How AI Systems Find and Recommend Local Businesses

How do AI systems decide which local businesses to recommend?

They interpret what the person wants and where they are, retrieve information from search results, business profiles, review platforms and directories, and then name businesses that appear both relevant and well evidenced across those sources.

No single source decides it. In our experience the businesses that get named consistently are rarely the ones with the most impressive website. They are the ones that are described the same way in a lot of places, have enough recent public feedback to be credible, and are unambiguous about what they do and where they do it.

What makes local recommendation harder than it looks is that a single short question contains at least six separate problems, each of which can be answered wrongly:

• What does the person actually want? A category, a specific service, an urgent solution or an opinion about one named business.
• Where are they? Explicitly stated, inferred from the session, or unknown, which changes the answer entirely.
• Which businesses genuinely provide it? Not which ones are nearby, which ones do this particular thing.
• Which of the available information is current? Hours, availability, whether a service is still offered, whether the business still exists.
• Which businesses are credible enough to name? Naming a business is a claim about it, and the system needs something behind the claim.
• Which sources support that conclusion? The evidence that can be pointed at, which is often not the business's own website.

A business can be a perfect answer to the first four and lose on the last two. It can also be a mediocre answer to the third and win everything else, which is how a general practice ends up being recommended for specialist work it barely does.

The rest of this article walks through the steps in roughly the order they happen, and then covers what a business can actually do about each one.

Query interpretation

The first thing that happens is that the question gets turned into an intent, and small changes in wording produce genuinely different candidate sets. Four questions about dentistry make the point:

• Best dentist near me. A broad category question with an implicit location. Almost any general practice qualifies, so the deciding factor becomes reputation and evidence rather than relevance.
• Emergency dentist open Sunday. Two hard constraints. Most practices are eliminated by availability before quality is considered at all, and a practice that does handle Sunday emergencies but has not said so is eliminated with them.
• Best dentist for nervous patients. A service-quality question with no clinical category attached. It rewards whoever has written something specific about sedation, appointment length or how they handle anxiety, and ignores everyone who has not.
• Is Harbour Dental good for implants? A branded verification question about one named business and one named procedure. Nobody else is being considered, and the answer depends almost entirely on what third parties say.

All four belong to dentistry. They have almost nothing else in common. The first is a reputation contest, the second an availability filter, the third a content gap and the fourth a due-diligence check on a single entity.

The phrasing also determines whether a shortlist is produced at all. Some questions return general advice with no companies named, some return a category explanation, and some return three names and a sentence each. A prompt set that only contains the recommendation-shaped questions will overstate how visible you are; one that only contains informational questions will find nothing to measure.

It follows that half the value of building a prompt set is discovering which phrasings produce recommendations in your category. That is a finding in itself, and it is usually the first surprise.

Location

Location is resolved from whatever the system has, and there are more inputs to it than most businesses realise. In rough order of reliability:

• An explicit place in the question. The strongest signal, and the only one a business can plan around.
• Inferred or session location. Whatever the product has established about where the user is, which is neither visible nor controllable.
• Your registered address. What appears on your profile and in structured data.
• Your stated service area. The places you say you serve, which is a different claim from where you are based.
• Neighbourhood and district names. The granularity people actually use locally, and the one businesses most often omit.
• City and region. Broad enough that many businesses compete, so rarely decisive on its own.
• Proximity. Distance between the user and the business, where both are known.
• Opening hours and availability. Which turn a geographic question into a filtered one whenever the question has a time in it.

Compare London dentist with emergency dentist near King's Cross tonight. The first is a broad category question in a city of several thousand practices, where geography barely narrows anything and reputation decides. The second is almost entirely a filter: a small radius, an urgent service and an availability window, in which most of the field is eliminated before quality is considered at all. They are different retrieval problems and a business can dominate one while being invisible in the other.

We would not put numbers on how any specific system weights these inputs, because nothing published establishes that and the honest answer is that it varies by product and by question. What is safe to say is that the inputs a business controls are the explicit ones: the address, the stated service area, the named neighbourhoods, the hours.

The practical consequence is that your service area needs to be stated in words on your own site, not implied by an address in the footer. A practice in a suburb that serves three neighbouring towns should say which three. And because the boundaries of visibility are geographic, testing has to be too: being recommended in your own postcode and invisible two towns over is a common pattern, and it is invisible to any measurement that reports one national number.

Service relevance

The system has to establish that you provide the specific thing being asked for, and it can only do that from what you have actually said. This is where vague positioning becomes expensive, and the cost is usually invisible because nothing appears to be broken.

Consider two sentences describing the same practice. The first: we offer comprehensive solutions for every smile. The second: we provide dental implants, emergency appointments, root canal treatment and Invisalign from our Hampstead clinic.

The first resolves nothing. It is not badly written and it reads perfectly well to somebody who already knows the practice, but it cannot be matched to a request for implants, or emergency work, or a clinic in Hampstead, because none of those things are in it. The second resolves the category, four specific services and the location in a single sentence, in text.

The operational version of this is short. Name the services individually rather than describing a philosophy of care. Use both vocabularies, the one customers use and the one the profession uses, because they frequently differ (patients say invisible braces, clinics say clear aligners). State where the service is delivered from and which areas it covers. And say what you do not do, because it prevents a category of wasted enquiry and because a stated exclusion is a more specific claim than silence.

Search retrieval

Many assistants perform a live search and read the results before answering. When that happens, conventional ranking becomes a direct input into who gets considered.

This is the most concrete link between SEO and AI recommendation, and it is why treating them as separate disciplines tends to produce worse outcomes in both. If you do not appear in results for the equivalent query, you are not in the pool of material that gets read.

The nuance worth carrying over is that ranking for the literal question is not the requirement. Systems reformulate: a conversational question becomes several retrieval queries, and material can be found through a synonym, a narrower sub-question or a supporting fact rather than the phrase that was typed. Large-scale evidence bears this out, and we go through it in the AEO and SEO comparison.

Not every answer works this way, and no product tells you which mode it used. But when the retrieval path runs through search, search visibility is close to a prerequisite, and neglecting it on the assumption that AI has replaced it is a straightforwardly bad trade.

The relationship is not one to one, and the disagreements are informative. We have seen businesses rank well and still be absent from the equivalent AI answer, usually because the entity behind the ranking page is unclear or unsupported elsewhere. We have also seen the reverse, where a business with modest rankings is named consistently because the wider web describes it very clearly. Measuring both, and noticing when they diverge, tells you more than either on its own, and it is one of the questions our research programme is currently looking at.

Website information

Your site is where the system confirms the specifics: what you do, where, for whom, and how to proceed. It is also the source most likely to be quoted directly.

The facts worth stating explicitly, and the ones most sites omit, are narrow enough to list:

• The service, named. One page or one clearly headed section per significant service, using the term a customer would use.
• The price, or an honest range. With the variables named if a fixed figure is impossible.
• Availability. Whether you are taking new customers, and what the current wait is.
• Who provides it. Named people, where the category makes that relevant, which is most professional services.
• Qualifications and registrations. Stated in a form somebody could verify against the issuing body.
• Geography. Where the service is delivered from, and the areas covered, in place names.
• The process. What actually happens, in order, including the first appointment.
• Opening hours. Matching the business profile exactly, including how urgent work is handled outside them.
• Restrictions and exclusions. Who a service is not suitable for, and what is not included.
• How to book. The actual mechanism, not only a contact form.

The pages that do well here are unglamorous, and the reason is worth stating plainly: every item above is a fact a general-purpose model could not know without being told. Everything else on a typical service page is explanation the model can already produce for itself. What does not help is a site where the substance is diffused across marketing copy that never quite states anything checkable.

Business profiles

Map and business profiles are among the most heavily used sources for local questions, and they are usually the fastest thing to fix.

The essentials: correct primary category, complete service list, accurate address or service area, genuinely current opening hours, and a description that matches your website rather than one written years ago for a different positioning.

The primary category deserves particular attention, because it is unusually consequential and unusually neglected. It is the single field that most directly answers what kind of business is this, it is chosen once during setup, and it is frequently chosen by whoever happened to create the listing. A practice whose primary category still says cosmetic dentistry will lose emergency questions to a practice whose category says dentist, whatever either of them actually does.

The wider principle is consistency rather than completeness. A profile that is fully populated but disagrees with the website about the address, the hours or the service list has added a conflict rather than resolved one. It is worth being explicit that no individual field guarantees anything: a correct primary category does not cause inclusion in an AI answer. What it does is remove one reason to be excluded, which is the realistic standard for most of this work.

Reviews

Reviews are the most readily available evidence of whether other people found a business good, which makes them useful to a system that is being asked to vouch for somebody.

Volume, recency and rating all appear to contribute. We would not put a precise weighting on them, because we have not seen a credible public study that establishes one and our own work on this is not finished. What is clear is that a thin or stale review profile gives a system very little to justify a recommendation with.

The content is underrated. Reviews that name the treatment, the location or the outcome give far more usable material than a five-star rating with no text.

That caveat is worth repeating because the chart invites the wrong reading. These are the things consumers say make a review persuasive. They are not weights inside any AI system, and nobody has published weights inside any AI system. What the figure is genuinely evidence of is the shape of the reputation layer these systems are drawing on: a body of reviews that is consistent, recent, descriptive and responded to reads as credible to people, and it is the same body of text a system has available when it is asked to justify naming somebody.

AI recommendation is often the start of the decision, not the end

Being named by an assistant does not close the sale, because most people go and check. BrightLocal's 2026 consumer research found that the overwhelming majority of AI users verify what they were told before acting on it.

The strategic implication is the most commercially useful point in this article. AI visibility starts the consideration; it does not finish it. A business that wins the recommendation and then loses the verification step has gained nothing, and the verification step happens somewhere it does not control: a review platform, a map profile, a directory, a forum thread.

This changes what the work is for. Improving your own pages raises the odds of being named. Improving the wider reputation layer decides whether being named converts. The second is slower, less satisfying to report on, and the reason a strong AI visibility number can sit alongside flat enquiry volume.

It also argues against a specific kind of optimism. The businesses most exposed here are the ones whose recommendation rests on a well-written website and a thin public record: named in the answer, then checked, then found to have eleven reviews from 2023. The recommendation was real. It just did not survive five seconds of scrutiny.

Directories, mentions and what actually gets cited

Third-party sources decide a large share of local answers, and the useful skill is identifying which ones rather than being present everywhere. A small number carry real weight in each category and a great many carry none at all, and the difference is testable rather than a matter of opinion.

Choosing the directories that matter

A directory or third-party source is worth your attention when it meets several of these:

• It appears for your category. Search the category and the location and see whether it surfaces at all.
• It ranks for relevant intent. Not for its own brand name, for the questions your customers ask.
• It appears in the citations behind AI answers. The most direct test available, and the one almost nobody runs.
• Customers genuinely use it. Ask a few. In some categories the answer is a body nobody outside the trade has heard of.
• It has category authority. A licensing authority or professional register has standing that a general business listing does not.
• Its data is accurate. Including yours, and including whether it has quietly syndicated an address you left in 2021.

What tends to count, on those criteria, is the sources a human would actually consult: professional registers, trade bodies, licensing authorities, established local guides, sector-specific comparison sites. What tends not to count is the bulk submission service sold as an SEO product, which is measured in volume precisely because none of the individual entries can be justified.

One entry on a body that genuinely governs your profession is worth more than fifty listings nobody consults, and the fifty carry a cost: each one is another place your details can go stale and contradict the others.

Ordinary mentions

Mentions elsewhere contribute to the sense that a business exists in the world and is regarded in a particular way. This includes local press, community pages, supplier and partner sites, sponsorships and event listings.

None of these are links to be acquired in the old sense. They are evidence of activity, and they are more persuasive when they are genuine, which is inconvenient for anyone hoping to buy them.

For most local businesses this accumulates slowly as a side effect of operating, and the useful intervention is simply making sure your name is spelled and described consistently when it happens. A mention that calls you by a former trading name is doing less work than one that does not, and correcting it is usually one email.

Reading the citations

Where the assistant shows its sources, read them. This is the closest thing to seeing the working, and across a set of questions you will usually find a small group of sources doing most of the work in your category. That group is your actual competitive landscape for retrieval, and it is frequently not what the business expected.

Once identified, a vague objective becomes a concrete list: be present, be accurate, and understand why those sources are trusted. The full workflow for doing that systematically, including how to categorise what each cited page is actually supplying, is in how to improve your visibility in ChatGPT, and the research on what characteristics make a page citable in the first place is in what makes AI cite a page.

Entity clarity

The single most common fixable problem we encounter. A business is described one way on its site, another on its map profile, and a third on a directory entry created years ago.

Every inconsistency is something a system has to resolve, and resolving it in your favour is not guaranteed. Two trading names, a moved address that still appears somewhere, or a category that no longer reflects the business all cost more than they appear to.

The fix is an audit and a cleanup rather than anything clever. It is boring and it works.

Trust signals

Systems being asked to recommend a business tend to favour ones with visible, checkable credibility, and the effect is most pronounced in the categories where a bad recommendation does the most damage: healthcare, legal, financial and professional services.

In those categories the specific things worth having in place are:

• Named practitioners. Real people with names, not a team page of first names and job titles.
• Stated credentials. The actual qualification, awarding body and where relevant the registration number.
• Licensing and registration. Confirmed as current on the register itself, rather than asserted on your own site.
• Professional body membership. Listed both ways, on your site and in the body's own directory.
• A real address. Where the category implies premises, an address that resolves to them.
• Published policies. Complaints, cancellation, regulatory disclosures, whatever the category requires.
• Contact details that work. A phone number that is answered and an email that is monitored.

We would not describe any of these as hidden ranking factors, and it would be dishonest to imply they are individually rewarded. The mechanism is more mundane. Each one is a claim that can be independently checked, and a business made of checkable claims is one a system can justify naming. A business whose credibility rests entirely on its own assertions is not obviously untrustworthy; it is simply harder to vouch for than the alternative.

For regulated professions in particular, presence on the relevant register is worth confirming rather than assuming. Registers go stale, practitioners move between practices without the record following them, and a lapsed or misfiled entry is both a compliance issue and a visibility one.

Measuring local visibility

A single visibility percentage is close to meaningless for a local business, because local visibility is geographic and a national average hides exactly the variation that matters. What you want at the end is a map, not a number.

Two things make aggregation mandatory before geography even comes into it. Recommendations differ between identical asks, because these systems are probabilistic, because different products retrieve differently, because location and phrasing change the result, and because models change without notice. Two colleagues in the same office can get different answers to the same question and neither result is the truth. The pattern across a hundred questions is real; any individual answer is not. There is more on what moves a single answer, and by how much, in how to improve your visibility in ChatGPT.

Given that, the useful design is to test across several dimensions at once and keep them separate in the reporting:

• By town. Each place you actually serve, tested as its own question set rather than assumed to follow from the head office location.
• By neighbourhood. The granularity locals use, which is frequently where a business is strong or absent for reasons nothing else explains.
• By service. Each named service separately. Being the answer for general work and invisible for the specialist work is common and commercially important.
• By modifier. Best, cheapest, near me, open now, for nervous patients. Each modifier is effectively a different question.
• By urgency. Emergency and same-day questions behave differently from planned ones, because availability filters before quality.
• By comparison. You against a named competitor, which is where a weak public record shows up fastest.
• By reputation query. Is this business any good, which tests the verification layer rather than the discovery one.

Cross those and you get a grid rather than a headline, and the value is in the gaps. A practice recommended for routine work across four towns and absent for implants in all of them has a content problem. One recommended in its own town for everything and absent one town over has a geography problem. One named in discovery questions and poorly reviewed in reputation questions has a verification problem, which is the hardest of the three to fix and the most expensive to ignore.

The grid also makes the work finishable, which a single number never does. Local intent narrows the field considerably, and that is good news: it is far easier to become a consistent answer for one city and one service than for a whole country, and the grid tells you which one to pick.

How a business can improve its chances

In the order we would usually do it:

1. Establish a baseline across a realistic set of local questions.
2. Fix entity consistency: name, address, service area, category and services everywhere.
3. Complete and correct the map profile, starting with the primary category.
4. Make sure the site states services and areas explicitly, in text.
5. Answer the top customer questions directly, one page or clear section each.
6. Build reviews steadily, and encourage specifics rather than just a rating.
7. Get present and accurate on the sources that actually get cited in your category.
8. Re-run the same questions and compare.

None of it is dramatic. Most of the businesses we see improve do so because a lot of small ambiguities got removed, not because of one clever intervention. There is more on how we measure this on Work.