ChatGPT and Gemini cite different sources by location, study finds
A new audit of 31,200 ChatGPT and Gemini answers shows citations shift by location through two different mechanisms, making single-location brand tracking unreliable.
Key takeaways
- ChatGPT localizes citations using the app's declared location, down to city level, and ignores network IP entirely.
- Gemini does the opposite: it ignores declared location and localizes citations based on the network's country of origin.
- Both engines alter cited sources even for questions with no geographic dependency.
- Single-location GEO audits miss most of the variation because the two leading engines respond to different location signals.
- Multilaterals, banks, and standards bodies risk inconsistent citation visibility across regions for the same evergreen content.
A researcher in Austin asking ChatGPT about the best project-management software gets a different set of cited sources than the same researcher asking from a declared location of Mumbai, even though the question has nothing to do with geography. That is the finding buried in a new arXiv paper, "Location, Location, Generation," and it should unsettle any enterprise marketing team that thinks a single GEO audit from head office tells them how they appear worldwide.
The researchers ran 300 questions through ChatGPT across 19 declared locations and through ChatGPT and Gemini from five network locations, each twice, producing 31,200 answers. The repetition matters: it lets them separate genuine location effects from the ordinary noise of generative output. What they found is not noise. It is structure, and the structure differs by engine in ways that have nothing to do with the content of the question.
Two engines, two different notions of "where you are"
ChatGPT, the paper reports, takes its location cue from whatever the application declares, down to city level, and ignores the network entirely. A user on a Tokyo VPN who has set their app location to Chicago gets Chicago-flavoured citations. The assistant writes the declared place into its own search queries and its backend localizes results even for questions that never mention a place. Ask about enterprise cybersecurity vendors, and ChatGPT will quietly route the question through a local lens anyway.
Gemini does the opposite. It disregards the declared location and instead follows the country the network traffic originates from, per the paper, changing most of its cited sources even on queries that have no plausible local dimension. Two engines, two completely different plumbing systems, both invisible to the end user and both invisible to most brand-tracking dashboards.
This is the detail that should make a CMO at a bank or an industrial group sit up. GEO monitoring tools, as the paper notes, typically run from a single place: one office, one IP address, one assumption about what "the AI's answer" looks like. If that office is in London and the client base is split across the EU, Gulf, and Southeast Asia, the brand is auditing a citation pattern that may apply to almost nobody else asking the same question.
Why multilaterals and financial institutions cannot treat this as a UX quirk
Consider a development bank publishing guidance on microfinance regulation, or a standards body like ISO publishing a technical framework. These organizations operate on the premise that their authority is portable: a document vetted in Geneva should carry the same evidentiary weight when cited in Lagos, Jakarta, or São Paulo. The arXiv findings suggest that assumption does not survive contact with how generative engines actually work. If Gemini is swapping cited sources by network geography even on questions with no local dependency, then a UN agency's guidance document might be a top citation for users querying from Western Europe and absent entirely for users querying from a VPN exit node in Southeast Asia, for reasons that have nothing to do with the quality or relevance of the document itself.