How AI Chooses Which Businesses to Recommend: The Mechanism, Step by Step

Three roofers named, and nobody came in fourth

The roofer who was left out will want to know his rank. He does not have one. A search results page is an ordered list that holds still long enough to measure. An AI answer is a paragraph written on demand, and it either contains your name or it does not. Ask the same question again in a fresh chat and the paragraph may name a different set of companies. We made that argument at length in there is no number one in an AI answer; this page explains the machinery that produces the variation, because the machinery is where an owner gets some say.

Every major assistant follows roughly the same pattern when a question needs current, local or commercial facts. First it searches. Then it reads what the search returned. Then it writes. Each engine uses different search plumbing and different local data, so the details below differ by engine, but no engine can name a business it never retrieved. That single fact explains most of what follows.

A model cannot recommend a business it never found. Most of GEO is making sure it finds you, and likes what it reads.

Step one: the question becomes a handful of searches

A customer types one sentence. The engine rarely searches for that sentence alone. Google says AI Overviews and AI Mode may use query fan-out, issuing several related searches across subtopics and data sources at once. A request for a hail-damage roofer near Boerne might quietly become searches for roofers in Boerne, roofers that work with insurance claims, roofing reviews in Kendall County and so on. Each search is a separate chance to be found, or missed.

Which index gets searched depends on the engine. That is why a business can be visible in one assistant and absent from another. The table summarizes what each engine has published or what credible testing has found; the engine pages carry the full detail.

Step two: retrieval builds the shortlist you have to be on

The searches return pages: business websites, directory listings, review pages, news stories, forum threads. That pile is the shortlist. In BrightLocal's November 2024 study of 800 local searches in ChatGPT, 58% of sources were business websites, 27% third-party mentions and 15% directories, and Yelp, Facebook and Google Maps did not appear as directory sources at all. Different engines, different piles: Gemini leans on Google Maps, Perplexity on licensed Yelp data, Siri on Apple Maps place cards.

This is the step where technical problems quietly disqualify a business. If robots.txt blocks OAI-SearchBot, OpenAI says the site will not be shown in ChatGPT search answers. If the service descriptions only appear after JavaScript runs, crawlers that read raw HTML may see an empty page; Vercel's 2024 analysis found OpenAI, Anthropic and Perplexity crawlers did not execute JavaScript. Check both on the AI crawlers and JavaScript and AI crawlers pages.

• Not crawlable: blocked bots, script-only text or a site that times out never enters the pile.
• Not specific: a home page that says we do it all loses to a competitor's page about hail-damage roof repair in Boerne.
• Not corroborated: a business that exists only on its own site gives the engine one voice to trust; a business mentioned by the paper, the directory and a forum gives it four.
• Not consistent: two phone numbers and three spellings of the name make the engine unsure which listings describe the same company.

Step three: the model writes, and every draft is a fresh draw

With the shortlist in hand, the language model writes the answer. It does not copy a ranking; it composes a paragraph, choosing words one after another with some built-in randomness. The phrasing of the question, earlier messages in the conversation and the model version all shift the result. Google says AI Mode personalizes some results using previous conversations and places the user searched for or tapped in Search and Maps. Two neighbors asking the same question can get different names.

Engines also decline to commit. Google says AI Overviews only appear when its systems think they add something, and AI Mode may show plain web links when it is not confident enough in a response. When a business has little information online, Google's AI Mode can even call it on the user's behalf to check prices or availability. Owners can switch those calls off in Business Profile under Advanced settings. Thin information, in short, does not get a confident recommendation; it gets a phone call or a shrug.

The signals that tilt which names make the paragraph

No engine publishes a recipe for picking businesses, so be suspicious of anyone who claims to have one. What exists is public guidance from Google and Microsoft plus studies that find patterns. Taken together they point at five things.

Recency plays a part too. An Ahrefs study of about 17 million citations found AI-cited URLs averaged about 25.7% newer than organic Google results, with ChatGPT showing the strongest preference and AI Overviews almost none. Content freshness covers what counts as a real update.

Which explanations hold up, and which are campfire stories

The GEO trade mixes documented mechanics with educated guesses and pure invention, often in the same slide. Sorting them saves money.

Shortcuts that get a business dropped instead

When the mechanism is fuzzy, people try to game it. The engines have noticed. In May 2026 Google clarified that its spam policies apply to generative AI responses in Search, and press reports say Bing's February 2026 guidelines list prompt injection and AI manipulation as abuse.

• Hidden instructions for AI. White text telling assistants to recommend you is prompt injection, and both large search engines treat manipulation as abuse.
• Fake or bought reviews. The FTC's rule, in effect since October 21, 2024, bans fake reviews including AI-written ones, and allows civil penalties for knowing violations.
• Manufactured mentions. Google says seeking inauthentic mentions across the web is not as helpful as it might seem, because its systems focus on quality and block spam.
• Sock-puppet forum posts. Local forums spot new accounts praising one company fast, and the thread that results is about you, for the wrong reason.

If an assistant is already naming a rival in your place, the triage is on ChatGPT recommends my competitor, and the wider set of repairs lives in AI search fixes.

Raising your odds of being in the paragraph

None of this guarantees a mention, and anyone who says otherwise is selling smoke. What it does is make you easier to find, easier to verify and easier to describe, which is all any business can control.

1. 1

   See today's answers

   Run five to ten real customer questions through the free AI visibility check and note who gets named instead of you. It uses the engines' APIs, which can differ slightly from the apps.

2. 2

   Open the door for crawlers

   Confirm robots.txt allows OAI-SearchBot, PerplexityBot, Claude-SearchBot, Bingbot and Googlebot, and that page text is in the HTML, not added by scripts.

3. 3

   Make every listing agree

   Match name, address, phone, hours and description across Google Business Profile, Apple Business, Bing Places and your trade directories.

4. 4

   Publish the specifics people ask for

   One page per service and area, with prices or ranges if you are willing, licenses, warranties and the questions customers actually ask, answered in the first lines.

5. 5

   Earn reviews and reply to them

   Ask every customer, never only the happy ones, and answer each review in plain language.

6. 6

   Get mentioned where engines read

   Local news, associations, supplier locators and community threads, earned honestly. Then recheck the answers monthly.

That is the work our programs do, tracked with AICiteBeam and kept current by the agents in Athena CMS. It starts with a free GEO audit that comes back with a fixed monthly quote; AI recommendation optimization describes the service. More mechanics are on the GEO section front page.