The short version
- In one study of 1,909 English-language questions, three Chinese AI models named the brand asked about in 88.9% of answers and three Western models in 58.3% (Huang and colleagues, whose firm shares a name with the brand they featured).
- The same study found Chinese brands named in 31.2% of Western answers and 96.8% of Chinese ones, while Western brands moved only 9.2 points.
- Across 66 European brands, asking in a brand’s home language instead of English raised local champions’ recommendation share by 0.80 on a 0-to-1 scale, against 0.15 for global brands (Żatuchin, who runs an AI visibility company).
- University of Toronto researchers who translated 100 shopping questions into five languages found ChatGPT switched to different websites in each language, while Claude kept citing English-language ones.
- In a fake-brand test of 12 Chinese and Western models, every one could be fooled, so neither group is safer from manipulation.
Do Chinese AI models mention brands more often than Western ones?
In the only direct test we found, yes: Chinese models named brands far more often. Huang and colleagues (opens in a new tab) asked three Western models (GPT-4o, Claude and Gemini) and three Chinese ones (Qwen3, DeepSeek and Doubao) about 30 collaboration and productivity software brands. They kept only questions written entirely in English.
Across those 1,909 question-and-model pairs, the Chinese models mentioned the brand in 88.9% of answers and the Western models in 58.3%, a gap of 30.6 percentage points. The gap depended on where the brand came from.
| Brand origin | Western models | Chinese models |
|---|---|---|
| Western brands | 72.1% | 81.3% |
| Chinese brands | 31.2% | 96.8% |
| Global or mixed brands | 71.6% | 88.6% |
The gap also depended on the kind of question. For “best tools for X” questions it was 46.9 points; for “What is X?” questions it was 12.5. When the Chinese models did name a brand, they described it more warmly: +0.71 against +0.42 on a scale running from −1 (negative) to +1 (positive).
How much weight should you put on that finding?
Some, but not much on its own, because the study has a clear conflict of interest and a narrow scope. The three authors work at OmniEdge (Zhizibianjie) AI Consulting in Shenzhen. Their case-study brand, Zhizibianjie, was named in 65.6% of 32 questions to Chinese models and in none of 32 to Western ones. The paper declares no conflict, but the shared name is plain to see.
The scope is narrow too. It covers one software category, six models and one collection window, and the paper’s own dates do not agree with each other. To keep every question in English, the authors removed 891 question-and-model pairs (31.8%), mostly about Chinese brands whose names are written in Chinese characters. The Chinese-brand figures therefore rest on a filtered subset.
The authors also say their design cannot show cause and effect. Their 18-month plan for building visibility in other markets is advice, not a tested result. We would not base a budget on this paper alone.
Is it the AI model or the language of the question?
Both matter, and language alone changes which websites and brands an assistant reaches for. Chen and colleagues at the University of Toronto (opens in a new tab) translated 100 English shopping questions into Chinese, Japanese, German, French and Spanish. They put them to ChatGPT, Perplexity, Gemini and Claude.
Under non-English questions, the assistants cited more websites written in that language. ChatGPT switched to an almost entirely different set of sites in each language. Claude kept citing much the same English-language authorities whatever the language. Chinese questions were strongly localized, and changing the language moved results more than rewording a question did.
Żatuchin (opens in a new tab) ran 35,640 answers from GPT-5.4, Gemini 3.1 Pro and Perplexity about 66 brands in 12 European languages. For local champions, asking in the home language instead of English raised the share of answers naming them by 0.80 on a 0-to-1 scale; for global brands, by 0.15. The author is affiliated with Rankfor.AI, an AI visibility company, so treat the paper as vendor research.
Even within English-speaking markets, the country changes the answer. In our country study, two ChatGPT answers from the same country shared 0.594 of their brands; two from different countries shared 0.429.
Which brands are most affected?
Brands that are strong in one home market but thin in English-language sources are affected most. In the Huang study, the gap between Chinese and Western models was 9.2 points for Western brands and 65.6 points for Chinese brands. In the Żatuchin study, local champions were rarely named in English buyer questions and named in almost every home-language one.
The Toronto team saw the same pattern by category. Brand lists stayed more alike across languages where a few global brands dominate, such as cameras and laptops, and differed more in localized categories such as home appliances.
Local publications matter at the margin. Żatuchin found Wikipedia was the most-cited website in 11 of the 12 languages. In Lithuanian, the national business daily vz.lt edged it out, with 4.38% of citations.
Are Chinese AI assistants easier to manipulate?
Not on the available evidence: in one test, Chinese and Western models were similarly easy to fool. Luo and Chen (opens in a new tab) describe a March 2026 broadcast on China Central Television. It exposed paid operators seeding fake reviews online to push a fake brand into the top picks of mainstream Chinese AI assistants within hours.
Their own test rewrote real products into fake ones in the web pages an assistant reads. Across 12 commercial and open models, including GPT-5.4, Claude, Gemini, Qwen and DeepSeek, a single polluted page fooled models up to 27% of the time. Replacing the top three pages raised this to 73.8%.
Their main test used Chinese pages. When they repeated it with English pages, 8 of 12 models landed within 10 points of their Chinese results. Models resisted best in categories whose real brands they knew well, which again points to how much an assistant already knows about you.
What should you do about it?
Treat each AI ecosystem and each language as a separate market, and measure each one directly. Our guide to planning GEO market by market sets out the wider approach for global brands.
- Check your visibility in the languages your buyers use, not only in English. An English-only visibility check can understate a locally strong brand.
- If you sell in China, test Chinese assistants such as Qwen, DeepSeek and Doubao separately. Do not assume results from ChatGPT or Gemini carry over.
- Earn coverage in respected local-language publications in each market. Translating your own site is a start; the Toronto researchers found that answers in other languages lean on local sources.
- Watch “best X for Y” questions most closely, since that is where the Chinese-versus-Western gap was widest.
- Monitor your category for fake or misleading pages, which can sway assistants in both ecosystems.
If you want help setting up multilingual tracking, see our generative engine optimization service.
What does the research not tell us yet?
No one has independently repeated a Chinese-versus-Western brand comparison across several product categories.
- The only direct comparison covers one software category and comes from authors with a stake in the result.
- It used English questions only. How Chinese assistants treat Western brands in Chinese-language questions is untested in the papers we reviewed.
- The language studies compared languages and Western-built assistants, not Chinese and Western models side by side.
- Whether bilingual content causes a brand to appear more often has not been tested; the evidence is correlation.
- AI models change often, so every figure here describes one moment in time.
Frequently asked questions
Do DeepSeek and Qwen recommend different brands than ChatGPT?
In one study, yes. Chinese models named the brand asked about in 88.9% of English answers against 58.3% for Western models, but the authors have a conflict of interest and studied one software category.
Is an English-language AI visibility check enough for a global brand?
No, not if you sell in several languages. Across 66 European brands, switching to the home language raised local champions’ share of recommendations by 0.80 on a 0-to-1 scale, against 0.15 for global brands.
Will translating my website get my brand into Chinese AI answers?
Nobody has tested that. The Toronto study suggests that answers in other languages draw on local-language sources, so earned coverage in local media probably matters more than translation alone.
Are Chinese AI models more positive about brands?
One study says yes. When they named a brand, Chinese models scored +0.71 on a −1 to +1 tone scale against +0.42 for Western models, from the same conflicted study.
Sources
- Huang Junyao, Situ Ruimin and Ye Renqin (2025), Cultural Encoding in Large Language Models: The Existence Gap in AI-Mediated Brand Discovery (opens in a new tab), arXiv:2601.00869.
- Mahe Chen, Xiaoxuan Wang, Kaiwen Chen and Nick Koudas (2025), Generative Engine Optimization: How to Dominate AI Search (opens in a new tab), arXiv:2509.08919.
- Dmitrij Żatuchin (2026), The Language Blind Spot: How Query Language and Brand Recognition Tier Shape AI-Constructed Brand Reputation Across Twelve European Languages (opens in a new tab), arXiv:2606.23165.
- Minghao Luo and Liang Chen (2026), One Polluted Page Is Enough: Evaluating Web Content Pollution in LLM Recommenders (opens in a new tab), arXiv:2606.13610.
- Underneath (2026), Same question, four countries: do AI recommendations change?