Teckel AI · Primary measurement

Why ChatGPT answers with Google Maps listings instead of websites

Updated: September 21, 2026 · By Teckel AI

Because for questions it reads as local, ChatGPT stops answering from web pages and starts answering from business data. We measured it: across 246 non-empty answers collected on 22 measurement days between July 23 and September 21, 2026, ChatGPT replied with only map listings and zero websites in 28 of them. Never a mix. In the same window, on the same queries, Claude did it 0 times out of 79 and the Google results page 0 times out of 93.

The practical consequence is blunt: when that mode fires, no website is a candidate. Not yours, not your competitor's, not the best article ever written on the subject. A business listing is. And here is the part that the Spanish edition of this research cannot contain: the mode is not a Mexico thing. It fires on English queries too, and when it does, most of the businesses it hands you have no location at all.

Teckel AI does not appear in any of those listings either. Across 224 listing slots served in this measurement, we were named zero times. This is research published against ourselves.

What exactly was measured?

A fixed set of 12 queries, run on working days against three engines, with every answer stored as raw JSON including the full list of URLs returned. Eight queries are in Spanish, three in English, one is a brand defence query. The engines are the OpenAI API with web search and Mexico geolocation, the Anthropic API with web search, and the google.com.mx results page through SerpAPI. The numbers below come from re-reading those stored files, not from remembering what the answers looked like.

One definition matters for everything that follows. We call an answer a listing answer when every URL in it is a google.com/maps/search/ link, and a web answer when none of them is. Runs where an engine returned no URLs at all are excluded from the denominators, which is why the engine totals below do not add up to the number of runs.

Two readings of that table are worth separating. The first is that roughly two out of five ChatGPT answers with results in this query set were listings. The second is that no other engine did it once. This is a product decision by one vendor, not a property of AI answers in general, which also means it can change without notice.

Is the switch really binary?

Yes, and that is the finding with the most consequences. In 246 answers there is not one case of an answer that offered some listings and some websites. It is eight business cards and no links, or it is links and no cards. A ranking you can climb gradually is one thing. A door that is either open or shut is a different thing to plan around.

The clearest series is the Spanish commercial query mejor consultora de IA en México, run 21 times in this window: 16 listing answers and 5 web answers. The same question, the same engine, the same geolocation, flipping between two entirely different answer machines depending on the day.

Does it happen in English?

It does, and this is the section the Spanish edition cannot have. The English query who builds AI-ready websites returned eight map listings and zero websites in 4 of its 5 runs. Nothing in that question mentions a city, a country or a neighbourhood. It asks who does a job.

So the trigger is the shape of the question, not the language and not an explicit place name. If your category can be read as "a provider you might hire", it can fall into the mode, in English, from anywhere.

The strangest result: most of the English listings have no location

Read the last row of that table again. Buildly AI, BuildInADay, ATF WebWorks, SiteIt.ai, 1MillionWebs: no city, no state, no country. That is not our omission. On September 21 ChatGPT gave the location of all eight businesses as N/A and still handed over a Google Maps search link for each one.

Counted across the whole measurement, 25 of the 32 listing slots served on the English query named a business with no stated location, and 0 of the 192 slots served on the Spanish queries did. In Spanish every single card carried a real city: Ciudad de México, Guadalajara, Mérida, Puebla. In English the local mode is mostly placeless.

We do not know the mechanism and we are not going to invent one. What we can describe is what it does to a buyer: they are handed the interface of a local answer, eight nearby-looking options, for businesses the system cannot place on a map. The form promises proximity that the data does not contain.

Who occupies those slots, and how stable is it?

Across the measurement, 224 listing slots were served to a total of 55 distinct businesses: 41 on the Spanish queries and 14 on the English one. That is heavy rotation for eight positions. Nobody owns the set, but a few names recur often enough to be worth naming.

The English side of the table is thin on purpose: 32 slots split across 14 businesses means almost nobody repeats. The most frequent name in the English listings has three appearances. In Spanish the leader has twenty-two. Whatever source feeds the English local mode, it is far less settled.

The same business can come through either door

One case in the data is worth isolating, because it is the closest thing to a control. ApexSites appears as a map listing with no stated location on August 3 and again on August 31. On September 7, the same query in web mode returns apexsites.ai as a website. Same business, same question, two different retrieval paths, two weeks apart.

That kills the tempting conclusion that you must choose between being a listing and being a site. The two modes draw on different sources, and a business that exists cleanly in both gets a candidacy in both. It also means the work is additive, not a bet: content does not help you in listing mode, a listing does not help you in web mode, and the mode is not yours to pick.

Why does ChatGPT answer this way at all?

Because for local queries it is not using its index of pages, it is using business and map data. OpenAI launched the local mode in ChatGPT Search in December 2024, with listings, a map, and the ability to tap a business to call it or open its site. In July 2026, Search Engine Land reported the agreement that brings Yelp's reviews, ratings, photos and business data into ChatGPT.

"The OpenAI partnership adds ChatGPT to Yelp's distribution network, with Yelp branding and links accompanying its content."

Jeremy Stoppelman, chief executive of Yelp, quoted by Search Engine Land, July 23, 2026.

Glenn Gabe, the search consultant who documented the launch of the local mode, framed it by what it does to local businesses rather than by the technical novelty:

"Google Search is incredibly important for local businesses, many have noted that local listings often act as their websites in the SERPs, ranking in the local pack is super important, etc., so ChatGPT now having a 'local mode' is important."

Glenn Gabe, G-Squared Interactive.

Our measurement adds one thing to that framing. Gabe is describing local businesses, the restaurants and dentists the local pack was built for. What the data shows is that the mode now fires for professional services bought remotely: AI consultancies, website builders. Categories where nobody expects to walk in.

What does this change if you sell services in English?

Four things, in the order we would do them, with the honest caveat attached to each.

• Treat the business listing as a candidacy, not housekeeping. Claim the Google Business Profile, pick the primary category that matches what you sell rather than the one that sounds bigger, and fill it completely. This is the entry ticket to the mode. It does not rank you inside it.
• Do the same on Yelp. Since July 2026 that data feeds ChatGPT directly. For a B2B services firm this feels absurd, which is exactly why most of your competitors have not done it.
• Keep publishing anyway. In our own series the web mode still accounts for the large majority of answers, 218 of 246. Abandoning content to chase listings would trade the common case for the rare one.
• Measure your own category before believing any of this. Our 12 queries are our categories. Run yours, on your engines, and store the raw answers. The mode may never fire for you, in which case none of this is your problem.

What this does not fix

Being in the listings is not the same as being chosen from them, and we have no evidence about what happens after the card appears. We measure what the engines return, not what buyers click, and not what they buy. The rotation we see also means a listing slot is not a position you hold, so a week of appearances proves less than it feels like it proves.

There is also a limit worth stating plainly: 28 listing answers is a small denominator, and the English side of it is 4 runs of a single query. The binary pattern is strong because it has zero exceptions in 246 answers, not because the sample is large. We keep running it, and we will say so here if it breaks.

How we work on this at Teckel AI

We build sites that answer engines can read, and we measure which agents actually read them. That is the AI Websites work, from 2,500 USD, and the category itself is explained in what is an agent-ready website. This article exists because the same measurement rig that tracks whether we get cited also produced a finding that is more useful to you than to us. The first action it put on our own list is the Google Business Profile we still have not set up.

If you want your own version of this measurement for your category, the diagnosis is free and we will hand you the raw answers, not a score.

Or reach us directly on WhatsApp: +52 55 3216 1578 · george@teckel-ai.com

Frequently asked questions

Sources

• Teckel AI primary measurement, July 23 to September 21, 2026: 12 fixed queries run on working days against the OpenAI API with web search and Mexico geolocation, google.com.mx through SerpAPI, and the Anthropic API with web search. 246 non-empty answers, stored as raw JSON.
• OpenAI, launch of the ChatGPT Search local mode with listings and a map, December 2024, documented by Glenn Gabe at G-Squared Interactive.
• Search Engine Land, the OpenAI and Yelp agreement bringing reviews and business data into ChatGPT, July 23, 2026, including the quoted statement from Yelp chief executive Jeremy Stoppelman.
• Muck Rack, Generative Pulse, May 2026, analysis of more than 25 million links cited by ChatGPT, Claude and Gemini across 17 industries: 84% of AI citations are earned media on third-party sites rather than the brand's own site.

Keep going: how to get ChatGPT to recommend your business, what an agent-ready website is, and the Spanish edition of this research with the Mexican listing detail: cuando le preguntas a ChatGPT por un proveedor en México.