Same near-me question, two US cities. ChatGPT named 8% of the same businesses.

Mahe Chen and Nick Koudas posted the paper on arXiv on Oct 16, 2026. They wrote 300 questions in ten commercial topics, from home services to retail to travel. A third were near-me questions like “best plumber near me”. A third depended on the place, like prices and laws. A third were not local, like “best robot vacuum”. No question named a place.

They asked ChatGPT with 19 declared locations: 15 US cities, plus London, Toronto, Sydney and Dublin. They also asked ChatGPT and Gemini from five network locations. Every question was asked twice, to measure the noise between identical requests. In total that is 31,200 answers.

According to the paper, ChatGPT takes its location from what the application declares, down to the city. It ignores the network the request came from. It wrote the place into 100% of its searches for near-me questions and 48% for place-dependent ones.

Even when the searches did not name a place, the sources still differed by city. The paper gives an example: for a user in Boise, ChatGPT cited Oregon State University’s extension service on when to plant tomatoes.

Gemini did the opposite. It ignored the declared location and followed the country of the network. New York and Los Angeles got the same sources. Other countries got different ones, including for questions that name no place.

The paper’s example is “which store has the cheapest groceries”. The first three names in the answers from the US East network were Aldi, Walmart and Costco. From the UK, they were Aldi, Lidl and Tesco. These are output from the paper’s test, not statements about those stores.

For near-me questions, a request declared in one US city recommended 8% of the businesses that the same request recommended in another of the 15 cities. Repeating the request in the same city gave 32%. And 96% of the businesses recommended for near-me questions showed up in only one of the 15 cities.

So how many places does monitoring need? The authors’ curve says one city captured 9% of the businesses ChatGPT recommended across the 15 cities, five cities 36% and ten cities 69%. It does not level off, because every added city brings mostly new businesses.

The authors list their limits. They used the engines’ APIs and small models, which need not behave like the consumer apps. The check on a full-size model covers ChatGPT only. VPN exits are data-center addresses, and Gemini may localize more finely for home connections. The questions are in English, and the results describe October 2026.

The size of the effect also depends on the question. For not-local questions, ChatGPT’s sources barely depended on the declared city. The 8% and 32% figures are for near-me questions, which matter most for brands with shops. For an online-only brand, the Gemini result is the more relevant one, because Gemini changed most of its sources by country even for questions that name no place.

PYMNTS Intelligence reported on Oct 13, 2026, that nearly 61 million US consumers now begin their online retail research with AI assistants. That comes from an August 2026 survey of 2,191 American adults. The same report says 45% of AI shopping users were introduced to online stores they had not visited before.

Those shoppers are not all asking from your office. A single check from one desk shows what one shopper in one place might see.

Five steps.

Shopthru.ai is an AI agents team for consumer brands. It asks ChatGPT, Gemini, Google AI, Perplexity and Copilot what your shoppers ask, finds where AI gets your products, prices or policies wrong, and drafts each fix into an approval inbox. It ships only on your yes, then checks again. Instagram and Amazon are next.