Type: Out-of-band primary read. Full text supplied by Adam after PDF extraction failed on the scheduled route.
1. Provenance
Betello, F., Missault, P., Czinczoll, T., Beliuha, M. — Investigating Brand Origin Geographic Bias in E-Commerce Rankings. FAccT 2026, June 25–28 2026, Montreal. Pages 8093–8105. Surfaced by the digest-01 sweep (axis 2, brand bias); listed there as tier C from an ACM DL link. Now tier A — full text and all appendix tables read, including Tables 2–8.
2. Scope, and the limit on transfer
Models: Llama (US), Mistral (EU), Qwen (Asia). Three open-weight, non-commercial models. Task: e-commerce product ranking. Language: English only. Origins are model-inferred, not verified — the authors are explicit that no authoritative ground truth for brand origin exists (they cite Ferrero's Italian founding versus Luxembourg registration).
This is the governing caveat: none of the engines LBOX actually measures were tested. No ChatGPT, Claude, Gemini, Perplexity or Copilot. Findings below are tier A for these three models and tier D — inference only — if applied to commercial engines. Do not carry these numbers into a client deliverable about ChatGPT behaviour. They justify a test, not a claim.
3. Findings
3.1 Models hold a belief about where your brand is from, and it is often wrong
Sanity-checked against Wikidata on the 200 most frequent brands: ~80% agreement across all three models. So roughly one in five well-established brands has a misattributed origin — worse for less familiar or ambiguous brands. This is a misattribution class LBOX does not currently test for.
3.2 Baseline origin representation differs wildly by model, and model nationality predicts nothing
Top-1 share by inferred brand origin, base condition:
| Model | North America | Europe | Asia |
|---|---|---|---|
| Mistral (EU-built) | 48.04% | 37.92% | 12.36% |
| Llama (US-built) | 34.17% | 17.72% | 32.03% |
| Qwen (Asia-built) | 71.53% | 9.96% | 17.16% |
The Chinese-built model is the most North-America-biased by a wide margin. Authors attribute this to English-language training bias. A model's country of origin does not predict its brand preference — this must be measured per engine, never assumed.
3.3 Stating brand origin changes rankings, and the direction flips by model
Change in top-1 share when brand country of origin is named in the prompt:
| Model | NA | Europe | Asia |
|---|---|---|---|
| Mistral | −5.63% | +8.73% | −3.40% |
| Llama | +1.41% | +5.05% | −3.88% |
| Qwen | +5.39% | −9.48% | −16.70% |
There is no universal advice here. Declaring European origin gains 8.7% on Mistral and loses 9.5% on Qwen. Declaring Asian origin loses 16.7% on Qwen. Origin disclosure is a per-engine strategic decision, and any brand publishing country-of-origin in schema markup or page copy is making that choice implicitly, for every engine at once.
3.4 User location moves rankings by up to 59%
Top-1 change when the prompt states both user location and brand location (Qwen, the most sensitive):
- German user → European brands +59.16%, Asian −21.80%
- Chinese user → Asian brands +38.90%, European −24.45%
- US user → NA brands +14.25%, Asian −36.46%, European −34.40%
Llama behaves as expected but more mildly (US user → NA +10.88%; German user → EU +21.37%). Mistral inverts: with a North American user, NA brands are penalised (−3.44%), and with a Chinese user, Asian brands are penalised (−1.84%). Same-region matching is not a safe assumption.
3.5 The bias is invisible in relevance metrics
Across every location condition, NDCG@1/5/10/15 barely moves and no difference is statistically significant at p < 0.05. Rankings stay relevant and well-formed. But set overlap collapses: Mistral's Jaccard@1 between base and brand-location conditions is 0.51 — roughly half of all top-1 results change identity while relevance scores stay flat.
You cannot detect this by judging whether the answer is good. It is only visible through counterfactual comparison: identical query, varied stated location.
4. Extracted rules
MUST
- R12. Test each engine's believed country of origin for every client brand, and correct it where wrong. ~20% error rate on major brands; a misattributed origin silently reprices the brand in every geographic condition. (A for Llama/Mistral/Qwen; test required for commercial engines.)
- R13. Geo-stratify every visibility measurement. A measurement taken from one location is a measurement of that location only; the same query can shift top-1 composition by up to 59%. (A, FAccT 2026)
- R14. Use counterfactual location probes as standard method — same query, varied user and brand location — because relevance metrics do not reveal the effect. (A)
- R15. Measure origin sensitivity per engine before recommending origin disclosure. The same disclosure gains on one model and loses double digits on another. (A)
MUST NOT
- R16. Do not infer an engine's brand-geography bias from where the engine was built. The Chinese-built model was the most North-America-favouring at 71.5% of top-1. (A)
- R17. Do not assume local-to-buyer framing helps. Mistral penalised same-region brands for both NA and CN users. (A)
- R18. Do not carry these effect sizes into client-facing claims about commercial engines. Open models were tested; ChatGPT, Claude, Gemini and Perplexity were not. (governance)
5. Revalidation of standing rules
| Rule | Status |
|---|---|
| Citation Readiness section 3 scope invariant (diagnostic unit = product variant × market × language × intent cluster × time window) | CONFIRMS, strongly. This paper is empirical justification for an invariant LBOX adopted on reasoning alone. Market and language are not administrative fields — they change top-1 composition by up to 59%. |
| R6 (state model set and date on every finding) | CONFIRMS. Three models, three different directions of effect. |
| R1 (compete on parity-breaking signals; brand identity only 1.2% of variance — digest 01) | QUALIFIES. Digest 01's variance decomposition treated brand as near-irrelevant. This paper shows brand geography carries real weight, and it is a brand attribute rather than a product parameter. The two are reconcilable — inferred origin may act through a different channel than brand name recognition — but the 1.2% figure should not be read as "brand attributes don't matter." |
| R10 (no single-run measurements) | CONFIRMS and extends — single-geography is as defective as single-run. |
6. Open queue additions
The reference list is a better source than keyword search. Priority reads, all cited by this paper:
- Itzhak et al. 2025, Planted in Pretraining, Swayed by Finetuning: Origins of Cognitive Biases in LLMs (arXiv 2507.07186) — speaks directly to the unresolved parametric-versus-tuning mechanism question. Highest value in the list.
- Lichtenberg, Buchholz, Schwöbel 2024, LLMs as Recommender Systems: A Study of Popularity Bias (arXiv 2406.01285) — corroboration or challenge for digest 01's incumbency finding.
- Cao 2025, WSDM, Writing Style Matters: Bias and Fairness in Information Retrieval Systems — directly actionable for content recommendations.
- Wang et al. 2024, Do Large Language Models Rank Fairly? (arXiv 2404.03192).
- Dai et al. 2024, KDD, Bias and Unfairness in Information Retrieval Systems: New Challenges in the LLM Era — survey; good map of the field.
- Ziems et al. 2024, ACL Findings, Measuring and Addressing Indexical Bias in Information Retrieval.
- Kamruzzaman et al. 2024, Global is Good, Local is Bad? (arXiv 2406.13997) — upgrade from digest 01 tier C.