Geographic Dependency of ChatGPT AI Output: Independent Test Results #1

Does geolocation affect AI output results?

With the growing habit of turning to ChatGPT instead of Google, we, as SEO specialists who understand the principles of search result formation well, had questions about how exactly LLM logic works. First and foremost, we were interested in whether the user’s location affects the content of responses. It turned out that there are few such public studies.

People from our team received different answers from ChatGPT for the same query, and it seemed to us that these answers took into account the GEO of those who asked. This became the trigger for launching our entire series of experiments – we didn’t understand how AI output works and wanted to figure it out.

Hypothesis #1: AI responses depend on user location

Before starting the experiment, we formulated a key hypothesis based on observations: ChatGPT may adapt responses depending on the user’s geolocation.

The logic is simple: if search engines have long used IP addresses and other signals to personalize results, it’s quite possible that language models apply a similar approach.

We assumed that the greatest differences would appear in local queries, but we were ready to test less obvious cases as well to understand the scale of geoposition influence on AI output.

How we tested geographic dependency of AI responses: Research methodology

To check whether user location affects AI output results, we built a maximally simple but very illustrative experiment.

Input conditions:

• Using 2 logged-in ChatGPT accounts (with GPT-5 or GPT-5 Plus version)
• Using 2 incognito sessions to avoid query history influence (GPT-4o mini version)
• Connecting 5 proxy servers with different geolocations
• Selected GEO: English-speaking only – USA, Britain, Canada, Australia, New Zealand.

For 1 query: 4 results for each GEO, total 20 results per query.

PS: Actually, following this logic, several people conducted the experiment (so we have a much larger sample of responses), but the conclusions were the same for all participants.

Categories of queries we prepared for testing:

1. Obviously geo-dependent — for example, where rent a car, Should I take an umbrella today?, dentist near me and others.
2. Not obviously geo-dependent — for example, purchase gaming chair, buy air conditioner, where to buy wetsuit and others.
3. Definitely geo-independent — for example, best coffee shop in London, Apple store in Tokyo, vegan restaurants in Berlin and others.
   

Other testing features worth noting:

• For each aforementioned category — 10 unique queries (Total – 30 queries)
• Each query was launched in a new chat to avoid influence of previous responses
• All results were recorded immediately in an Excel table, without additional clarifications to AI
• Did not change syntax or question structure between GEO, did not add language markers.

How geolocation affects AI output: Test results

In principle, if you ask ChatGPT directly: “Do your responses depend on user geolocation?”, it will say absolutely not 🙂

But in reality, it monitors your session data, which is why we can observe the dependency of responses on GEO.

It’s worth noting that this is relevant for queries from accounts where you allowed it to receive this information and it has dialogue history with you. If questions are asked in incognito sessions, ChatGPT almost never reports geo-dependency, asks clarifying questions if location information is important for the answer. Although in reality, results sometimes still differed.

For many results from accounts, ChatGPT immediately loaded a map:

Query from account: UK (logged in)

The map doesn’t look like standard Google Maps or Bing maps, but it’s interesting where exactly these results are pulled from (but this is a topic for future tests).

For the same query in the same GEO in an incognito session, the result was more restrained:

Query from account: UK en (incognito)

Also, one of the experiment participants asked ChatGPT to respond as if for another region, and it turned out this works quite well.

Sometimes responses differed under absolutely identical test conditions:

After analyzing all hundreds of query results, we formed the experiment conclusions:

1. Geographic dependency of responses
   • The analysis showed that ChatGPT’s output is geo-dependent for queries that are geo-dependent or not obviously geo-dependent. 
   • For geo-independent queries, the results are independent of the user’s location.
2. Location determination
   • ChatGPT can determine a user’s approximate location, including through session data, conversation history, and IP address, which affects the responses to location-based queries.
3. Usage without account
   • The responses for a paid user logged into an account turned out to be significantly better than in free and incognito mode (they include accurate maps, and more specific and accurate answers).
   • When accessing ChatGPT without an account, the system usually gives simple clarifying answers.
   • In some cases, the model correctly determined the user’s location and gave a geo-dependent answer even in incognito mode.
4. Other features
   • Geo-dependency does not manifest itself in the same way in all queries. For example, for some queries (like about a pizzeria), GPT correctly identifies the city the user is in, but can additionally request a more precise location. And in some cases, it immediately offers results within the city.
   • Even when using the same IP address and an absolutely identical prompt, the system can generate excellent answers.
   • Authorization is not always required to determine the location.
   • If you ask to answer as if I am in region A, it works. The results will sometimes be absolutely the same, and sometimes they will be slightly different.
   • In the case of the query “emergency plumber in Sydney”, even in incognito mode, GPT almost always gave specific answers with contact information for companies, which leads to the idea that the “urgency” in the query affects the behavior of the model.
     

ChatGPT clearly demonstrates the ability to adapt responses to user geolocation only in cases of queries where this information is relevant, while for general queries, geo-dependency is mostly absent.

You can review the queries and received responses in Queries and their output:  Experiment #1

Conclusions and further AI output tests

Результаты эксперимента доказывают гипотезу геозависимости при формировании The experiment results prove the hypothesis of geo-dependency in AI output formation, which has a direct impact on local SEO and business promotion strategies.

During testing, we had new questions that we’d like to check in future experiments. In particular, it’s interesting to investigate how exactly AI response is formed in cases when it displays a map, since this doesn’t look like Google Maps or Bing Maps, as seen in the screenshots above.

The logic of branded results formation also requires separate analysis: it’s evident that search is used, however, ranking principles differ from classic Google or Bing.

More to come!

Author

Dmytro Kovshun

Dmytro Kovshun is the founder of Luxeo Team – an SEO Outsourcing Company. As a leading specialist in the industry, he is recognized as an expert in SEO promotion of websites. With years of experience and a deep understanding of the field, Dmytro continues to drive success and innovation in SEO strategies, helping businesses achieve their online goals.