How AI assistants decide which businesses to recommend

It is worth understanding the pipeline before trying to influence it, because most of the advice circulating about AI visibility targets the wrong stage.

Here is what happens between the question and the answer.

Stage one: the assistant decides whether to search

Not every question triggers retrieval. The model makes a judgement about whether its internal knowledge is sufficient.

Questions about stable facts usually do not trigger a search. Questions with location words, time words, superlatives about a specific place, or anything the model treats as volatile usually do.

"Best neighbourhood cafés in Lisbon for working" almost certainly triggers retrieval. "What is a flat white" almost certainly does not.

This matters because if the question your customers ask does not trigger retrieval, you are competing on training data alone, and your leverage there is slow and indirect.

Stage two: query reformulation

The assistant rarely searches for the literal question. It rewrites it, often into several queries at once.

A question about quiet cafés with good wifi might become searches for the city plus "best cafes for working," the city plus "laptop friendly coffee," and the city plus a neighbourhood name. Each returns a different result set.

The practical implication: you are not optimising for one phrase. You are optimising for the family of phrases a model would generate from a customer's real question. Listing the actual attributes people care about, in plain words, on a page a search engine can rank, covers far more of that family than trying to guess an exact keyword.

Stage three: retrieval

The searches return results and the assistant reads some of them. Usually the top handful per query, sometimes only the snippets.

What gets read is ordinary web content. Your site if it ranks. Directories. Review platforms. City guides. Local press. Google's own index, when the assistant is Gemini or an AI Overview. Reddit and forum threads, which appear more often than most businesses expect.

Stage four: entity extraction and filtering

From the retrieved text, the model pulls out candidate businesses and whatever attributes it can attach to each: neighbourhood, price band, opening hours, what it is known for.

Then it filters against the constraints in the original question. Open on Sunday. Walking distance from a landmark. Under a certain price. Suitable for a group of eight.

This is the stage where most businesses lose, and they lose quietly. Not because the model dislikes them, but because the constraint could not be verified. If nothing in the retrieved text says you are open on Sunday, you get dropped from a Sunday query even when you are open.

Every attribute a customer might filter on is an attribute worth stating explicitly somewhere crawlable.

Stage five: corroboration weighting

When several candidates survive filtering, agreement across sources acts as a tiebreaker.

A business described the same way by four independent sources reads as a safe recommendation. A business described one way on its own site and differently on a directory reads as uncertain, and models tend to route around uncertainty rather than resolve it.

This is the single most under-appreciated lever in the whole pipeline. Fixing contradictory information across your listings is unglamorous, cheap, and moves more than most content work.

Stage six: composition

The model writes the answer, usually naming three to five places with a line each.

That line is assembled from the descriptive language in the retrieved sources. If your reviews consistently mention the courtyard, the courtyard shows up. If nobody has ever written down what makes you distinct, the model writes something generic, and generic entries get cut when the answer needs to be shorter.

What this means you should work on

Reading back up the pipeline, the leverage points in rough order of value:

Be retrievable. A crawlable site with real text about what you are. This is table stakes and a surprising number of hospitality sites fail it because everything is inside an image or a booking widget.

Be filterable. State every attribute a customer would constrain on, in plain language and in structured data. Hours, price band, capacity, accessibility, whether you take bookings, what you are suited to.

Be corroborated. Get the same facts stated accurately on three or four independent sources, and remove contradictions from the ones that already exist.

Be describable. Give people, including reviewers and journalists, the specific detail that distinguishes you. The distinctive phrase in your reviews is the phrase that survives into the answer.

Be current. Stale third-party listings are what keep closed businesses in AI recommendations. The same mechanism keeps your old opening hours there too.

None of this requires guessing at a ranking algorithm. It requires making a business easy to find, easy to classify, and easy to describe accurately, which is the same thing every discovery system has ever rewarded.