How does ChatGPT choose which businesses to recommend?

If you want ChatGPT to name your business, you need to understand how ChatGPT arrives at an answer. The process is less mysterious than it looks, and it explains exactly why some businesses get named again and again while others never do.

Step 1: the question becomes search queries

When you ask ChatGPT "who's a good accountant for a restaurant in Alexandria," the model first decides whether it needs information from the web. For questions about businesses, it almost always does. It then writes one to five search queries, for example "restaurant accountant Alexandria," "best accounting firm for restaurants Northern Virginia" and "restaurant accountant reviews." We call these search queries fan-outs.

Fan-outs are gold for GEO, because they show which search terms you need to own to make it into the answer. They often differ from the user's literal question: the model adds words like "best," "comparison," "reviews" or a year.

Step 2: the model reads a handful of pages

For each query, the model pulls a limited number of results, usually between five and ten, and reads their content. In practice, ChatGPT strongly prefers pages that put several options side by side: comparison sites, "best X in Y" lists, trade media, industry associations and forum threads. Your own website gets read if it ranks well for the fan-out, but it's rarely the only source.

Step 3: recognizing and weighing entities

From the pages it reads, the model extracts business names (entities) and the context around them: specialty, location, price range, strengths and weaknesses. A business that appears in three of the five sources it read, with consistent information, is almost certain to be named. A business that appears in one source, with a different name or without concrete details, often gets dropped.

Consistency carries a lot of weight. If your site calls you "Hollis Mechanical," your state business registration says "Hollis Mechanical Services LLC" and a comparison site lists "HMS Fairfax," the model may see three different businesses, and name none of them.

Step 4: the answer, and why it varies

The model writes an answer with the three to five most convincing options, often with a short reason for each and links to its sources. Because the fan-outs and the chosen sources differ slightly every time, so does the answer. Ask the same question three times and you get three versions, in which your business might appear once, twice or three times.

That's why "I tried it myself" isn't a measurement. A reliable statement about your visibility requires running the same prompts several times a day, over several days and across several models, and reporting your share of mentions as a probability.

What to do with this

• Find the fan-outs for your most important customer questions and check whether you show up for those terms.
• Make sure you're listed in the sources the model reads, often three to five sites per topic.
• Make your business name, address and specialty identical everywhere.
• Write pages that answer a question with concrete information: prices, lead times, terms.
• Track repeatedly, per model, and steer on the trend.