Local GEO: how ChatGPT and Google AI answers change by neighborhood

We asked ChatGPT for an orthopedic clinic near three Seoul stations and three New York neighborhoods, and every place got its own clinics. How to measure local AI visibility, place by place.

What local GEO is

Local GEO is generative engine optimization for questions about places: getting a clinic, a restaurant or a store into the answer when people nearby ask an AI assistant for a recommendation. Questions like “dentist near me”, “where to eat lunch” or “a pharmacy open now” depend on where the person is, and ChatGPT, Gemini and Google answer them for the user's location.

So the same question has a different answer in every neighborhood. In the test below, one question asked from three stations brought three different sets of clinics. Measuring local visibility means asking from each place that matters to the business.

The test: one question from three places

In October 2026 we sent one ChatGPT prompt from three stations in Seoul, with country set to KR and location set to each station. Each answer recommended clinics around its own station, with walking distances:

One clinic chain shows the effect at branch level. From 홍대입구역 the answer recommended the chain's 홍대 branch, 45 m away. From 강남역 it recommended the same chain's 교대역 branch, 1.1 km away.

New York gave the same result. “orthopedic clinic near me”, asked from Hoboken, Times Square and Long Island City, brought clinics near each place with their street addresses:

Each table shows two of the clinics each answer recommended.

Google AI Mode, AI Overviews and Gemini

Google answers for the precise location a phone or a browser shares with it. In tests with a location set, AI Mode named the requested place in all 50 answers, across eight Korean cities and ten places in the United States, Japan, Germany and the United Kingdom. AI Overviews reflected the place in all 18 located answers, and Gemini answered for the requested city in every completed answer from Germany and the United Kingdom. The changelog has the details.

Google Search results take the same location, so one address can be checked across the search results, the AI Overview and the AI Mode answer.

How to ask an AI engine from a place

Set payload.location on the task. Write the place the way people there write it, up to 200 characters, inside the task's country: a station, a neighborhood, a district or a full street address. For a name that several cities share, add the district or the city.

curl -X POST https://api.querying.ai/v1/async/task \
  -H "Authorization: Bearer $QUERYING_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "taskType": "CHATGPT",
    "payload": {
      "prompt": "근처 정형외과 추천",
      "country": "KR",
      "location": "강남역"
    }
  }'

The result carries the place the engine answered for in location:

{
  "location": {
    "query": "강남역",
    "name": "강남역, 강남대로, 역삼1동, 강남구, 서울특별시, …",
    "latitude": 37.49993,
    "longitude": 127.02697,
    "precision": "railway"
  }
}

• query: the text you sent.
• name: the place it matched, written the way the market writes it.
• latitude and longitude: the point the engine answered for.
• precision: the kind of place matched, such as railway for a station, suburb for a neighborhood or city.

Location works on CHATGPT, GEMINI, GOOGLE, AIMODE and GOOGLE_SERP tasks, and a located task costs the engine's usual credits. The API checks the request when you submit it: other engines return LOCATION_UNSUPPORTED and a place that matches nothing returns LOCATION_NOT_FOUND, so every task that runs answers for the place you chose.

How to measure local AI visibility

1. List the places your customers ask from: each branch's address, the nearest station and the neighborhoods you serve.
2. Write the local questions: “near me” questions, the service with the area's name, and what a first-time customer asks.
3. Send every question from every place to every engine. POST /v1/async/task/batch takes up to 500 tasks per request.
4. Score each answer for your branch's name and your competitors' names, and keep the place with every result.
5. Repeat on a schedule and compare place by place, so a gap near one branch shows up in that branch's number.

Local rates need enough answers too. Five questions on two engines, run three times, give 30 answers per place, the floor for reading a rate. The sample size guide has the numbers, and the AI visibility guide covers the scoring.