A location field is not a universal control
An AI visibility report labeled “New York” does not guarantee that ChatGPT and Gemini were tested under equivalent geographic conditions. In an October 2026 API audit, ChatGPT’s local citations tracked the city explicitly passed to its web-search tool. Gemini’s citations instead tracked the country of the originating network connection. Passing the same declared location to both systems failed to put them on equal footing.
Researchers Mahe Chen and Nick Koudas evaluated 300 English-language commercial prompts split across three categories: explicit “near me” queries, queries where answers naturally depend on location, and non-local queries. None of the prompts named a specific place. The authors independently varied the declared API locations and the network exit points, measuring how much cited website domains overlapped across different locations compared to repeated runs under identical settings. That repeat baseline separated true geographic sensitivity from normal generation variance.
Location influences more than just search query wording
In the primary ChatGPT tests, altering the declared city shifted cited domains mainly for “near me” and location-dependent prompts; non-local prompts barely moved. When no location was declared, switching the network connection yielded no detectable difference beyond normal repeat variance.
One mechanism was obvious in the search queries ChatGPT generated: every search for a “near me” prompt explicitly appended the declared city. Yet query rewriting did not explain everything. Cited domains still diverged beyond baseline noise even when neither compared response included a location name in its search queries. The authors point to localization inside the underlying search backend for that residual shift. Auditing only the model’s generated search terms will therefore miss part of how location shapes citations.
Gemini responded to a different lever. Supplying declared coordinates did nothing in pilot testing, whereas routing requests through different network countries altered its cited domains—even on non-local prompts. Switching network exits between New York and Los Angeles produced no measurable city-level effect. However, that does not prove Gemini is blind to sub-national geography: the network tests relied in part on data-center VPN exits, and the authors note that residential connections might trigger finer-grained localization.
Adding markets and adding repeats solve different problems
The business recommendation findings illustrate how much a single-city audit leaves out. Across ChatGPT’s “near me” prompts, sampling one randomly chosen US city surfaced just 9% of the distinct businesses seen across all 15 tested US cities; sampling five cities captured 36%. Each city was queried twice per prompt. Those percentages reflect coverage within the study’s collected sample, not the total universe of businesses ChatGPT could ever surface.
Still, watching the list of recommended brands grow as you add cities does not prove geography caused the expansion. On non-local prompts—where ChatGPT showed minimal sensitivity to declared cities—testing one city still captured only 19% of the total observed recommendations, while five cities captured 53%. The authors trace that accumulation to ordinary run-to-run variance. Querying more cities inherently means running more total prompts.
For teams buying or building AI visibility tracking, the necessary methodology questions are specific: What location parameter is passed to the search tool? Where does the network traffic exit? How many repeat runs are executed per condition? A simple market tag on a dashboard answers none of these, requiring explicit documentation from the tool provider.
While the paper clarifies how to measure cross-market visibility, it is strictly an auditing study. The authors did not alter business websites or test whether adding geographic signals to page content increases citation rates.
The source and evidence
Location, Location, Generation: Where You Ask Matters for Generative Engine Optimization
Study limitations and disclosures
- Primary experiments ran on gpt-5.4-mini with web search and gemini-3.1-flash-lite with Google Search grounding via API, which may behave differently than consumer web interfaces. A follow-up check on a full-sized model tested ChatGPT only.
- Business recommendation counts depended on LLM-based extraction; a manual audit found 29 of 36 extractions were completely accurate, leaving some margin of error in the coverage metrics.
Evidence behind this briefing
[c1] The audit tested 300 researcher-written questions across ten commercial topics, equally divided among near-me, place-dependent and not-local questions. E1 varied ChatGPT’s declared location across 19 cities and an unspecified-location condition, using two repeated requests and optional versus prompted web search. E2 tested ChatGPT and Gemini without declared locations across five network locations, with a sixth session repeating the Canadian connection to assess temporal drift.
Section 3, Method — Questions and locations; E1: declared location; E2: network location · Read the study
[c2] The localization index subtracts cross-location cited-domain Jaccard overlap from same-setting repeat overlap, so it measures location-related differences beyond run-to-run noise rather than accuracy or a percentage of sources changed. Confidence intervals are bootstrapped over questions. Recommended-name extraction used a language model; only 29 of 36 manually checked extractions were fully correct.
Section 3, Method — Measures · Read the study
[c3] Declared location changed ChatGPT’s cited domains mainly for local questions: localization indices for two US cities versus two countries were 0.14 versus 0.28 for near-me questions, 0.17 versus 0.43 for place-dependent questions, and 0.00 versus 0.03 for not-local questions. Gemini’s cross-country network indices were 0.43 for near-me questions (95% CI 0.40–0.46), 0.40 for place-dependent questions and 0.23 for not-local questions (95% CI 0.20–0.27). These are noise-adjusted overlap differences, not accuracy scores.
Section 4, Results — RQ1: Location changes the sources of local answers; Figure 1 · Read the study
[c4] The tested engines responded to different location channels. ChatGPT’s network-location indices were within noise, while Gemini followed network country rather than declared coordinates. Gemini showed no detectable New York–Los Angeles difference (index 0.00; 95% CI −0.03–0.04). ChatGPT explicitly inserted the declared place into 100% of near-me searches and 48% of place-dependent searches. Monitoring therefore needs engine-specific location controls rather than assuming a location parameter or VPN affects both systems identically.
Section 4, Results — RQ2: The two engines use different channels · Read the study
[c5] ChatGPT’s citation localization persisted when neither compared answer’s queries named its location: index 0.17, with a 95% CI of 0.09–0.25. The authors attribute this to search-backend localization; when both queries named the place, the index was 0.37. Near-me localization was similar for US cities less than 1,000 km apart and more than 3,000 km apart, with Spearman correlation 0.01. The closest tested cities were 293 km apart, so this does not establish behavior between nearby neighborhoods.
Section 4, Results — RQ2: The two engines use different channels, paragraph following Table 1 · Read the study
[c6] Single-city monitoring substantially undercovered the businesses recommended by ChatGPT for near-me questions in the tested 15-city US sample. One city captured 9% of the businesses observed across all 15 cities, five captured 36%, and ten captured 69%; 96% of near-me recommended businesses appeared in only one city. Repetition also mattered: not-local questions had 19% coverage from one city and 53% from five despite little location effect. These are coverage results within the sampled answers, not estimates of all businesses the engine could ever recommend.
Section 4, Results — Implications for monitoring; Figure 3, based on two requests per question per city · Read the study
[c8] The findings are an October 2026 snapshot of APIs and mostly small models, not a validation of consumer applications. The full-size check covered ChatGPT only. Network testing covered five locations and used data-center VPN exits, so Gemini could behave more locally on residential connections. Questions were English-language and researcher-written.
Section 5, Discussion and Limitations — Limitations · Read the study