How do AI assistants decide which local business to recommend?

Retrieval

The step where an assistant gathers material about your area before it writes anything. If you are not in what it gathers, you cannot be in what it writes.

Grounding

Tying the answer to retrieved sources rather than to the model's memory. Local recommendations are almost always grounded, because a model's own memory of a small business goes stale fast.

Entity consistency

Every source describing the same business the same way: name, address, phone, hours, categories. It is what lets a model treat scattered mentions as one business rather than three uncertain ones.

What the assistant is actually working with

For a local question, an assistant rarely relies on memory. It goes and looks: a search index, maps and directory data, review platforms, local publishers, and your own pages. Then it composes a short answer out of whatever came back and agreed with itself.

Ask "best HVAC company near Austin" and the businesses that survive into the answer are usually the ones that appeared in several of those sources at once, described identically in each. Not the ones with the best homepage.

What survives into the answer

Businesses that several independent sources describe the same way. Contradictions do not average out; they make you ambiguous, and an ambiguous business gets dropped rather than argued about. That is the whole finding, and it is duller than the tactics people expect.

The corollary is uncomfortable for anyone who has spent a decade on their own website. Corroboration you do not own (what directories, review platforms, local press and suppliers say about you) carries more weight than anything you write about yourself. Your site settles ties. It rarely wins them.

Which of those sources is actually being leaned on differs by city, by category and by engine, and it changes. That is the part worth measuring rather than assuming, because the answer for a plumber in one metro is not the answer for the same trade three states over.

What does not appear to move it

• Keyword stuffing. Models read meaning, and a padded page mostly succeeds at signalling that the page is padded.
• Dozens of thin location pages. They dilute the entity rather than multiply it.
• Paying for placement. There is no ad slot inside most answers today, and anyone offering to buy you one is selling something else.
• Bought reviews. Platforms filter them, assistants inherit the filtering, and if you are caught the damage lands on the source of truth you most depend on.

Why more than one engine

ChatGPT, Gemini, Perplexity and Claude do not retrieve from the same places and do not agree with each other. It is entirely normal for a business to be named in most answers on one assistant and in none on another. An average across them hides precisely the fact you needed, which is why Apex reports every engine separately and always shows the counts.

Where to start

Fix the contradictions first. They are cheap to fix and they are the single most likely thing keeping you out of the retrieval set. Then measure, so that the next change you make is a decision rather than a guess.

A free Apex scan puts real customer questions to every engine we measure and comes back with your AI Visibility Score and its counts, the questions we asked, and the businesses being named instead of you. That last list is usually where the work becomes obvious.