How do AI assistants choose which businesses to recommend?

The rough process

When someone asks “who’s a good accountant for a small business in Edmonton?”, an assistant with search will usually turn that into one or more web searches, pull a set of pages, and summarise what it finds. Some lean on their own index, some on third-party search providers (Microsoft’s Copilot uses Bing), and Google’s products draw on Google’s index and local data.

The model then writes an answer. It appears to favour businesses that appear in several of the sources it found, and that clearly match the details in the question: the service, the place, any specifics like “weekend hours” or “speaks Punjabi”.

Signals that seem to matter

• Clear relevance. Your site says plainly that you offer that service, in that area, for that kind of client.
• Consistency. Your name, location and services match across your site, Google Business Profile and directories.
• Reviews. Volume and what people actually say, especially on widely used platforms.
• Third-party mentions. Local news, association listings and “best of” lists give independent confirmation.
• Readable pages. Content in HTML that crawlers can access.

Why answers vary

Ask the same question twice and you may get different businesses. Wording, the user’s location, whether search was triggered, what the search returned that moment and the model’s own randomness all play a part. Assistants can also be wrong: naming a business that has closed or mixing up details.

That variability is why there’s no stable “AI ranking” to buy or guarantee. The best you can do is be the clearest, best-evidenced match for the questions your customers ask, across all the sources these tools are likely to read.

I treat it like being referred by a well-read stranger. They’ll only mention you if they’ve come across you in several places and the details line up. So I make sure the website states the specifics, the Business Profile matches, and the client is listed where their industry is listed.

Then I test real customer questions in ChatGPT, Gemini and Perplexity and note which sources they cite. Those sources are the to-do list.