Why does an AI recommend one business over another?

The uncomfortable version

Right now, for most local businesses, the model is guessing. Not randomly — it is drawing on real sources — but it is assembling a recommendation about your business out of pages you did not write, from data you did not check, with no way to ask you directly.

We can show you what that looks like. Ask a model about one of the businesses we work with by name and it answers accurately and warmly. Look at where it sourced that answer, though, and it is citing Top Rated Local, HomeAdvisor and the Better Business Bureau. Directories. Not the business's own record — because for most businesses there isn't one it can read.

Paid placement does not exist here

This is the part that catches advertising-heavy trades off guard. There is no sponsored slot inside an AI answer. No budget buys a mention, no agency has a relationship that places you there, and the auction that has protected some businesses for twenty years buys precisely nothing in that conversation.

For a trade paying two hundred dollars a click, that is either alarming or the best news in a decade, depending on whether there is anything of yours for the model to read.

What a model treats as trustworthy

Broadly: sources that are specific, structured, consistent with each other, and attributable. A page that says "we are the best plumber in town" is worthless to it. Two hundred dated, attributed customer reviews that independently mention the same technician by name and the same problem being solved are not.

This is why review data is unusually well suited to the job. It is high volume, written by third parties, timestamped, and impossible to fake at scale in a way that stays consistent.

Making yourself readable

There are three practical layers, and most businesses have none of them.

The first is structured data — your services, your location, your hours, your aggregate rating and your individual reviews marked up in schema.org types specific to your trade, so a machine reads facts instead of parsing prose.

The second is a queryable endpoint — an MCP server and an ai-plugin manifest, so an assistant can ask your record a question directly rather than hoping it scraped the right page.

The third is being in the indexes — submitted, crawled and retrievable, because a record nothing can find is a record that does not exist.