AI and the Local Business

What changed about the near-me question

A local business has always depended on being findable at the moment somebody needs it. What has changed is not that dependency but the shape of the moment. For roughly two decades the common pattern was a list: a map panel, a set of listings, several links, and a person doing the comparing. Several widely used systems now put a composed answer in front of that list, naming a small number of businesses in prose before anybody scans anything.

The size of that shift is measurable on at least one surface. Pew Research Center tracked the browsing of 900 US adults through March 2025, covering 68,879 unique Google searches, of which 12,593 produced an AI summary. Users clicked a conventional search result in 8 percent of visits where a summary appeared, against 15 percent of visits where none did, and clicked a link inside the summary itself in 1 percent of visits. That study measured Google searches in general rather than local queries specifically, it covers one panel in one month, and it should not be read as a finding about near-me behavior. It does establish that the answer layer changes what people do next.

The practical consequence for a local business is narrow and worth stating plainly: appearing in a listing database and being named in an answer are no longer the same event.

There is no single index behind near me

The most common error in this area is treating AI-driven local discovery as one system. It is not. Different assistants are publicly documented as drawing on different underlying records, and those records are maintained in different places by different people.

What can be established from operator statements:

• Google's own guidance for site owners on AI features lists keeping Business Profile information up to date among ordinary best practices, and Google's Business Profile Help documents relevance, distance and prominence as the factors behind local ranking on Search and Maps.
• Foursquare has publicly stated that its Places API, which it describes as covering more than 100 million points of interest across more than 200 countries, delivers real-world results to ChatGPT search.
• Perplexity's use of licensed Yelp data for local and restaurant results was announced and reported in March 2024, with Yelp links returned alongside the answer.
• Apple's Business Connect lets a verified business manage the place card that Apple says appears across Apple Maps, Messages, Wallet, Siri and other apps.
• Google's LocalBusiness structured data documentation describes markup on the business's own website as a separate, publisher-controlled record.

The consequence of many records

If several assistants read several different records, then a business can be complete and accurate in one ecosystem and thin, stale or entirely absent in another, without anything being broken. Two assistants asked the same near-me question can return different names, and the difference may say more about which places database each consulted than about either business.

This also means that the work of maintaining a local record is not finished when one profile is filled in. AIOFacts treats the local record as plural by default: a set of parallel statements about the same entity, held by different custodians, each of which can drift independently. Consistency across them is not a ranking tactic. It is the condition under which any reader, human or machine, can tell that the records describe one business.

Location is an input, and it is not always precise

Near me is not a fixed query. It is a query plus a location value, and that value is derived differently on different platforms, with different precision, sometimes not at all.

OpenAI documents that ChatGPT can optionally use device location to give more relevant results, including local recommendations, and distinguishes approximate location, based on IP address and reflecting a general area such as a city or region, from precise location. Device location sharing is documented as optional and off by default, and on mobile the precise component can be toggled separately. Google's Business Profile Help states that distance is measured from the customer who is searching, and that if a customer does not share where they are, Google uses what it knows about their location.

Two things follow. First, the same phrase can resolve to a street corner in one session and a city-scale area in another, which is a legitimate source of variation in what gets named. Second, any attempt to check whether a business is being recommended has to hold location constant to mean anything. Position checks taken from different devices, networks or accounts are not comparable, and a business that appears to have vanished may simply have been queried from a different starting point.

What the platforms document about selection, and what they do not

Google's Business Profile Help describes three factors for local results. Relevance is how well a profile matches what somebody searched for, and the stated remedy is complete, detailed business information. Distance is how far each business is from the searcher. Prominence is how well known a business is, described as based partly on information such as how many websites link to the business and how many reviews it has.

Google's guidance for AI features is separate and unusually direct about optimization: it states that there are no additional requirements to appear in AI Overviews or AI Mode, and no special optimizations necessary, and that the same foundational practices apply. The same document describes a query fan-out technique, issuing multiple related searches across subtopics and data sources, and says these features tend to display a wider and more diverse set of supporting links than classic web search.

The limits of that documentation matter as much as its content. These are descriptions written by the operator of the surface, not specifications, and they do not disclose how a response is assembled. No party outside a model's operator can observe its internals. Where operator documentation and circulating practitioner folklore disagree, the documentation is the more checkable of the two, and it is still an interested party describing itself.

The record the business itself controls

One record in this landscape belongs to the business outright: the machine-readable description published on its own website. Schema.org defines LocalBusiness as a subtype of both Organization and Place, giving it properties inherited from each, with dozens of narrower subtypes such as Restaurant, DaySpa and HealthClub. Google's implementation guidance requires name and address, recommends telephone, url, openingHoursSpecification, geo coordinates, priceRange, review and aggregateRating among others, and advises using the most specific subtype available and defining each physical location separately.

What markup does is make a claim explicit and unambiguous to any reader that parses it. What it does not do is verify the claim, and Google's own AI features guidance names no markup as an AI-specific requirement. Eligibility for a rich result is documented as eligibility, not a guarantee of display.

AIOFacts treats structured data as valuable for a reason that is independent of any ranking effect: it removes interpretation from a set of facts that a business is in the best position to state correctly. That is a position, not a measured finding.

Local facts expire, and expiry is not documented

Local information is unusual among web content in how much of it is time-bound. Hours change seasonally and on holidays, a phone number is ported, a service area is redrawn, a location closes. Google's guidance ties complete and accurate information to relevance, and ties review volume to prominence, both of which are maintained facts rather than one-time entries.

What is not publicly documented, for most systems, is refresh latency: how long a corrected fact takes to reach the surface where an assistant restates it, or how long a superseded fact can persist in an intermediate copy. This is a genuine gap rather than a detail omitted here. The available evidence indicates only that multiple copies of the record exist and are updated by different parties on schedules none of them publish.

The bounded implication is procedural, not clever: a fact that has changed should be corrected at the record itself, in each ecosystem that holds one, rather than only on the website, because the website is one of the copies and not necessarily the one being read.

What this changes for practice

None of the following is a guarantee of being named in any answer, and no honest account can offer one. Each is defensible on the narrower ground that it makes the record clearer and more consistent for anything reading it.

• Claim and complete the record in each ecosystem an assistant is publicly known to draw on, rather than treating one profile as the whole job.
• Keep name, address, phone number and hours identical across every record, including punctuation and suite numbers, so the records are recognizably about one entity.
• Choose the most specific category or schema subtype that is accurate, since a general one describes the business less clearly and an inaccurate one is a false statement.
• Publish the same facts on the business's own domain in LocalBusiness markup, so the publisher-controlled copy exists and agrees with the rest.
• Write pages that answer the question a person actually asks, in the words they use, rather than pages about the business in general.
• When checking visibility, hold the location value constant, and record which assistant, which account and which device produced each observation.

Where the evidence runs out

Three limits should be stated rather than smoothed over. No operator publishes how a local recommendation is composed, so any account of why a specific business was or was not named is inference. Widely circulated share figures for how much of one assistant's local results come from one data provider are practitioner measurements from small samples, not disclosures, and this reference does not repeat them as fact. And the arrangements described here are commercial data licences that can be renegotiated, so a mechanism accurate on the access date above may not survive the year.

The honest summary is that the inputs are documented and the selection is not. That is enough to act on the inputs, and not enough to promise an outcome.