How ChatGPT decides which local business to recommend

Ask ChatGPT, or Google’s AI Mode, or Perplexity, “who is a good carpet cleaner in Mill Valley” and you get a short answer naming two or three businesses, with a sentence on each about why. No list of ten links. No map. A recommendation, in prose, that most people will take.

Where do those names come from? That is the question behind “answer engine optimisation”, or AEO, and the honest answer is more ordinary than the term suggests.

The answer is assembled, not ranked

A language model does not have a ranked index of plumbers. When it is asked a local question, it goes and looks: it runs searches, reads the results, reads listings and review sites and the businesses’ own pages, and then writes a summary of what it found. The businesses that appear are the ones that were easy to find, easy to describe, and consistently described across the sources it read.

That means the inputs are mostly the same inputs that decide the Google map pack. A strong Business Profile, reviews with real content, one consistent name and address everywhere, and a website that says plainly what the business does and where. The weighting is different and nobody outside these companies knows it precisely, but the raw material is the same.

What the model is reading

In practice, for a local service question, the sources that show up again and again are:

• The Google Business Profile, or the Maps data derived from it: category, address, hours, rating, review count
• Yelp, and to a lesser degree the other directories, because they have a page for nearly every business and the pages are structured
• The business’s own website, in particular whichever page describes the service asked about
• Review text, not just the star rating, because the model is looking for something to say about each business
• Local press, chamber of commerce listings, and anything else that mentions the business by name in context

A business that is thin or contradictory across those is a bad candidate for a confident answer, and the model will pick a different one.

Consistency is most of the game

The most common reason a good business is missing from an AI answer is not that it lacks authority. It is that the sources disagree with each other. Three versions of the name. An old phone number on Yelp. A category on the profile that does not match what the website says. A service area that is listed differently in three places.

A model trying to say “Burrous Brothers is a carpet and rug cleaner in Mill Valley with strong reviews” needs every source it reads to agree with that sentence. The citation clean-up that local SEO has always required is, it turns out, the foundation of showing up in AI answers too.

Reviews as text

Star ratings get a business onto a shortlist. Review content gets it recommended. The model is reading the reviews to find something to say, so “great service” a hundred times is less useful than twenty reviews that mention the actual service, the town, and what happened.

You cannot write your customers’ reviews, and you should not try. But the way you ask shapes what comes back. Asking right after the job, with a mention of what was done, gets reviews that describe the work. That is good for the customer reading them, and it is what the model is looking for.

Pages that answer the question

The website’s role changes a little here. For the map pack, the site mostly needs to match the profile. For AI answers, it also needs to contain sentences that a model can lift. A page that says, in plain words, what you do, where you do it, who it is for, and what to expect, gives the model its summary ready-made. A page that is a wall of “premium solutions” gives it nothing.

This is the one place where AEO asks for something that older SEO advice did not: write the page as if answering the question a person would actually ask, in the order they would ask it.

For software companies the same logic holds at a different scale. The model is looking for a clear claim about what category the product is in and what job it does. That is the same problem B2B SaaS SEO has always been about, and the sites that struggle in search for lack of a clear claim struggle in AI answers for the same reason.

What I do not know

The weighting. Nobody outside OpenAI, Google and Perplexity knows how much a review counts against a citation against a page, and anyone selling you a specific formula is guessing. The systems also change quickly.

What I do know is that the work that helps is the same work that helps in the map pack, done with slightly more attention to whether the sources say the same thing and whether the pages contain plain sentences. So I do not sell AEO as a separate engagement. It is part of local SEO, measured the same way, with one addition.

Measuring it

Once a month, I ask the assistants the questions your customers ask. “Who is a good tile cleaner in Marin.” “Emergency plumber in the Sunset.” Five to ten questions, the same ones each month, and I record who is named and what is said about them. It is manual and it is imprecise, but it is the only direct measurement available, and over six months it shows movement.

Alongside that, leads. If the phone rings and the caller says “ChatGPT recommended you”, that goes in the count, and it has started happening.

If you are weighing up who to hire for any of this, this piece is the checklist I would use.