How AI Answer Engines Decide Which Local Business to Recommend

Ask ChatGPT for a good physiotherapist, a reliable boiler engineer, or the best coffee within walking distance, and you do not get ten blue links to pick through. You get an answer: a short list of names, each with a sentence on why. For a local business that is a far harsher cut than page one of Google ever was. There is no position two that still gets a few clicks. Either the machine says your name, or you do not exist in that conversation at all. This post is about how that decision gets made, and the three things that actually change whether it goes your way.

The question changed, and most local SEO advice did not keep up

For twenty years the job was legible. You wanted to be in the top three of the Map Pack and on page one of the organic results, because that is where the clicks were. Rank was a number, the number had a position, and you could watch the position move. Nearly every local SEO playbook still in circulation is built around that idea.

An answer engine does not produce a ranked list of options for the searcher to evaluate. It produces a recommendation. When someone asks ChatGPT or Gemini for a good osteopath in their town, the output is one to five names, usually with a short justification each, and sometimes a citation or two. The searcher reads it the way they would read a friend's text back: as a vetted answer, not as a list to research further. That is the part that should worry you and motivate you in equal measure. The shortlist is shorter, and being on it is worth far more per slot.

You are no longer competing for a position on a page. You are competing to be one of the three names a machine is willing to put its credibility behind.

That reframing has a practical consequence. Optimising for “rank” assumes a stable, observable position you can nudge upward. Optimising for recommendation means making yourself the obvious, low-risk, well-evidenced choice, because the engine is not scoring you against a rubric, it is deciding whether it trusts you enough to say your name out loud.

What actually happens between the question and the answer

It helps to see the pipeline. When a local question lands, a modern answer engine rarely treats it as a single lookup. It interprets the intent, then fans the question out into several narrower sub-questions, gathers evidence for each one (from its training memory, from a live web search it runs on your behalf, and from structured sources like maps and business listings), and only then synthesises a single answer.

Two of those branches are about building the list (who even counts as an option here), and the rest are about vetting it (is this one any good, does it actually do the thing, will recommending it embarrass me). A business that shows up in the first stage but fails the vetting stage gets quietly dropped before the searcher ever sees a word. You were a candidate. You did not make the cut. There is no log file that tells you this happened.

Three levers actually move whether you get named

We have run a lot of these prompts, watched a lot of answers form, and traced a lot of citations back to their source. When you strip away the per-engine quirks, the same three factors keep deciding who gets named and who gets skipped. We think of them as levers because, unlike a ranking position, you can actually pull on each one directly.

Lever one: entity clarity, or can the machine tell exactly who you are

This is the unglamorous one, and it is the one that quietly eliminates most businesses. An answer engine is, underneath everything, a system trying to resolve “the thing the user means” to “a specific entity in the world”. If it cannot do that resolution cleanly and confidently, it will not gamble on you. It will reach for a business it is sure about instead.

Entity clarity breaks when your own facts contradict each other across the web. A suite number on the website that is missing from the Business Profile. A trading name on the directory that does not match the signage in your photos. An old phone number that still lives on three citation sites. To a person, these are trivial. To a machine deciding whether two records describe the same place, every mismatch is a reason to lower its confidence, and low confidence reads as “do not recommend”.

1. 1

   Fix the canonical facts first

   Pick the single correct version of your name, address, phone and opening hours, then make every surface match it character for character. Start with the three that carry the most weight: your own website, your Google Business Profile, and your largest industry directory. Consistency here is not a nicety, it is the signal that says “these records are all the same business”.
2. 2

   Make the page state plainly what you are

   Somewhere prominent, in words, say what you do, where you do it, and who for. A machine should not have to infer that you are a podiatry clinic in Leeds that treats sports injuries. It should be able to read it. Vague hero copy (“helping you feel your best”) is invisible to entity resolution.
3. 3

   Add the structured layer

   LocalBusiness structured data, with your address, geo, opening hours, and a sameAs list pointing to your verified profiles, gives the machine an unambiguous, machine-readable version of the facts it would otherwise have to guess at. It does not replace clean on-page facts, it confirms them.

Lever two: content extractability, or can it lift the answer straight off your page

Once you are a candidate the engine knows is real, the next question is whether you are easy to quote. Answer engines are, by temperament, lazy in a useful way: faced with two equally good businesses, they will favour the one whose page hands over the answer in a clean, liftable form, because that lets them respond with confidence and a citation they can stand behind.

Extractability is mostly about removing friction. The fact a searcher wants (do you treat running injuries, are you open on Sunday, what does an initial consultation cost, do you cover my postcode) should exist somewhere on your site as a plain sentence, not buried in a graphic, a downloadable price list, or a booking widget that only renders after three taps. If a machine has to execute JavaScript, parse an image, or open a PDF to find your answer, assume it will not bother.

A practical test: take the ten questions your front desk actually gets asked, and check that each has a clear, standalone answer in real text on your site. Not implied, not downloadable, not behind a widget. Just there, in a sentence a machine could copy and attribute to you without having to interpret anything. That single exercise does more for answer-engine visibility than a month of conventional content work.

Lever three: consensus, or does the rest of the web back you up

The first two levers get you into contention. Consensus usually decides it. An answer engine is acutely aware that it can be wrong, and that recommending a bad business has a reputational cost. So when it has several plausible names and must choose, it leans towards the one the wider web most clearly and consistently vouches for.

Consensus is not one number. It is the agreement across independent sources: a healthy volume of recent reviews that say roughly the same positive things, listings whose facts line up, and mentions on pages the engine already trusts (local press, genuine industry bodies, partners) that corroborate what you say about yourself. The signal is not “this business claims to be good”. It is “multiple sources that do not control each other independently arrive at the same conclusion”.

This is also why review work pays off twice over. Reviews matter to the searcher directly, but they also matter as the richest, freshest, most independent body of text describing your business that an engine can read. Recency counts: a wall of five-star reviews from three years ago reads as a business that used to be good. A steady trickle of recent, specific reviews reads as one that is good now. If you want a structured way to keep that trickle going, our guide on how to get Google reviews covers the mechanics.

Entity clarity and extractability earn you a place in the room. Consensus is what makes the machine comfortable saying your name once it gets there.

What we saw testing this across the four big answer engines

In May 2026 we ran a structured sample of 600 local-intent prompts, 150 against each of the four engines most people actually use, spread across twelve service categories and four cities. We were not trying to crown a winner. We wanted to know how often each engine is willing to commit to naming a real business, and what the businesses it names have in common. Two findings stood out.

First, the engines behave differently, and the difference is mostly about how much they lean on live retrieval. The more an engine grounds its answer in a real-time search, the more willing it is to name a specific business, and the more your day-to-day web presence matters versus what the model happened to memorise during training. Second, and more useful, the businesses that got named were not the ones with the cleverest content. They were overwhelmingly the ones that were easy to identify and well corroborated. The boring levers won.

A two-hour audit you can run this week

You do not need a tool to start. You need a quiet two hours, the questions your customers actually ask, and a willingness to read your own business the way a machine would.

1. 1

   Ask the engines the real questions

   Write down the ten ways a customer would describe needing you, with their real constraints (location, urgency, speciality, budget). Put each one to ChatGPT, Gemini, Perplexity and Copilot. Record who gets named, who does not, and which of your competitors keep recurring. That recurring set is your real competition on this surface.
2. 2

   Audit your entity facts

   Open your website, your Business Profile and your biggest directory side by side. Check name, address, phone, hours and category match exactly. Every mismatch you find is a confidence deduction you have been paying without knowing it.
3. 3

   Pressure-test extractability

   For each of your ten questions, find the plain-text answer on your own site. If the answer is in an image, a PDF, a booking widget, or simply missing, that is a page to fix. Aim for a clear sentence a machine could quote verbatim.
4. 4

   Check your consensus

   Look at review volume and, more importantly, recency and specificity. Are recent reviews mentioning the things you want to be known for? Are there any independent pages (local press, real bodies) that corroborate your story, or is every source ultimately you talking about yourself?
5. 5

   Fix the cheapest gaps first

   Rank what you found by effort. Detail mismatches and missing plain-text answers are usually a morning's work and move you straight into contention. Consensus is slower, so start it now but expect it to compound over months.
6. 6

   Re-run the prompts in a month

   This surface is noisy, so a single check tells you little. Re-running the same prompts monthly turns anecdote into a trend you can actually manage. That ongoing measurement is exactly what an answer-engine visibility tracker is for.

What this is not

A lot of the early advice on this topic is confident and wrong, usually because it borrows the mental furniture of traditional SEO and assumes it transfers. It mostly does not. A few corrections worth holding onto.

Where this fits, and what to do on Monday

Keep this surface separate in your head and in your reporting. Your performance inside ChatGPT, Gemini, Perplexity and Copilot is not the same thing as your position in Google's Map Pack, and it is not the same thing as Google's own AI features either. Folding all of it into one “AI” line in a report is how teams end up unable to explain what is actually happening. Track answer engines as their own scoreboard.

Then pull the levers in order. Get your identity clean so a machine can tell exactly who you are. Make your pages easy to quote so you are the path of least resistance. And invest, patiently, in the independent corroboration that makes recommending you feel safe. None of it is exotic. It is the same instinct that has always driven local search, applied to a reader that happens to be a machine summarising on a customer's behalf. The businesses that get this right early will be the names that get said while everyone else is still trying to rank.