Key findings
- Profound analyzed 3.25 billion citations across seven global engines and 14 countries, showing that query language can materially reshape citation patterns; its study did not include mainland China or Chinese answer engines.
- In Geolix.ai's low-code SaaS case, 1,233 valid answers to ten Chinese buyer-intent questions produced 15,353 citation records. Brand mention rates ranged from 10.3% on Perplexity to 81.9% on DeepSeek.
- Chinese-engine and Western-engine source pools overlapped only modestly in that case: 68 shared domains and a 12% Jaccard overlap.
- Chinese-language sources represented 90.4% to 100% of citations across all six engines in the low-code SaaS case, including ChatGPT, Perplexity and Google AI Mode.
- These two Geolix.ai cases are complementary evidence, not a matched Chinese-versus-English experiment: they cover different categories, prompts, dates, repetition levels and engine sets.
What global research establishes
Profound’s March 2026 study covered ChatGPT, Claude, Google AI Mode, Google AI Overviews, Gemini, Microsoft Copilot and Perplexity. Social sources accounted for 15.3% of citations in Google AI Overviews and 14.5% in AI Mode, compared with 9.1% in ChatGPT, 3.99% in Claude and 3.6% in Gemini. The useful conclusion is not that one source type always wins; citation behavior varies by engine and market language.
Sources: Profound citation study | Peec AI language study
Peec AI provides a second, independently collected vendor dataset. Across more than ten million prompts and twenty million fan-out searches, it reported that 78% of non-English sessions included English supplementation and that 43% of fan-out searches for non-English prompts were conducted on the English-language web. Peer-reviewed multilingual retrieval research also documents language preference and high-resource-language bias. Together, these studies explain why translation alone cannot guarantee equivalent retrieval.
Sources: ACL 2025 multilingual RAG research | ACL 2026 language-bias research
The low-code SaaS case: the same Chinese questions, six different visibility outcomes
From June 15 to 19, 2026, Geolix.ai tested one low-code SaaS brand using ten Chinese purchase-intent questions across ChatGPT, DeepSeek, Doubao, Google AI Mode, Perplexity and Qwen. Each question was scheduled five times per day per engine. After invalid runs were removed, the dataset contained 1,233 valid answers and 15,353 citation records.
| Engine | Valid answers | Mention rate | Top-1 rate | Chinese-source share |
|---|---|---|---|---|
| DeepSeek | 210 | 81.9% | 5.2% | 96.5% |
| Doubao | 199 | 71.4% | 4.5% | 100.0% |
| Qwen | 206 | 70.9% | 34.5% | 100.0% |
| ChatGPT | 207 | 45.4% | 3.4% | 96.3% |
| Google AI Mode | 208 | 30.8% | n/a | 90.4% |
| Perplexity | 203 | 10.3% | 1.0% | 97.1% |
The gap is operationally large: mention rate differed by roughly eight times, while measurable Top-1 rate ranged from 1.0% to 34.5%. Google AI Mode’s rank field failed to populate, so its Top-1 result is reported as unavailable rather than zero. At question level, the same content gap could reverse across engines. For example, two questions received zero ChatGPT mentions while DeepSeek recorded 76% and 71%; Qwen recorded 10% and 100%. A single blended GEO score would hide those differences.
Source ecosystems diverged even when the prompt language stayed fixed
The six engines did not simply produce different rankings from one common source pool. DeepSeek drew 62.5% of its citations from cloud and official developer ecosystems. Doubao drew 31.1% from user-generated and video sources, especially Douyin. Qwen drew 72.6% from user-generated and video sources and 25.1% from search self-references. By contrast, ChatGPT’s largest category was vendor websites at 68.6%. These are case-specific observations, not universal rules for every category.
When Geolix.ai grouped DeepSeek, Doubao and Qwen as Chinese engines and ChatGPT, Google AI Mode and Perplexity as Western engines, the two groups shared only 68 citation domains. Their unique-domain Jaccard overlap was 12%. Yet more than 90% of citations in every engine were Chinese-language sources. This is strong evidence that Chinese questions activate a predominantly Chinese evidence environment, while engine architecture still determines which part of that environment is used.
What the English fintech case adds
A separate Geolix.ai case monitored nine English purchase-intent questions for a GEO service brand serving fintech in Singapore and APAC. Between July 20 and 31, 2026, ChatGPT API, Gemini and Perplexity produced 15,495 valid answers and 223,196 citation records. Perplexity mentioned the brand in 75.3% of answers, Gemini in 38.6% and ChatGPT in 6.9%. Most citations were non-Chinese sources: 89.0%, 99.9% and 91.5%, respectively.
The case also shows that citation and brand visibility are different metrics. ChatGPT cited the target brand’s own domain in 22.9% of answers but named the brand in only 6.9%—a 3.3-to-1 gap. Perplexity, meanwhile, cited Chinese-language versions of the brand’s pages 3,149 times, representing 42.5% of its citations to paired pages; ChatGPT and Gemini did not cite those Chinese versions. Even under English prompts, engines can treat localized pages differently.
What APAC fintech teams should do
- Maintain separate English and Chinese buyer-question libraries based on real discovery, comparison, trust, licensing and market-access intents.
- Report mention rate, Top-1 and Top-3 recommendation rate, own-domain citation rate and third-party citation rate separately by engine.
- Map each engine’s cited domains and source types before deciding where to publish or conduct outreach.
- Keep regulated facts—legal entity, licence, market availability, fees, eligibility and product limits—consistent across language versions and third-party profiles.
- Treat before-and-after changes as observational unless the test holds prompts, engines, timing and other content actions constant.
- Add an independent factual-accuracy review; the current Geolix.ai export measures mentions, rankings and citations but does not contain a factual-accuracy field.
Methodology and limitations
Both cases are production-monitoring exports, not randomized experiments. The two cases differ in category, brand, question set, date range, repetition level and engine availability, so their aggregate percentages should not be used to calculate a causal language effect. In the low-code SaaS case, DeepSeek and Doubao region values were not stored. In the English fintech case, Google AI Mode succeeded on only 1 of 450 scheduled runs and Google AI Overviews on 62 of 450; both were excluded. The dataset contains no factual-accuracy field, and negative-sentiment counts were uniformly zero, so neither metric is used as evidence.
Frequently asked questions
Does this prove that Chinese content always performs better for Chinese prompts?
No. It shows that Chinese sources dominated one controlled Chinese-language case and that engines used materially different subsets of those sources. A strict language-effect estimate would require the same brand and same prompts tested in both languages during the same period.
Can ChatGPT performance predict DeepSeek or Qwen performance?
Not safely. In the low-code SaaS case, the same brand and prompts produced large differences in mentions, recommendations and source types across engines.
What is the minimum useful multilingual GEO dashboard?
Prompt-level answers and cited URLs, segmented by engine, language, location and date, with mention, recommendation, own-domain citation and factual-accuracy metrics kept separate.
References
- Profound — How query language reshapes AI citations: https://www.tryprofound.com/blog/how-query-language-reshapes-ai-citations
- Peec AI — ChatGPT searches in English, even when you don't: https://peec.ai/blog/chatgpt-searches-in-english-even-when-you-don-t
- ACL 2025 — Investigating Language Preference of Multilingual RAG Systems: https://aclanthology.org/2025.findings-acl.295/
- ACL 2026 — All Languages Matter: Understanding and Mitigating Language Bias in Multilingual RAG: https://aclanthology.org/2026.acl-long.338/
- ACL 2024 — Language Bias in Multilingual Information Retrieval: https://aclanthology.org/2024.mrl-1.23/
- Geolix.ai — Official website and engine coverage: https://geolix.ai/
- Geolix.ai — GEO bilingual case dataset: Geolix.ai internal production-monitoring export, anonymized, July 31, 2026.