How does ChatGPT choose which local businesses to recommend?

Nobody outside OpenAI can see exactly how ChatGPT chooses which local businesses to recommend. There’s no published formula, and I don’t have one. I observe it instead: I ask the questions and record the answers and the sources, then look for patterns. Those patterns have limits.

Two ways it can answer

Sometimes ChatGPT answers from training. The model draws on the text it learned from, which stops at a cutoff date, so this is recall, not lookup. A large national brand appears in that text many times. A twelve-person electrical company in La Mesa may barely appear, and what is there may be out of date: an old phone number, or a service the company no longer offers.

Other times it searches. At the time of the question it runs web searches and reads some of the pages that come back, then writes an answer from them, usually with links to its sources. Here a small business has a fair chance, because the answer is built from what can be found today.

Which of the two it uses depends on the question, the version and the user’s settings, and that changes as the product changes. For a local business the difference is large.

What can be observed

When it searches, it shows its sources, and that is the most useful fact in this whole subject. For “who should I call” questions, those sources are ordinary web pages: business websites, directories, review sites and local “best of” lists. A business missing from the pages it read can’t appear in the answer. A business described the same way on several of them is a much easier one for it to name.

Repeated testing also shows that answers vary: ask twice and the names can change. The place in the question matters too. “Electrician in La Mesa” and “electrician near Grossmont” can return different companies. The reasons it gives (“well reviewed”, “offers same-day service”) come from what those pages say, so it repeats claims it can find and leaves out the ones it can’t.

What I won’t claim

Nobody outside the company has a list of “ChatGPT ranking factors” with weights, so I don’t offer one. I can’t tell you how it orders the pages it reads, how much a review counts against a directory listing, or what the next version will do differently. Anyone presenting those as certainties is guessing. And nobody can guarantee that ChatGPT will recommend a particular business.

What a business can do about it

Work on what it can find. In practice:

• Make the facts consistent: the same name, phone number, services and service area on your site, your Google Business Profile and every directory.
• State plainly on your own pages what you do and where, with one service or one question per page.
• Treat reviews and citations as AI signals: earn reviews in more than one place, with customers describing the job.
• Get listed, accurately, on the specific sources that appear when your questions are asked. The test shows which ones those are.

This is GEO (Generative Engine Optimization): the work of being named and cited by generative AI. It overlaps almost entirely with sound local SEO, so most of it isn’t new work.

How I measure it

I write a fixed set of questions your customers would ask, run them repeatedly, and record which businesses are named and which sources are cited. Then I change one thing and run them again about two weeks later. If the names and sources don’t change, I undo it and try something else.

That’s the method behind my ChatGPT visibility work, and the ongoing version of the same test is AI visibility tracking.