THE RECOMMENDATION LAYER · JULY 28, 2026 · 10 MIN READ

How does ChatGPT choose which local businesses to recommend?

ChatGPT rewrites your buyer's near me into a named place, then answers from what it can verify there. The local mechanics, in plain English.

By Jacquie Baker
TL;DR: WHAT TO REMEMBER

ChatGPT chooses local businesses by first turning your buyer's question into a different question. It rewrites "who should I call near me" into a search for a named place, sends that to search providers, then builds an answer from what it can retrieve and verify about the businesses it finds there. OpenAI documents that rewriting step in its own help material. What happens after it is a working model rather than a published mechanism, because OpenAI has not put a local ranking formula on the record. Your buyer asked about their street. The machine went looking for a city.

Which is why the result can feel so unfair. A customer three streets from your shop asks for a plumber who can come tonight. Three names come back, and one of them subcontracts to you when the job gets complicated. It reads like a verdict on the work. It is closer to a filing problem.

Here is the useful way to hold it. Local is a fact about you. Locatable is a fact about your record. Your neighbors know exactly where you are and what you will drive out for. The record is vaguer than they are, and the record is the only thing in the room when the question gets asked.

This guide follows one local question through the machine. How the place gets resolved, how the candidate list gets built, why a service area has to be looked up rather than assumed, what reviews are actually doing in a local answer, and why the answer for your town behaves nothing like the answer for your industry.

Your Buyer's Question Is Not the Question That Gets Asked

Start with the mechanic OpenAI has actually published, because it is more useful than anything anyone has guessed. In its help material for ChatGPT search, OpenAI says that when it works with search providers, "ChatGPT search typically rewrites your query into one or more targeted queries that it sends those providers."

Then it gives a local example, in its own words. "If you type into ChatGPT 'What are some good restaurants near me?' and ChatGPT determines from your IP address that you are in the San Francisco area, ChatGPT may rewrite your prompt into the search query 'top restaurants San Francisco.'" OpenAI adds that it collects general location information based on your IP address, and may share that general location with search providers to improve the accuracy of results.

Read what that does to your buyer's question. In that example, "near me" leaves the conversation and a city name arrives in its place. What gets retrieved next is retrieved for the city, not for the street corner your customer is standing on. A user who has switched on precise location sharing can get something tighter than that.

Which raises the obvious question of how precise any of this is. OpenAI's own answer to whether it uses your IP address is that it "may use your IP address to estimate your general location, such as your country, state, or city," that this "is approximate and based on your internet connection," and that ChatGPT "does not access your device's GPS or precise location unless you enable location sharing." Device location is optional and off by default. Plenty of local questions are therefore answered at city or region level, with no doorstep in the picture at all.

So the practical move is unglamorous. Say the names. The towns, the neighborhoods, the counties, spelled the way locals spell them, on pages that also say what you actually do there. If the rewritten query carries a city and none of your material carries it back, you have handed the machine nothing to match.

How the Candidate List Gets Built

Once there is a place-shaped query, the assistant has to find businesses for it. Two documented details are worth your attention here, and both cut against the way this work usually gets sold.

The first is access. OpenAI's guidance on getting a site into ChatGPT search is blunt: "Ranking in ChatGPT Search is based on a number of factors designed to help users find reliable, relevant information. There is no way to guarantee top placement." It then names the entry requirement. You allow its search crawler to reach your site, and your host or content delivery network permits traffic from its published addresses. Nobody sells the answer. Somebody may nevertheless be blocking the door, and a well meaning security rule set long before any of this existed can do it without anyone noticing.

The second is that retrieval and citation are different lists. OpenAI's developer documentation for the same web search capability describes responses carrying "annotations for the cited URLs," alongside a separate sources field listing "all URLs retrieved during a web search." A machine can pull your page, read it, and still write the answer without you in it. Being retrieved is not being named.

Then there is the furniture around the answer, which is licensed rather than built. Mapbox announced in December 2024 that it worked "with OpenAI to integrate Mapbox maps into ChatGPT search." Mapbox called it the first large language model service to combine text answers with tailored auto generated maps. Review and listing data has followed the same pattern, most visibly through the Yelp licensing arrangement with OpenAI, which our piece on whether a small local business can get recommended at all covers with its dates. Those arrangements change. Your record is the part that does not have to.

Now the honest part, since this is where local AI advice usually goes soft. We have not found a primary source stating that ChatGPT queries your Google Business Profile or your Bing Places listing directly. Ask anyone who claims otherwise to show you the document. What is documented is that answers are assembled from sources the system retrieved. Structured profiles are one place your business facts exist in a checkable, standardized form, which is reason enough to keep them accurate. Build them because they are your facts written down, not because a vendor promised you a pipeline.

Your Service Area Is Something It Has to Look Up

Here is the split that local content almost always skips. Take two ordinary examples. A dentist has an address. A plumber has a base and a territory, and those are different problems to solve.

Google documents its rules for this more plainly than most platforms do, so borrow its shape while remembering it is describing its own product rather than ChatGPT's. Google tells service area businesses to remove the address and enter the service area instead, when customers are not served at the location. It allows up to 20 service areas. It says you "must specify your service area by city, postal code, or another type of area," and that "the boundaries of your overall area shouldn't be more than about 2 hours of driving time from where your business is based."

None of those numbers are ChatGPT's rules. What transfers is the shape. Coverage is a list of named places with an outer edge, and it is written down or it is inferred. A territory that lives only in your dispatcher's head is harder for anybody to verify, customer or machine, than the identical territory published as a list of towns.

A hidden address is a policy, and often the correct one. An unnamed service area is a gap, and gaps get filled with whoever did write it down.

What Reviews Are Doing in a Local Answer

Reviews do unusual work in local answers, and not for the reason most owners assume. For a small professional service firm, they are often among the few third party descriptions of the work that exist in writing anywhere. That can make them the corroboration layer whether you intended them to be or not.

Google states its own position plainly: "More reviews and positive ratings can help your business's local ranking," and it also says there is "no way to request or pay for a better local ranking on Google." That is Google describing Google, and no engine speaks for another. It still tells you what the category treats as prominence.

The part worth your attention is the text rather than the average. A review saying a technician replaced a tankless water heater the same afternoon in a named neighborhood contains four things a machine can read: the service, the urgency, the place and the outcome. A five star review saying "great service" contains a number. Only one of those is any use to a buyer who asked for something specific, and plenty of local buyers ask exactly that kind of question. They ask for a dentist who is good with an anxious adult. They ask for an attorney who has handled a bicycle case. Specific beats sunny.

Two boundaries, both worth saying out loud. We have found no published study measuring how much review wording changes a ChatGPT recommendation, so treat this as reasoning about what is available to be read rather than a proven lever. And ask for reviews, never buy them, and never make an incentive conditional on a positive one. Fake reviews and sentiment-conditioned incentives run into platform policies and US consumer protection rules, and a record built that way is a risk sitting permanently where every machine can read it.

Why the Local Answer Is Not the General One

Ask an assistant a general question about your industry and it is working with a deep written record: years of articles, comparisons and forum threads. Ask it who to call in your town and that record thins out fast. Far less tends to have been written about the six plumbers within ten miles of your buyer than about your industry as a whole.

That thinness is our best explanation for why local behaves differently. With less to read, an answer has fewer places to look, and the places left tend to be structured records and the handful of pages that actually name the town. National reputation does not automatically carry a local slot, which is how a firm can be well known in its category and still be missing from the answer in your county. It also means the specificity a small local firm can write is worth more here than almost anywhere else. The eligibility question goes into that properly.

Engines differ here, and the reason is supply rather than philosophy. Place and business data reaches each product through its own partners and indexes, so the same firm can be described well in one assistant and be missing from another with nothing changing at the business. Google also publishes far more guidance about its own local surfaces than anyone else does about theirs, which is one reason a well kept Google Business Profile earns its keep in more places than Google. Never generalize one product's behavior into a claim about "AI." We test the five engines your buyers actually ask, ChatGPT, Gemini, Perplexity, Claude and Grok, with the questions they actually ask. One screenshot from one product on one afternoon is a sample, not a finding.

There is one more mechanic worth knowing, because it is the closest thing to a preview of where this goes. OpenAI says that when memory is on, ChatGPT may use what it remembers about a user to improve the rewritten query, and its example is a vegan user in San Francisco whose question becomes "good vegan restaurants San Francisco." The question a buyer types is not the question that gets searched, and it is getting more personal, not less. Publish the specifics of who you serve and where, and there is more surface for that to land on. The underlying job looks much the same across engines, and the plain English version of it is entity identity. The general, non local version of the work is laid out in the step by step guide to getting recommended by ChatGPT, and the decision itself is walked through in how ChatGPT decides who to recommend.

So where does that leave you this month? Look first. Ask two or three assistants the exact question a good customer would ask, in your town, in their words, and write down who gets named. Then fix the locatable things in order: your service area written as a list of real places, your details agreeing everywhere they appear, and pages that answer one local situation each. The free 109-Point AI Visibility Diagnostic scores that across eight zones and shows what five engines answer today when buyers ask who to hire in your category. You start from what is actually missing.

One boundary, since somebody will promise you otherwise this quarter. OpenAI itself says there is no way to guarantee top placement, and no outside provider controls what a model says. What you can do is make yourself easier to place, easier to verify and harder to leave out, then measure what changes. Readiness is buildable. Outcomes are earned.

Want to Learn More?

Probably Genius is a done-for-you operation for expert-led firms. It starts with a conversation that gets your judgment out of your head and onto the record. That becomes 12 done-for-you thought-leadership articles every month, planned to produce approximately 60 clear Knowledge Entries, each scored 0 to 100 through the Integrity Gate, where nothing publishes under 80. Alongside it we strengthen the record beyond your own pages. For a local business that means the listings, citations and reviews describing where you work and what you do there. Your town already knows. Probably time the machines did too.

How does ChatGPT choose which local businesses to recommend?
It starts by rewriting the question. OpenAI says ChatGPT search typically rewrites a query into one or more targeted queries before sending them to search providers, and its own example turns "what are some good restaurants near me," from an IP address read as San Francisco, into "top restaurants San Francisco." From there the answer is assembled from what the system can retrieve about businesses in that place, and a business whose details agree across those sources is easier to describe than one whose details contradict each other. OpenAI has not published a ranking formula for local recommendations and says there is no way to guarantee top placement, so treat any account of the remaining steps, including this one, as a working model built from documented behavior rather than a leaked mechanism.
Does my Google Business Profile affect what ChatGPT says?
We have found no primary source stating that ChatGPT queries Google Business Profile directly, and you should be skeptical of anyone who claims otherwise without producing the document. What is true is that your profile is one of the few places your business facts exist in a structured, standardized form, and that keeping them accurate serves customers and machines alike. Google says businesses with complete and accurate information are more likely to appear in relevant local results, and that there is no way to request or pay for a better local ranking. It is also the surface Google's own AI features sit alongside, so the work is rarely wasted even where its effect on any one assistant is undocumented.
Why does ChatGPT recommend businesses further away than mine?
Often because the question was never resolved to a radius in the first place. OpenAI says it may estimate general location from an IP address at country, state or city level, that this estimate is approximate, and that precise device location is off unless a user turns it on. A city-level query has no way to prefer the business four minutes closer. What it can prefer is a business whose material clearly names that place and that kind of job. Service area businesses are especially exposed, since a territory that has never been published cannot be weighed against a competitor who published theirs. OpenAI does not publish a local ranking formula, so no single explanation can be assigned to any one answer without controlled testing.
Do different AI assistants recommend different local businesses?
Yes, and local is a place where they visibly differ. The likeliest reason is supply rather than judgment: place and business data reaches each product through its own partners, indexes and licensing arrangements, so the same business can be described well in one assistant and be missing from another without anything changing at the business. Mapbox, for example, published its integration of maps into ChatGPT search in December 2024, a reminder that the local layer around an answer is often licensed from other companies. The practical response is not to chase whichever platform signed a deal most recently. It is to make the underlying record accurate, then sample several engines on a schedule so you can tell a trend from an afternoon.

CITATIONS

  1. "ChatGPT Search" (OpenAI Help Center). OpenAI's own account of how ChatGPT search behaves, stating that it "typically rewrites your query into one or more targeted queries" for its search providers, giving the worked local example of "What are some good restaurants near me?" becoming "top restaurants San Francisco" from an IP-derived location, confirming that IP-based location is approximate at country, state or city level while device GPS is used only if a user enables location sharing, and stating that "there is no way to guarantee top placement" and that inclusion starts with allowing OpenAI's search crawler to reach your site. The single most useful published document for any local business trying to understand this. help.openai.com
  2. "Manage your service areas for service-area & hybrid businesses" (Google Business Profile Help). Google's guidance for businesses that travel to customers: remove the address and enter the service area instead, specify the area "by city, postal code, or another type of area," with up to 20 service areas, and boundaries that "shouldn't be more than about 2 hours of driving time from where your business is based." Google's rules for Google's product, cited here as the clearest published model of writing a service area down as a list of real places. support.google.com
  3. "Web search" developer guide (OpenAI). OpenAI's documentation for the web search tool, describing an approximate user location built from a country code, an IANA timezone and free text city and region fields "like Minneapolis and Minnesota respectively," and responses that carry annotations for cited URLs alongside a sources field listing "all URLs retrieved during a web search." Cited for the documented gap between what a system retrieves and what it names, and as the developer-side view of an approximate location. developers.openai.com
  4. "Tips to improve your local ranking on Google" (Google Business Profile Help). Google's statement that local results are based primarily on relevance, distance and prominence, that "businesses with complete and accurate info are more likely to show up in local search results," that "more reviews and positive ratings can help your business's local ranking," and that there is "no way to request or pay for a better local ranking on Google." Cited for what the category treats as local prominence, not as a description of any other engine. support.google.com
  5. "Mapbox 2024: A Year of Innovation and Impact" (Mapbox, December 17, 2024). Mapbox's own account of working "with OpenAI to integrate Mapbox maps into ChatGPT search," described in Mapbox's words as the first large language model service combining text answers with tailored auto generated maps. Evidence that parts of the local layer around a ChatGPT answer are licensed from third parties, which is why place data supply lines are worth understanding and not worth chasing. mapbox.com
WRITTEN BYJacquie ("Jax") Baker

Founder of Probably Genius, an AI visibility firm helping professional service brands become the named answer in AI search. Nearly two decades across technology, digital strategy and branding, including 1,000+ digital projects through her previous agency, now focused on making experts visible, verifiable and recommendable to AI. Let's talk →

Related reading

Read the method.
Then see it run on you.

The 109-point diagnostic is this library, applied to your business. Free, about an hour to present, no obligation.

See where you stand →

A real diagnostic, not a sales call in disguise.