The answer
Yes. Location changes AI search answers for local questions. A University of Toronto study of 31,200 answers found that ChatGPT changes its sources and recommended businesses by the city in the request. Gemini changes them by the country of the network. So an AI visibility report from one location shows you only one market.
Does location change AI search answers?
Yes. For a "near me" question, two US cities shared only 8% of the businesses that ChatGPT recommended. Two runs of the same question in the same city shared 32%. Your AI visibility in one city tells you little about the next one, so measure each market on its own.
The study is "Location, Location, Generation" by M. Chen and N. Koudas of the University of Toronto. The researchers asked ChatGPT and Gemini the same 300 questions from many places and ran each request twice: 31,200 answers in total. A/GEO Research summarized the study on November 10, 2026. Every number here comes from that summary.
Which location does each AI engine use?
ChatGPT uses the city that you name in the prompt. Gemini uses the country of your internet connection. Each engine ignores the other signal.
| Engine | Uses | Ignores | How to test it |
|---|---|---|---|
| ChatGPT | The city in the prompt | Where the user connects from | To test a city, name the city in the prompt. A VPN (virtual private network) alone changes nothing |
| Gemini | The country of the connection | The city in the prompt | To test a country, connect from that country. Gemini gave the same sources in New York and Los Angeles |
The study scores each comparison with a localization index. This article shows it as "changes per 100 sources." A score of 43 in 100 means that location changes about 43 of every 100 sources the AI cites, after you remove the normal change between two identical runs. A score of 0 means location has no effect.
How many sources change with location? Pick a question type.
For a question like "best HVAC company near me," test ChatGPT city by city and Gemini country by country.
Rules, prices and services that differ by state or country change many sources in both engines. Give each state or country that you serve its own page.
For general questions, ChatGPT cites the same sources everywhere. Gemini still changes about a quarter of them by country, so global brands must test Gemini in each country.
Sources that change with location, per 100 cited sources, after normal run-to-run change. Data: Chen and Koudas, University of Toronto, via A/GEO Research.
Do you need content for every city?
No. You need content for each city that you serve, not for every city in the study. The researchers tested 15 US cities. That number is their test set, not a target for you.
Two findings make each city that you serve a separate market:
- 96% of the businesses that ChatGPT recommended for near me questions appeared in only one of the 15 cities. The competitors that AI shows in one city are not the ones it shows in the next.
- Distance made no difference. Cities less than 1,000 km (about 620 miles) apart changed 13 in 100 sources. Cities more than 3,000 km apart changed 14 in 100. A page for a nearby city does not cover yours.
For each city that you serve, do two things. Publish a page that names the city and the services that you offer there. Then get listed on the local directories and review sites that AI cites for that city. In US cities, 79% of ChatGPT's near me answers cited a source that names the city or the state.
How many cities do you need to track?
Track each city that brings you revenue. A report from one city shows only 9% of the businesses that ChatGPT recommends across the 15 test cities.
Pick a number of cities below. "100%" means every business found across the 15 test cities. It is the ceiling of the study, not a goal.
How much of the market does your tracking see?
A one-city report misses 91% of the near me market. Do not use it as a national score.
Three cities make a good first test. They still show less than a quarter of the near me market.
Each city that you add shows new businesses. No small set of cities stands in for the rest.
Ten cities show about two thirds of the market. If you serve ten or more cities, track each one.
100% is every business across the study's 15 cities. Your own 100% is the set of cities that you serve.
Share of ChatGPT's recommended businesses across the 15 US test cities. Data: Chen and Koudas, University of Toronto, via A/GEO Research.
Look at the "Not local" line. One city shows only 19%, even though location does not affect those questions. The gap comes from ChatGPT itself: it gives a different answer each time. Run each prompt at least twice, or one run can mislead you.
Which industries change the most by location?
Public services change the most, and telecom and utilities change the least. The numbers are for ChatGPT between two US cities. Each topic rests on ten questions, so treat them as a direction.
| Topic | Near me | Depends on place | What to do |
|---|---|---|---|
| Public services | 16 in 100 | 57 in 100 | Answers cite each state's government sites. If your offer depends on state rules, build a page for each state |
| Travel and leisure | 31 in 100 | 6 in 100 | Near me answers change a lot by city. Local guides and review sites matter most |
| Telecom and utilities | 12 in 100 | 2 in 100 | Answers cite national sources, such as the FCC (Federal Communications Commission). National pages matter more than city pages |
Cars, dining, personal finance, health care, home services, insurance and retail fell between these.
How do you run a three-city AI visibility test?
Pick five buyer questions, run each one twice in three of your cities, and record what the AI cites and recommends. Tick each step as you go. The page does not save your ticks.
Agencies and in-house teams must record the test city and the network country on every report. For the trust signals behind local citations, see the AEO Influence Chain.
What are the limits of the study?
The study used smaller AI models through developer access, so real users can see stronger local effects. A check with the full-size ChatGPT model changed 26 in 100 sources, against 23 in 100 for the smaller model. Other limits:
- The VPN connections came from data centers. Gemini can localize more closely on home connections.
- The questions were in English.
- The results show how the engines behaved in October 2026.
Common questions
Do I need to create content for 15 cities?
No. The 15 cities are the study's test set. Create a page for each city that you serve, and get listed on the local sites that AI cites there.
What does "43 in 100" mean?
It means location changes about 43 of every 100 sources that the AI cites, beyond the normal change between two identical runs. The study calls this the localization index and writes it as 0.43.
Does ChatGPT use my IP address to localize answers?
No, not in this study. ChatGPT used the city in the prompt. Users in two countries with no place named changed at most 1 in 100 sources.
Does Gemini use the city I type in the prompt?
No, not in this study. Gemini used the country of the connection. Users in New York and Los Angeles got the same sources.
Why run each prompt twice?
Two identical runs in the same city shared only 32% of ChatGPT's recommended businesses. One run can show you noise, not your real position.
Next step
Run the three-city test on five money prompts before the end of this month. Then list every directory and review site that the AI cites in each city, and get your business on them.
For help to set up a location-based AI visibility test, talk with Victor Ramayrat.
Sources
- Does Location Matter with AI Search? What 31,200 AI Answers Show, Victor Ramayrat, A/GEO Research, November 10, 2026.
- Location, Location, Generation: Where You Ask Matters for Generative Engine Optimization, M. Chen and N. Koudas, University of Toronto, arXiv 2610.24081, 2026.
- Issue 2: Where You Ask Changes AI Answers, A/GEO Research Weekly, October 19, 2026.
This article is based on the A/GEO Research summary of the study. Every number comes from that summary. The original research belongs to M. Chen and N. Koudas of the University of Toronto.























