How Do AI Chatbots Choose Which Businesses to Recommend?

AI assistants choose businesses to recommend in three steps: they read your prompt, optionally retrieve live web pages, then synthesize an answer from those pages plus training data. There is no ranking algorithm to game and no paid placement to buy. The businesses that get named are the ones described consistently across the sources the model reads.

The mechanism: prompt, retrieval, synthesis

Strip away the interface and every AI recommendation works the same way.

Step 1: the prompt. A customer types something like “who’s a good mortgage broker in Frisco for first-time buyers?” The wording matters — intent, city, and qualifiers all shape what happens next.

Step 2: retrieval (sometimes). If the system has search enabled — Perplexity always, ChatGPT search, Google’s AI answers — it runs a web search and reads a handful of the pages that come back: directories, review platforms, best-of lists, local press. If search is off, the model skips this step and answers purely from training data, a static snapshot of the web that is months old.

Step 3: synthesis. The model writes an answer from what it just read plus what it already “knows.” It names the businesses that the retrieved pages describe most clearly and most consistently for that intent.

That is the whole pipeline. No auction, no index of businesses with scores attached, no submission queue. Which means the leverage is entirely in step 2: what the pages say about you when a machine reads them.

It also means intent matters more than most owners expect. A discovery question (“plumbers in Mesa”), a recommendation question (“who should I call for a burst pipe”), a comparison question (“X vs Y, who’s better for repipes”), and a task-specific question (“who can replace a water heater this week”) each retrieve different pages and reward different evidence. A business can be visible for one intent and invisible for another. That is why our battery tests four intents separately instead of one generic query.

Why consistency beats any single profile

A search engine can rank you off one strong page. A language model synthesizing an answer behaves differently: it is aggregating evidence across several sources at once, and it is sensitive to agreement.

If your directory listing says “Smith & Co. Plumbing,” your review-platform profile says “Smith and Company,” and your site footer shows an old phone number, the model is reading three weak, partially conflicting descriptions instead of one strong one. Consistent name, address, and phone across sources — plus a complete Google Business Profile — is the boring foundation that shows up again and again in the audits we run. In our local-business audits, the pattern that predicts visibility best is presence on the sources AI answers already cite, plus exactly that consistency.

This is also why “I have a great website” is not sufficient. The comparison pages the model retrieves for “best X in Y” are rarely your website. Which domains those actually are in your market is a knowable fact — we cover how to find them in what sources ChatGPT uses for local recommendations.

Prominence: being named first

Not every mention is equal. When an AI lists five plumbers, the first name carries the recommendation; the fifth is an afterthought. Customers read AI answers top-down, the same way they read anything else.

That is why our Visibility Score does not just count mentions. It weights mention rate at 50%, prominence — where in the answer you appear — at 20%, sentiment and accuracy at 15%, and source presence at 15%. Two businesses can both “appear in ChatGPT” while one is the lead recommendation and the other is a trailing also-ran. The score separates them.

Prominence tends to follow the same inputs as mention rate, concentrated: the business that best-of lists rank first, that has the deepest review evidence, and that local press actually covers, tends to get named first in synthesized answers too.

Why answers change run to run

Ask the same question twice and you may get different businesses. Three reasons:

• Sampling. Models generate text probabilistically. Close calls between similarly-described businesses can go either way on any given run.
• Retrieval variance. Search results shift day to day, and the model reads only a handful of pages per answer. Different pages in, different names out.
• Phrasing sensitivity. “Best dentist in Raleigh,” “top-rated dentist Raleigh,” and “who should I see for a crown in Raleigh” retrieve and emphasize different pages.

None of this is malfunction. It is how generative systems work, and it is not going away. It also has a direct consequence for anyone trying to measure their AI visibility: variance is the default, so a single observation tells you close to nothing.

What it means for measurement

One screenshot of ChatGPT naming you (or not naming you) proves almost nothing. It is one draw from a distribution.

Measuring honestly means sampling. Our audit battery runs 20 localized customer questions — four intents (discovery, recommendation, comparison, task-specific) — against six AI systems: GPT-, Claude-, Gemini-, DeepSeek-, and Qwen-class models plus Google’s answer layer. That is 100+ timestamped answers per audit, each stored raw and quoted verbatim. We never edit, paraphrase, or fabricate a model output.

From that sample you get a stable number: how often you are named, where in the answer, with what sentiment, and which sources the answers drew on. Re-run the same battery monthly and you can see whether the presence work is moving the number. The full method is on how it works.

No, you cannot pay for placement

There is no advertising product that buys a spot in organic AI chat answers. Not from OpenAI, not from Google’s AI answers, not from Anthropic, not from Perplexity. Anyone selling “guaranteed AI mentions” is selling a guess.

What can be bought is labor: the work of getting present on the sources AI systems retrieve, and the measurement to verify whether it worked. That is what agencies charge for — entry-level AEO/GEO agency programs run $1,000–$2,500 per month — and it is what our audits measure. If you are being pitched something that sounds like a placement fee, ask the vendor which mechanism the money buys. There isn’t one.

What to do this week

1. Ask ChatGPT (with search on), Perplexity, and Google the same question a customer would ask about your category and city. Do it three times each. Note who gets named, in what order.
2. Write down every domain those answers cite. That list is your market’s source map.
3. Check your presence on each cited domain. Claim missing profiles, complete thin ones.
4. Standardize your name, address, and phone everywhere — site footer, directories, review platforms, Google Business Profile.
5. Ask five recent customers for reviews on the platform your market’s answers cite most.
6. Put a monthly reminder in your calendar to re-run step 1 — or let an audit do the sampling at proper scale.

If the systems name your competitors and skip you entirely, that has specific, diagnosable causes — see why doesn’t ChatGPT recommend my business.

The stakes, stated plainly

This channel is no longer fringe: 29% of adults now initiate daily searches via generative AI summaries, and among Gen Z adults, 82% have used AI chatbots. At the same time, about 84% of US consumers still use Google daily for local business searches, and only around 24% prefer AI chat for local queries. Both facts are true. The practical position is to be visible in both channels — especially since the distinction is collapsing as Google’s own answers become AI-composed. The same sources feed both.

To see who the AI systems actually name in your market right now, run a free AI visibility scan — three questions across two models plus Google, results emailed in about 30 seconds.