Where you ask changes which sources OpenAI and Gemini API models cite, a preprint finds

Analysis · Preprint, not peer-reviewed

Direct answer: In this preprint, where a question came from changed which sources the two API models cited for local questions, and each model took its location from a different signal: OpenAI’s from the location the application declared, Gemini’s from the network’s country. The authors conclude that monitoring from a single place misses much of this variation.

What happened: Mahe Chen and Nick Koudas of the University of Toronto posted “Location, Location, Generation: Where You Ask Matters for Generative Engine Optimization” to arXiv on Oct. 16, 2026 (version 1, 18:07 UTC) [1]. The four-page paper is a preprint; its arXiv listing gives no journal or conference publication. The same authors wrote the preprint on retrieval and page tactics that we covered on Oct. 28.

How was the study designed?

The paper labels the engines “ChatGPT” and “Gemini,” but the authors report testing API models, not the consumer apps: OpenAI’s gpt-5.4-mini through the Responses API with the web search tool, and Google’s gemini-3.1-flash-lite with Google Search grounding [1]. This article calls them the OpenAI model and the Gemini model.

• Questions. 300 questions in ten commercial topics, none naming a place: 100 “near me” questions (“best plumber near me”), 100 whose answer depends on the place (“how much is sales tax on a car”) and 100 not local (“best robot vacuum”) [1].
• Declared location. Each question went to the OpenAI model twice with each of 19 cities set as the tool’s approximate user location (15 US cities plus London, Toronto, Sydney and Dublin) and twice without a location, all from one network. A second arm prefixed “Search the web:” so that every answer had sources. This produced 24,000 answers [1].
• Network location. With no declared location, both models answered every question twice in six sessions: the authors’ own connection in Canada, VPN exits in New York, Los Angeles, London and Sydney, and the authors’ own connection again three days later. This produced 7,200 answers [1].
• Measure. A “localization index”: how much two identical requests share their cited domains, minus how much requests from two different places share them. Zero means location had no effect beyond run-to-run noise [1].

What did it find?

With a declared location, the OpenAI model’s sources for near-me questions differed by 0.14 on the localization index between two US cities and by 0.28 between cities in different countries; for place-dependent questions, by 0.17 and 0.43. For not-local questions the values were 0.00 and 0.03 [1].

The Gemini model’s sources differed between network locations in different countries by 0.43 (95% CI 0.40 to 0.46) for near-me questions and 0.40 for place-dependent ones, and by 0.23 (CI 0.20 to 0.27) for not-local ones [1]. Between New York and Los Angeles its index was 0.00 (CI −0.03 to 0.04). Its search queries named the country in 56% of near-me answers but never a city [1].

The OpenAI model ignored the network: between network locations in different countries its index was 0.00, 0.01 and 0.01, within noise [1]. It wrote the declared place into 100% of its searches for near-me questions and into 48% of its searches for place-dependent ones. Where neither of two answers’ queries named the location, its sources still differed (index 0.17, CI 0.09 to 0.25), which the authors attribute to the search backend localizing results. For near-me questions, its localization did not fade with distance: 0.13 between US cities less than 1,000 km apart and 0.14 between cities more than 3,000 km apart [1].

A declared location also changed whether the OpenAI model answered with sources at all. With one, it cited sources in 64% to 73% of its answers, against 46% without; without a location, it answered 63% of near-me questions by asking for a ZIP code or city [1].

What changed for the businesses recommended?

A language model extracted the businesses, brands and products each answer recommended; in the authors’ hand check, 29 of 36 extractions were fully correct [1]. A near-me answer from the OpenAI model named 4.3 of these on average. Two identical requests shared 23% of the names; requests with two different declared locations shared 4% [1].

For near-me questions, 96% of the recommended businesses appeared in only one of the 15 US cities. Monitoring one city captured 9% of the businesses recommended across all 15, five cities 36% and ten 69%; for not-local questions, one city captured 19% and five 53% [1].

What the study cannot show

• Preprint. arXiv lists version 1 only and no journal or conference publication [1].
• API models, not apps. The authors write that they used “the engines’ APIs and small models, which need not behave like the consumer applications,” and that the full-size model check (gpt-5.4, 720 answers) covers OpenAI only [1].
• Network locations. The VPN exits are data-center addresses and cover five network locations; the authors say Gemini “may localize more finely for residential connections” [1].
• Questions. All questions are in English and were written by the authors. With ten questions per topic and group, the authors call the topic-level estimates “indicative” [1].
• Recommendation counts. These rely on language-model extraction that was fully correct in 29 of 36 hand-checked cases [1].
• Timing. “The measurements describe the engines in October 2026” [1].
• Data. The paper says “We release the questions, code and answers,” but neither the arXiv listing nor its HTML version gave a repository address when we checked on Oct. 30, 2026.

Why it matters

Our view: For anyone tracking AI visibility, the paper treats location as a variable to record and hold fixed, and shows that the two models read it from different places. The authors write that “the location must be set on the channel the engine uses: a declared location for ChatGPT, a network location for Gemini,” and that comparisons over time or between engines “are valid only with the location held fixed” [1]. Whether the consumer apps behave the same way is untested here.

What we don’t know yet

• Whether the ChatGPT and Gemini apps, or Google’s AI Overviews and AI Mode, localize the same way. The study tested API models only.
• Whether Gemini localizes below country level on residential connections.
• Whether the results hold for questions in other languages.
• Whether peer review or a later arXiv version changes the estimates.

What to do now

1. When you compare AI visibility results across tools, dates or engines, record both the declared location and the network location used, and keep them fixed.
2. For local businesses, the authors suggest visibility on near-me and place-dependent questions comes through pages the engine finds for place-specific searches, “such as local directories, review sites and the sites of local institutions, rather than through national rankings” [1].
3. If you serve several countries, measure each one. In this data, the Gemini model’s sources changed by network country even for questions that did not depend on place [1].

Disclosure: AVJ pays OpenAI standard prices for a ChatGPT Pro subscription and for API access, receives no credits or special terms, and uses OpenAI models among its editorial tools (see our AI use policy). AVJ has no other commercial relationship with OpenAI.

Update log: Oct. 30, 2026 — First published.

Sources

[1] Chen, M., & Koudas, N. (2026, Oct. 16). “Location, Location, Generation: Where You Ask Matters for Generative Engine Optimization.” arXiv:2610.24081v1 [cs.IR]. Preprint, not peer-reviewed. Accessed Oct. 30, 2026. https://arxiv.org/abs/2610.24081

[2] (HTML version of [1], version 1, used for section and table locations.) https://arxiv.org/html/2610.24081v1

How this was made

Prepared with AI assistance from the sources cited above. Editorial review by Nhi Phan (editor) and fact-checking by Jackies continue after publication; any change will be recorded in the update log and in our Corrections Policy and Log.