Same question, 36 spots: how far AI local recommendations move

We asked ChatGPT and Gemini 15 “near me” questions from 36 spots in New York, Seoul, Austin and Tokyo, twice. Answers 3 km apart shared 4.9% of their places on ChatGPT and 15.1% on Gemini.

What we measured

Questions like “a dentist near me” or “a good place for dinner” are answered for wherever the person asking stands. We wanted to see how far that goes. If a customer walks 3 km, does an AI assistant send them to different businesses, and how much of that change comes from the place and how much from the assistant answering differently each time?

We asked 15 local questions from a 3 × 3 grid of points in four cities, with neighbouring points 3 km apart, on ChatGPT and on Gemini. Then we sent the same 1,080 tasks a second time. That gives 2,160 answers and two comparisons: the same spot asked twice, and different spots in the same round.

Three kilometres is enough to change the answer

Here is one question from the first round. From each of nine spots in New York, ChatGPT was asked “Recommend a dentist near me”. The nine answers named 28 different dentists, and none of them came up at more than one spot. Each pin shows the area ChatGPT said the user was in, its first pick, and how many more dentists the same answer named.

Seoul gave the same picture in Korean. Asked “근처 치과 추천해줘” (recommend a dentist nearby) from nine spots between Oksu-dong and Irwon-dong, ChatGPT named 31 different clinics, and none of them came up at more than one spot.

Across all four cities and 15 questions, the overlap falls as soon as the spot moves. Overlap here is the share of places named by both answers out of all places named by either, averaged over every pair of answers:

ChatGPT varies from one try to the next, so the same spot asked twice shared 35.1% of its places. Moving 3 km cut that to 4.9%, and 85.6% of the pairs from different spots had no place in common. Gemini answers one spot more consistently, at 68.5%, and drops to 15.1% at 3 km. On both engines the place explains most of the change.

Which questions change the most

Questions about independent businesses move the most: dinner, coffee, hair salons, dentists and real estate agents stay at 1.1% or less on ChatGPT. Pharmacies, urgent care and bank branches hold more in common, because chains and national banks have a branch in every neighbourhood and both engines name the brand. We count a place by its name with the branch label removed, so “Chase Bank, Hunters Point” and “Chase Bank” are one place.

At several spots ChatGPT answered the lawyer question with a question of its own, asking which kind of lawyer the user needed. Those answers named no place and are left out of the overlaps.

City by city

Dense cities change the most. In Seoul and Tokyo, ChatGPT answers from different spots shared under 2% of their places, and in Tokyo Gemini's places sat a median 0.5 km from the spot. Austin spreads out: Gemini reached a median 2.6 km for its places and named the same dentists and coffee shops across downtown, so its answers from different spots shared 27.8%.

ChatGPT and Gemini pick different places from the same spot

At the same spot in the same round, Gemini named 30.8% of the places ChatGPT recommended. Gemini reads its places from Google Maps and returns each one with coordinates, so the distances above come from Gemini. ChatGPT searches the web: it rewrote “a dentist near me” into searches such as “best dentists Upper West Side Manhattan NYC reviews” and “잠실동 송파구 서울 치과 추천 후기 진료시간”, naming the neighbourhood the point falls in.

For a business, that means two separate questions: whether it shows up near each of its locations, and on which engine. A national visibility rate averages every neighbourhood into one number and hides both.

How we collected the data

All 2,160 tasks were sent through the querying.ai API between 18:53 and 18:58 UTC on October 10, 2026 (October 11 in Seoul and Tokyo), and every task completed. Each grid is one request to POST /v1/async/task/grid, which runs the same task from each point of the grid with that point as the device location:

{
  "taskType": "CHATGPT",
  "payload": {
    "prompt": "Recommend a dentist near me",
    "country": "US"
  },
  "grid": {
    "center": {
      "latitude": 40.7484,
      "longitude": -73.9857
    },
    "size": 3,
    "radiusMeters": 3000
  }
}

ChatGPT's places come from the entities list of each answer and Gemini's from its map cards, which carry coordinates. Questions were asked in each city's language with country set to US, KR or JP. 35 ChatGPT answers and 13 Gemini answers named no place, and every overlap leaves those pairs out. Google AI Mode answers carry their places inside the text, so this study covers ChatGPT and Gemini.

The data is public. answers.csv has one row per answer with its point, the places, the cited domains and ChatGPT's search queries, and places.csv has one row per recommended place with its rank and, for Gemini, its coordinates and distance. Every number on this page is recomputed from those two files.

Measure your own locations

Put the grid center on a branch, pick a radius that covers the area its customers come from, and send the questions those customers ask. Run each grid twice or on a schedule, so the repeat overlap separates the engine's own variation from the effect of the place. The local GEO guide covers single locations, and the sample size guide covers how many answers a rate needs.