A commercial recommendation looks like a ranking problem, but this paper shows a more basic gate is often operating first. In a generative search interface, the language of the question can decide whether local suppliers are even eligible to appear, while the exit IP mainly shifts which market’s brands get named once that gate is open.
That distinction matters because “the top result changed” is a weak signal. If the underlying problem is market selection rather than ranking noise, then the right control is not just to ask whether the answer moved — it is to ask whether local suppliers were available at all.
The authors probe that directly across logged-out ChatGPT web sessions and the OpenAI API, using the same commercial prompts under controlled combinations of surface, exit country, and query language. They are not trying to prove that one brand should win. They are trying to show when a market is admitted into the answer space in the first place.
234 usable runs.
What the authors actually tested
The design is built to separate three things that usually get tangled together: surface, exit IP, and query language. The study uses 11 experimental cells and 234 usable runs collected on 29 and 30 August 2026, with six identical runs per prompt per cell. The web interface was used logged out, so account history could not shape the output.
That matters because the dependent variable is not free-form prose. Brand presence is coded against a fixed market-to-brand mapping assembled before scoring, and every brand is assigned to exactly one of three classes: the country of the exit IP, another named country, or global. In other words, the paper is not asking the model to explain itself; it is auditing which market content shows up.
The API arm adds another useful control. It runs 36 prompts with the web_search tool enabled and 36 without, so the authors can see whether the behavior depends on surface-level search access or persists regardless. That makes the setup less about one interface quirk and more about the system’s actual routing logic.
The recommendation is noisy, but the noise is not the main story
The top recommendation is unstable almost everywhere. Across six identical runs, it changed on 4 of 6 prompts in the Berlin interface arm, 4 of 6 in the Oslo arm, 4 of 6 in the API with search enabled, and 4 of 6 with search disabled.
That is a lot of variance, but it is not random variance in the sense practitioners usually mean. The same instability rate shows up in browser and API settings, and it survives whether web search is on or off. So if you only track the exact brand at rank one, you miss the more important pattern: the surface is noisy, but the market gate underneath is systematic.
The paper’s move is to stop treating ranking instability as the phenomenon. Instead, it asks whether the model can surface a local supplier at all. That is where the language and IP effects become separable.
Query language is the availability gate; exit IP is the market switch
The strongest result is simple: query language, not location, decides whether local suppliers appear at all. Across the four cells where the query language is the official language of the exit country, a global supplier appeared in 1 of 24 runs. In the two English cells on the Istanbul and Tallinn connections, a local supplier appeared in 0 of 12 runs.
That is a very different failure mode from ordinary ranking drift. If the user asks in English, the system can behave as if local suppliers do not exist, even when the connection points to a market where they plainly should be candidates. The answer may still be fluent and commercially plausible, but the supplier set is already constrained before the model gets to the “reasoning” people tend to focus on.
The other half of the separation shows up when the language is held fixed and only the exit IP changes. With Turkish held constant, the answer stays in Turkish in both cells, but the supplier set follows the connection. That is why the authors summarize the behavior as “language sets the register, location sets the market.”
The Estonian connection shows three tiers, not a clean binary
The Estonia cells are the most interesting because they do not collapse into a simple local-versus-global split. Instead, the three query languages on that connection sort into three tiers: Estonian, Russian, and English behave differently rather than all moving together.
The numbers make that visible. On the Estonia connection, Estonian queries name a local supplier in every run, Russian queries name an Estonian supplier in 4 of 6 runs and a global supplier in all 6, and English names no local supplier in 0 of 6. That is not just “local language helps.” It is a graded language effect, with a minority language occupying a middle tier.
That middle tier matters operationally because it is where many localization assumptions break. If you treat “non-English” as one bucket, you will miss that some languages may partially unlock local market content while others suppress it completely. The paper also notes a negative control in a second category with no language effect at all, which helps show that the pattern is not universal across every commercial domain.
What this means for commercial recommendation systems
If your product promise is “show users relevant local suppliers,” this paper says you should not treat geography as the main control knob. Language is at least as important, and in these experiments it is the stronger gate: it determines whether local suppliers can appear at all.
That changes how you should test these systems. A single-language QA pass from one location is not enough, because the failure mode is not just which brand ranks first. You need multilingual probes from multiple exit IPs, and you need to audit supplier presence, not just final ranking. Otherwise you can miss a system that looks stable while quietly excluding the local market.
It also changes how to think about localization logic in generative search. If the product team wants market-aware recommendations, language-aware gating should be treated as a first-order control, alongside any IP-based logic. The paper cannot tell you whether the gate is implemented by retrieval, instruction tuning, safety-adjacent list construction, or some mix of them. But it does show that the gate exists, and that is enough to justify testing for it explicitly.
Across 234 controlled runs with repeated prompts, the study finds systematic, separable effects: query language strongly affects whether local suppliers can appear, while exit IP mainly shifts which market’s brands are named.
For commercial recommendations, treat query language as a market-selection input (availability gate) and treat exit IP as a market switch (brand-set shift) — and test both, because ranking instability alone won’t reveal the failure mode.
Dmitrij Żatuchin (2026). The Language of the Question Selects the Market: Query Language and Exit IP as Separable Factors in Commercial Recommendations from a Generative Search Interface. arXiv:2608.30052