In early September, 36Kr reported that Chinese open-source models are becoming the "yardstick" for global AI, and that Alibaba's open-source "Octopus" model is opening new markets. Around the same time, Gartner projected that over 60% of searches would be handled by AI by the end of 2026 (cited by PE Daily, 2026-09-03). The conclusion for brands: AI search gateways are shifting from a Doubao-centric landscape to a crowded field of open-source models, and GEO strategy must move to a multi-engine mode.
Why the open-source wave changes AI search
Open source is not new, but the balance of power has changed. 36Kr data shows Chinese open-source AI held an average market share of about 13% from November 2024 to November 2025, up from 1.2% at the end of 2024. On Hugging Face's open-source leaderboard, eight of the top ten models are Chinese, with Qwen 3.5 at number one (source: Cover News, 2026-02-24). DeepSeek, Kimi, GLM and Qwen have all gone open source, making Chinese models the default foundation for developers worldwide.
For brands doing GEO, this means brand content is now cited by a growing number of AI products built on open-source foundations. CNNIC's 57th report puts China's generative AI users at 602 million, a 42.8% penetration rate. QuestMobile's 2026 AI application market report shows strong growth in monthly active users, with agents invoking skills becoming a new traffic gateway.
More engines, more volatile rankings
More engines mean more ranking volatility. Phoenix New Media's tests show significant exposure gaps for the same brand across engines: lululemon was mentioned in 71% of Qwen answers, 66% of Doubao answers and 70% of DeepSeek answers. Optimizing for a single engine abandons more than 30% of AI search traffic.
36Kr also flags a deeper risk: open weights make it easy to adopt a model and easy to swap it out, and when models are retrained and renamed, a Chinese brand can effectively disappear. The lesson: instead of betting on a temporary gateway, build source assets any engine can cite reliably. Xinbang Zhihui data (cited by PE Daily, 2026-09-03) shows third-party objective evaluations account for 80.9% of sources in AI answers.
Three changes for GEO in the open-source era
| Ecosystem change | Impact on brands | Recommended action |
|---|---|---|
| Surge in open-source models, fragmented gateways | Single-engine optimization fails, cross-engine gaps widen | Build multi-engine monitoring across Doubao, Qwen, DeepSeek, Kimi |
| Third-party evaluations dominate AI answers (80.9%) | Reputation and review content shape AI citations | Develop authoritative media, industry rankings, objective data |
| Agent-driven invocation rising | Content must be machine-readable | Structured data and fact bases, consistent messaging |
From single-engine to multi-engine coverage. Each engine's crawl logic, weighting and citation rules differ (see AI engine source preference comparison). Cover at least Doubao, Qwen, DeepSeek and Yuanbao, and use one unified brand fact base so every engine reads consistent information.
Third-party sources rise in importance. At 80.9%, what others say about you matters more than what you say about yourself in AI answers. Build third-party endorsements: authoritative media coverage, industry rankings, verifiable client cases and credentials, kept consistent with your website (see brand AI mention rate guide).
Content must feed machines. METR research shows the task duration at which frontier AI agents achieve 50% success doubles roughly every 89 days since 2024. Agents increasingly complete whole tasks for users. How structured your site is (Schema, FAQ, tables, parameter pages) determines whether agents can read and act on your information (see structured data and LLM inclusion).
A multi-engine GEO action checklist
- Run a multi-engine audit: ask your core brand terms in Doubao, Qwen, DeepSeek and Yuanbao, record mention rates and sources, and fix the largest gaps first.
- Build third-party sources: secure authoritative media coverage, industry rankings and verifiable client cases, creating a "website plus third-party" dual-source structure.
- Upgrade structured content: convert core pages to Schema plus FAQ plus tables so agents can read them programmatically.
- Set a monthly review cycle: following Gartner's pace, re-test each engine monthly and keep records (metrics are covered in how to measure GEO results).
- Watch new engine moves: open-source models iterate weekly; re-check brand content within 48 hours of engine updates (see GEO strategy after Qwen3.8-Max).
FAQ
What impact do Chinese open-source models have on GEO? They expand AI search gateways from a few players to dozens, so brands must stay visible across more engines. Objective third-party content now holds a larger share of AI answers, making reputation building more important.
Is optimizing only for Doubao enough? No. Exposure for the same brand can differ by more than five percentage points across Qwen, Doubao and DeepSeek.
Why do AI answers increasingly cite third-party evaluations? Third-party objective evaluations account for 80.9% of sources in AI answers, because objective, verifiable data is seen as more trustworthy than brand self-descriptions.
Is multi-engine GEO expensive? The core cost is content building and monitoring, not advertising. Follow a three-step path: restructure the website, add third-party sources, then run monthly reviews, starting with Doubao, Qwen and DeepSeek.
Sources: 36Kr (September 2026); Cover News Hugging Face leaderboard report (2026-02-24); 36Kr market share data (13%); CNNIC 57th report (602M generative AI users); QuestMobile 2026 AI application market report; PE Daily (2026-09-03, Gartner forecast and Xinbang Zhihui 80.9% data); Phoenix New Media cross-engine exposure tests.