fresh answer every time. Bay Area AEO's own demonstration on its site carries the line that real answers vary by platform and change over time. This is the reason measurement in AEO uses a fixed list of questions run across every major platform each month: one answer is a snapshot, while a month-by-month panel shows the trend. See how AEO progress is measured.

What an owner can do with this

Read the fan-out from the customer's side. Write down the questions customers actually ask, including the conditions they attach: a neighborhood, an insurance plan, a time of day, an occasion. For each condition, ask where an assistant would confirm it. If the answer is “nowhere”, that is the next page to write, the next listing field to fill, or the next review topic to invite. A free AI visibility check from Bay Area AEO runs those questions for a business and shows who the assistants name today.

One habit helps more than any single fix: reading an assistant's answer the way the assistant wrote it. Each sentence about a recommended business usually traces back to something the assistant retrieved, such as a phrase from a review, a line from a service page or a detail from a listing. When a competitor is described as “open Sundays” or “good with older homes”, that phrase came from somewhere, and the business that wants the same mention can publish the same kind of fact about itself, accurately and in the same places.

See also
Query fan-out · ChatGPT vs Google AI Overviews · Entity consistency · The five AEO signals