What an AI agent needs to know about a location

Earlier this year, a small company called VOYGR tested four leading language models on 345 local questions across more than 50 cities. One prompt asked for cantinas in Condesa, in Mexico City. The models returned real cantinas, in Centro and Coyoacán. In another test, every configuration offered booking advice for a Buenos Aires restaurant that had closed for good. Even the best performer recommended a place that did not exist, was permanently closed, or was in the wrong neighborhood 8 percent of the time. VOYGR sells a verification API, so it has a stake in that finding. The examples still match what anyone who manages physical locations has seen.

These are no longer edge cases. BrightLocal's 2026 consumer survey found that 45 percent of US adults had used an AI tool for a local business recommendation in the past year, up from 6 percent the year before. Google started rolling out Ask Maps, a conversational layer over Maps built on Gemini, in March 2026 in the US and India, drawing on more than 300 million places. An answer that used to be a list of ten blue links is becoming one sentence, followed sometimes by an action.

So the question is worth asking carefully: what does an agent actually need to know about a location to answer well? Our view comes from building Mapboard at WAOF, a technology, product and creative studio that started in Monterrey. Mapboard is local presence infrastructure for operators with many physical locations. It is a partial view, and we say where it stops.

Declared facts are necessary and not enough

The first layer is what the business declares: name, address, phone, category, regular hours. It sounds solved. It is not. In a late-2025 study of about 350,000 US locations, SOCi asked ChatGPT, Gemini and Perplexity for basic business profiles and found ChatGPT's data accurate 65.5 percent of the time and Perplexity's 69.8 percent, against 99.2 percent for Gemini. The gap tracks access. Gemini sits on top of Google Maps; the others assemble the picture from a fragmented ecosystem of directories, websites and review sites.

Two things follow. One is that "the" profile of a location is really many copies, kept by different parties at different speeds. The other is that the copy an agent reads depends on which agent is asking. Yext's analysis of 6.8 million AI citations, collected between July and August 2025, found that Gemini drew 52 percent of its citations from brand websites, while OpenAI drew 49 percent from listings. Yext also notes that Gemini uses Google Business Profile data without citing it, so citation counts understate how much an answer depends on the profile.

What changes is what matters

The second layer is where most errors live, and it is not marketing data. It is the operational state of the location. Holiday hours. A temporary closure for renovation. A pharmacy that is open 24 hours while its attached clinic closes at nine. A gas station whose car wash is out of service. A convenience store whose alcohol sales stop at a different hour than its doors.

None of this fits a profile written once a year. All of it is known first by someone in operations, usually at the store or district level, and usually long before it reaches whoever manages the listing. This is the gap behind Mapboard. For years, retail chains in Mexico told us the same thing: local-search tools are built for marketers, while the people responsible for hundreds of physical locations need something different. A physical location changes constantly. If its digital presence does not keep up, an agent will give a confident, well-formatted answer that was true last month.

Reputation is now a filter, not a score

The third layer is what customers say. For a human, a 4.1 rating is one input. For an AI assistant choosing three places to mention, it can be the difference between appearing and disappearing. SOCi's 2026 Local Visibility Index reports that ChatGPT recommended about 1.2 percent of brand locations in its test, against a 35.9 percent appearance rate in Google's local 3-Pack, and that recommended locations averaged 4.3 stars. These are one vendor's measurements under one test design, but the direction is consistent with what the platforms themselves describe.

That makes the network average a dangerous number. A chain that averages 4.0 has locations well below that, and those are the ones an agent quietly drops. What an agent needs is not the average but the pattern: the complaint that repeats at one branch, the line at checkout at peak hours, the restroom that keeps being closed. Those patterns are also exactly what an operations team needs to fix, which is why we think reviews belong as much with operations as with marketing.

Whose version is true

The fourth layer is the one most discussions skip: provenance. Google's own help pages explain that it gathers information about a business from user reports and licensed content and may update a profile when sources say it is wrong. When a user suggests an edit that needs the owner's review, the owner has four days to respond. After that, Google may publish it automatically if other public information supports it. That is a reasonable design for a map of the world. It also means the truth about a location is negotiated among the owner, the public and the platform, and an agent reading the result cannot tell which party last touched each field.

For an operator, that turns accuracy from a one-time setup into a standing responsibility with a name attached. Somebody has to own each location's record and respond when it is challenged.

Where an agent should stop and ask a person

If agents increasingly read locations, they will increasingly be asked to change them: reply to reviews, update hours, post an announcement across a network. Mapboard is evolving toward a more agentic way of working, and on its public site we have drawn one line clearly. Its assistant, Orbit, proposes replies and flags patterns; a person on the team edits and approves before anything is published, and the execution is logged. It is slower than full automation. We think that is the right trade for now, because a wrong hour published to a thousand profiles is not a typo, it is a thousand wrong answers.

There are limits to what any one tool covers. Mapboard publishes only to Google Business Profile today. BrightLocal's CEO points out that tools like ChatGPT cannot see inside Google's reviews, so a location's reputation elsewhere matters too. No single system, ours included, controls everything an agent reads.

The short answer

An AI agent needs four things from a location: what it is, what it can do for you right now, what people have been saying about it recently, and who stands behind that information. Models are getting better at the first. The other three come from the operation itself. The companies whose locations get described accurately will be the ones that treat that information as part of running the store, owned by the people who run it.