TL;DR
- An AI visibility score measured from one place is a score for that place. In a University of Toronto preprint (arXiv 2610.24081, version 1, 16 October 2026, not peer reviewed), ChatGPT, reached through its API, followed the city an application declared to it, and Gemini's API followed the country of the network the request came from.
- The limits are wide. English questions, 15 US cities plus London, Toronto, Sydney and Dublin, small models reached through their APIs and not the consumer apps, and data-center VPN exits. No Quebec city and no French question was tested. None of these percentages is a Quebec result.
- In that preprint, the businesses ChatGPT named changed with the place. For "near me" questions, a request declared in one US city recommended 8 percent of the businesses the same request recommended in another city, against 32 percent when it was repeated in the same city. Monitoring one city captured 9 percent of the businesses recommended across 15.
- Repeating a question matters even when the place does not move. On the paper's other count, two identical near-me requests to ChatGPT had 23 percent of the names they recommended in common.
- Our own check was 16 requests on one trade, too small to give a rate. The consumer ChatGPT interface we reached answered a question with no town in it for a different city on each run, and naming Sherbrooke in the question returned largely the same stores every time.
- Three questions to put to anyone who hands you an AI visibility report: which city was it measured from, on which engine and through which channel, and how many times was each question asked?
A business owner who is shown an AI visibility score is usually shown one number: you appear in so many percent of the answers. In our post on reading AI visibility scores we listed six things a score has to state before it can be checked: the questions, the engines, the date, the number of runs, whether it counts mentions or citations, and where the raw answers are. This post adds a seventh, and it is the one we had been treating too loosely ourselves: where the question was asked from.
The source is a short preprint by Mahe Chen and Nick Koudas of the University of Toronto, the same two authors as the 90-day study we covered in what drives AI citations. Its limits come before its figures. It is version 1 and not peer reviewed. We read it in October 2026 and have not re-run it. The questions are in English and were written by the authors. The places are 15 US cities and four cities in other English-speaking countries. The engines are two small models reached through their APIs (gpt-5.4-mini and gemini-3.1-flash-lite), which need not behave like the ChatGPT and Gemini apps a customer opens on a phone. Every figure below is the authors' own unless we label it as ours.
What did the study test?
The study tested whether the same question, asked from different places, gets different cited sources and different recommended businesses, and which signal of location each engine reads. The authors wrote 300 questions in ten commercial topics (home services, health care, finance, insurance, telecom, dining, retail, cars, travel, public services). No question names a place. They come in three groups of 100:
- Near me. Questions that ask for something nearby, such as "best plumber near me".
- Place-dependent. Questions whose answer depends on where you are (prices, laws, providers, weather), such as "how much is sales tax on a car".
- Not local. Questions with nothing local in them, such as "best robot vacuum".
Two experiments follow, 31,200 answers in all.
- Declared location (24,000 answers). Every question went to ChatGPT through the API with its web search tool, twice in each of 20 settings: each of 19 cities passed as the tool's approximate user location, and no location at all. All requests left from one network. A second pass put "Search the web:" in front of each question so that sources exist for every question.
- Network location (7,200 answers). Every question went to ChatGPT and to Gemini (with Google Search grounding) with no declared location, twice each, in six sessions: the authors' own connection in Canada, VPN exits in New York, Los Angeles, London and Sydney, and the Canadian connection again three days later.
The measure is overlap: for one question, how many of the websites cited in one answer are also cited in another. Two identical requests (the authors call them twins) give the noise. The paper's localization index is the overlap between twins minus the overlap between two places. Zero means that place changes nothing beyond what repeating the request already changes.
Where do ChatGPT and Gemini get your location?
In this study, run through the two engines' APIs, ChatGPT and Gemini read location from different signals, and at different levels of detail. ChatGPT's API followed the location the application declared and ignored the network. Gemini's API ignored declared coordinates in a pilot and followed the network, to the country and no further. Our own small check, further down, saw the consumer ChatGPT interface behave differently.
| What the study observed | ChatGPT (API, web search tool) | Gemini (API, Google Search grounding) |
|---|---|---|
| Signal it followed | The location the application declares | The network the request comes from |
| Finest level observed | City | Country |
| Signal it ignored | The network: between network countries the index was 0.00, 0.01 and 0.01 | Declared coordinates: no effect in a pilot |
| Place written into its own searches | In 100% of searches for near-me questions and 48% for place-dependent ones | The country in 56% of near-me answers, never a city |
| With no location declared | Asked for a ZIP code or city on 63% of near-me questions. When it searched, it answered for an unspecified user in the United States, from national sources | Used the country of the network |
The authors sum it up in one line: "to ChatGPT every other city is a different place; to Gemini, only another country is." Distance did not soften it. For near-me questions, ChatGPT's sources differed as much between US cities less than 1,000 km apart as between cities more than 3,000 km apart (an index of 0.13 against 0.14). Between New York and Los Angeles, Gemini's index was 0.00.
Inside ChatGPT, the place got in by two routes. The model wrote it into its own searches: asked for "credit unions in my area" with Boise declared, it searched for credit unions near Boise, Idaho. And the search service behind it localized results without being asked: where neither answer's searches named the place, the sources still differed (an index of 0.17). The paper's example is a user in Boise asking when to plant tomatoes and getting Oregon State University's extension service.
For anyone who buys or sells monitoring, the consequence is in the authors' discussion: "A tool that sets only a location parameter measures ChatGPT for that city but Gemini for the country it runs from; a tool that sets neither measures ChatGPT for an unspecified US user and Gemini for its own country."
How much do the sources change from one place to another?
In the preprint, the websites ChatGPT and Gemini cited for local questions changed with the place, and on Gemini they changed between countries even for questions with nothing local in them. The table gives the paper's localization index for cited websites. The further a value is from zero, the more of the difference comes from the place and not from chance.
| Localization index of cited websites | Near me | Place-dependent | Not local |
|---|---|---|---|
| ChatGPT, two declared US cities | 0.14 | 0.17 | 0.00 |
| ChatGPT, two declared cities in different countries | 0.28 | 0.43 | 0.03 |
| ChatGPT, two network countries, no declared location | 0.00 | 0.01 | 0.01 |
| Gemini, two network countries | 0.43 | 0.40 | 0.23 |
A few details sit behind that table.
- Gemini's last cell. An index of 0.23 on not-local questions means that a question like "best robot vacuum" drew different sources depending on the country the request came from. A brand measured from a US server and sold in Canada is the obvious case (our example).
- The topic matters. Among place-dependent questions, the effect between two US cities was strongest for public services (0.57, each state's government sites) and weakest for telecom (0.02, national sources). Those rest on ten questions per topic, and the authors call them indicative.
- A declared location changes whether ChatGPT cites at all. With one, 64 to 73 percent of its answers carried sources. Without one, 46 percent did.
Do the businesses ChatGPT recommends change from city to city?
Yes. In the study, the businesses an answer recommended moved with its sources, and for near-me questions two cities had almost none in common. A near-me answer from ChatGPT named 4.3 businesses, brands or products on average. The paper reports two counts of those names. It does not define how they differ. We read them as two different measures, so we give both.
| ChatGPT, declared location | Same place, request repeated | Two different places |
|---|---|---|
| Near me: names two answers have in common | 23% | 4% |
| Near me: share of one request's businesses that another request also recommends (15 US cities) | 32% | 8% |
| Place-dependent: the same share (15 US cities) | 34% | 31% |
For near-me questions, on either count, two repeats in the same place had several times as many names in common as two different places did. For place-dependent questions the two were close (34 percent against 31). The paper adds that 96 percent of the businesses recommended for near-me questions appeared in only one of the 15 US cities. With two requests per city, part of that is repetition and not place (our caution). Its own examples show what that looks like.
| Engine and question | Place | First three businesses named in both repeats |
|---|---|---|
| ChatGPT, declared city: "credit unions in my area" | New York | New York University FCU, Neighborhood Trust FCU, Brooklyn Cooperative FCU |
| Boise | Idaho Central CU, CapEd FCU, Connections CU | |
| Miami | Educational FCU, University CU, Miami Postal Service CU | |
| London | London Mutual CU, London Capital CU, Your Community Bank | |
| Gemini, network country: "which store has the cheapest groceries" | Canada | No Frills, Walmart, Food Basics |
| United States, East | Aldi, Walmart, Costco | |
| United States, West | Aldi, Walmart, WinCo Foods | |
| United Kingdom | Aldi, Lidl, Tesco | |
| Australia | ALDI, Coles, Woolworths |
How many cities does a monitoring tool have to cover?
In the preprint, for near-me questions, a tool that monitors one US city sees a small part of what ChatGPT recommends elsewhere, and each added city brings mostly new names. The authors took everything ChatGPT recommended across the 15 US cities and counted how much of it a tool would capture by monitoring fewer of them.
| Cities monitored (of 15) | Near-me questions: share of recommended businesses captured | Not-local questions: share captured (reference) |
|---|---|---|
| 1 | 9% | 19% |
| 5 | 36% | 53% |
| 10 | 69% | not stated |
| 15 | 100%, by definition | 100%, by definition |
Read the right-hand column before drawing a conclusion from the left. Place has no effect on not-local questions, and one city still captured only 19 percent of the names. That shortfall comes from repetition alone: thirty requests name far more businesses than two do. So the table carries two separate lessons. A tool has to ask from the right place, and it has to ask more than once.
For a business that serves one town, the first lesson is good news. It does not need 15 cities. It needs its own, set on purpose, because in the authors' words "no city stands in for another, not even a nearby one". A business that serves several towns has several markets to measure.
What does this not tell you?
This study does not tell you how often ChatGPT or Gemini names your business in your town.
- Anything about Quebec, or about French. Toronto is one of the 19 declared cities, and the authors' own Canadian connection is one of the network locations. No Quebec city, no province and no French question was tested. Whether ChatGPT treats a French question with a Quebec town declared the same way is unmeasured. The per-city result makes it plausible. It does not show it.
- Whether two neighbouring towns differ. The closest pair of cities in the study is 293 km apart. Towns closer together than that were not tested.
- How the consumer apps behave. The engines were reached through their APIs, with small models. A check with the full-size gpt-5.4 on 60 questions at six locations (720 answers) localized at least as strongly, and it covers ChatGPT only. The apps were not tested, and an app on a phone may read signals the API does not see.
- How Gemini behaves on a home or office connection. Four of the five network locations were VPN exits on data-center addresses, and the fifth was the authors' own connection. The authors note that Gemini may localize more finely on residential connections.
- Anything about other engines. Perplexity, Google's AI Overviews and AI Mode were not measured.
- Which page wins. Overlap is counted on websites, not on pages or on position in the answer.
- How long it holds. The measurements describe the engines in October 2026. The authors say they release the questions, code and answers. The version we read gives no link to them.
What did our own 16-request check on ChatGPT show?
Our check is not the paper's experiment, and it is far smaller. On 20 October 2026 we sent 16 requests through a commercial scraping service that collects answers from the consumer ChatGPT interface with web search forced on. That service can set a country and cannot set a city. We asked for hardware stores, in English and in French, with no town and then with Sherbrooke named, twice per combination.
| Question | Country set to United States (two runs) | Country set to Canada (two runs) |
|---|---|---|
| "best hardware stores near me" | Answered for Stockton, California, then for Paxtang, Pennsylvania. No local business in common | Answered for Gatineau, Quebec, then for Winnipeg, Manitoba. No local business in common |
| "best hardware stores in Sherbrooke, Quebec" | 6 of 6 businesses in common | 5 of 6 in common, the same six Sherbrooke stores as the US runs |
| The near-me question asked in French: "Quelles sont les meilleures quincailleries près de moi?" | Answered in French for Greensboro, North Carolina, then for Phoenix, Arizona. No business in common | Answered in French for Toronto, then for Greater Sudbury, Ontario. No business in common |
| The Sherbrooke question asked in French: "Quelles sont les meilleures quincailleries à Sherbrooke?" | 7 of 8 in common | 5 of 6 in common. The same five Sherbrooke stores led all four runs |
What came out of it, with the caution sixteen requests deserve:
- The interface we reached followed the service's country setting. With no town in the question, ChatGPT answered for what it called the user's "estimated location", a different city of the chosen country on each run. That behaves like a location read from the network. We did not see how the service applies its setting. The paper found that the API ignores the network. The two are different products, which is the paper's own first limitation.
- Asking in French did not move the answer to Quebec. Neither of the two French runs set to Canada landed in Quebec (Toronto, then Greater Sudbury). The one Quebec answer, Gatineau, came from an English run. Two runs cannot show that language never matters.
- A town in the question pinned the market. With Sherbrooke named, both country settings and both languages returned largely the same stores, with differences too small to separate from the noise between two runs.
We also sent the near-me question to gpt-5.4-mini through an API route with no location setting. It asked for a ZIP code or a city both times. Told to guess, it picked New York, then Chicago, which is consistent with the default the paper describes. We ran a handful of Gemini requests the same day. They were too few and too mixed to report.
What does this suggest about getting named locally?
The preprint's authors suggest that for local questions, the pages that count are the ones an engine finds when it searches with the town's name in the query. With a declared location, 79 percent of ChatGPT's near-me answers in US cities cited a source whose URL named the city or the state, and 92 percent of those in the other countries cited a source on that country's domain. A control that counted sources naming one of the study's other locations stayed below 0.2 percent.
The authors read this as a business becoming visible "through pages that the engine finds for place-specific queries, such as local directories, review sites and the sites of local institutions, rather than through national rankings". That is their interpretation of the pattern. The study did not test any page-level change. It does line up with the citation and listing work that local SEO has always involved, and with the third-party mentions we put early in a mandate after the 90-day study. We would not sell it as a measured lever.
What should you ask before trusting an AI visibility report?
Ask where it was measured from, on which engine and through which channel, and how many times each question was asked. The six items in our earlier checklist still apply. These three sharpen two of them and add the missing one.
| Ask | What a usable answer looks like | What a missing answer means |
|---|---|---|
| Which city was it measured from? | A named town for each market you serve, and how it was set: declared to the engine, written into the question, or taken from the tool's network | By the paper's default-location result, a ChatGPT measurement through the API with no location describes an unspecified user in the United States. By the authors' conclusion, a nearby city is not a stand-in for yours (their closest pair was 293 km apart) |
| Which engine, and through which channel? | Each engine by name, app or API, with the location signal used for each one | One "location: Canada" setting may reach one engine and not the other. In the study, ChatGPT's API read the declared location and Gemini's read the network |
| How many times was each question asked? | At least two runs of every question, with the overlap between runs shown beside any comparison | A difference between two businesses, two towns or two dates that is smaller than the gap between two identical runs is noise |
You can run a rough version yourself. Ask the question with your town in it, in a new conversation, several times, and count how often each business is named. If the person running the check is not in your town (an agency in another city, a tool on a server in another country), the town has to be set on purpose, and the report should say how.
What we changed
We changed our own measurement procedure on 20 October 2026, after reading this paper and running the check above. None of it guarantees a mention.
- The old line was too loose. Our procedure used to say that Quebec questions run in French "with a Canadian location". By this paper, a country is too coarse for ChatGPT, and a location parameter is the wrong channel for Gemini. Measurements we made under that line set a country at most, and we do not present them as city-level results.
- The location goes in the report header. For each engine, the header now says how the place was set (declared city, place in the question, country filter, or not set), next to the number of questions.
- Every live question runs at least twice. The overlap between the two runs is reported as the noise floor beside any comparison. It is the same no-change rerun we ask for in rewriting a page for AI answers.
- The town goes in the question. We no longer send a near-me question, or any question with no town in it, through a tool that can set only a country.
- "Not measured" is an allowed answer. Where no instrument we have can set the market, the row says so.
If someone hands you an AI visibility report that does not name the city it was measured from, it may describe someone else's market. Ask for the city, the engine and the number of runs. We sell this kind of measurement ourselves, so we have an interest here, and the same three questions apply to ours. Where a measurement sits among the rest of the work is laid out in the AI visibility funnel and on our SEO audits page.