Which sources Chinese AI assistants cite

Chinese AI assistants don’t publish how they choose sources, so what’s known comes from vendors and consultancies sampling answers. All of the studies below are industry reports. Most are not public. The Xsignal report can be downloaded from the CEIBS repository, and one study has press coverage; you can’t check the others yourself. They are dated observations, not documented behaviour. Each platform has changed its sourcing within a year at least once.

Summary by platform

Assistant	Group	Observed preference	When	Study
Doubao	ByteDance	Toutiao, Douyin, Kuaidong Baike; Douyin video in local results	2026-06, 2026-08	Pharma; Pilates (Hefei)
Yuanbao	Tencent	WeChat official accounts, Tencent News and Tencent sites	2026-06, 2026-08	Pharma; Pilates
Qwen	Alibaba	Dianping for local questions; versions differ widely	2026-06, 2026-08	Pharma; Pilates
Wenxin	Baidu	Baijiahao, Baidu medical Q&A	2026-06	Pharma
DeepSeek	—	Narrowed to academic and government sites in late 2025, broader again in 2026; complaint and business-registry sites in local results	2025-H2 to 2026-08	Pharma; Pilates
Kimi	—	Zhihu, Bilibili, mainstream portals, more overseas sites (more so in Agent mode)	2026-06	Pharma

The pattern holds across studies and categories. The exact sites do not. They depend on category and city.

Each platform favours its own group

In a study of Pilates studio questions in Hefei (consumer apps, 3,469 citations), Doubao’s most cited source types were local classifieds (854), Douyin (244) and Toutiao (111). These are Doubao’s own counts; the 324 citations of local review platforms the report gives elsewhere are a total across all models. Yuanbao’s were Toutiao (110) and Tencent’s portal (106). Qwen’s top source was Dianping (207). The report concludes that each model clearly leans towards content inside its own ecosystem, so one set of articles for all models doesn’t work.

Observed in practice · observed 2026-08 · 合肥图灵象人工智能科技有限公司 (unpublished report, not publicly available; readers can't check it)

A pharma study by IQVIA and a partner (three cases, 300 prompts each) found the same split on medical topics. Doubao drew roughly equally on news media, medical media and academic or government sites, leaning to ByteDance’s Toutiao accounts, Douyin and Kuaidong Baike. Yuanbao leaned to WeChat official accounts and Tencent News, and Wenxin to Baijiahao and Baidu’s medical Q&A sites. Kimi favoured Zhihu, Bilibili and mainstream portals, with more overseas sites than the others, especially in Agent mode.

Observed in practice · observed 2026-06 · 艾昆纬（IQVIA）· 上海零假设信息科技 (unpublished report, not publicly available; readers can't check it)

A vendor study of infant formula questions points the same way, though it discloses no sample size.

Observed in 2025-06, more than 9 months ago. Answer engines change often; re-test before relying on this.

Xsignal’s monitoring of infant formula questions on DeepSeek, Doubao, Yuanbao and Kimi (1–20 June 2025) found different sources leading on each. Kimi relied on Zhihu (24.5%) and general portals. Doubao relied on Babytree (16.3%) and health shopping sites. DeepSeek “clearly leaned towards” mother-and-baby media (Babytree, Yaolan and others, over 21% combined). The report does not say how many prompts were asked, how often, or whether apps or APIs were used. (Quotes translated from Chinese.)

Observed in practice · observed 2025-06 · Xsignal（奇异因子）

Kimi’s lean towards Zhihu matches the pharma study above. DeepSeek’s use of vertical media in June 2025 came before the narrowing the pharma study reports. That study places the shift towards academic and government sources in the second half of 2025.

Few sources are shared across platforms

Observed in 2025-10, more than 9 months ago. Answer engines change often; re-test before relying on this.

A 2025 study by Growthbox and Percent Technology monitored seven platforms from July to October 2025. Of 51,000 sources cited, only 9.34% were cited by four or more platforms. In one category (electric toothbrushes, 1,240 answers, 22,015 sources) the share was 4.17%.

Observed in practice · observed 2025-10 · 增长黑盒 · 百分点科技 (unpublished report, not publicly available; readers can't check it)

A 2026 guide from GDMS and Miaozhen reports that Doubao’s citations were more concentrated and stable, while Qwen, DeepSeek and Yuanbao were more scattered and changeable. It adds that the same-named model on different platforms may share only 30–40% of its sources. Optimising for one form of a model doesn’t necessarily carry over.

Observed in practice · observed 2026-09 · GDMS全球数字营销峰会 · 秒针营销科学院 (unpublished report, not publicly available; readers can't check it)

What kind of content gets cited

DataStory asked the same 18–22 questions per category on six Chinese AI platforms, three times each (web and app versions, 12 monitoring endpoints in all) (648–792 answers per category, nine consumer categories, 13–15 August 2026). Across categories, cited content was mostly ranking and recommendation lists (22.8%–72.0%), then product descriptions (3.6%–42.2%). User reviews were “under 3%” in every category. The report gives these shares for all platforms combined, not per platform. (Quotes translated from Chinese.)

Observed in practice · observed 2026-08 · 数说故事 (unpublished report, not publicly available; readers can't check it)

This matches DATA17’s finding that category list pages were the most cited sources (see Claim: seeding brand mentions across the web boosts AI visibility).

Sourcing changes over time

DeepSeek’s sourcing changed several times. In the second half of 2025 it sharply narrowed to academic and government sites, favouring authority over freshness. From the second quarter of 2026 it widened again to large portals such as Sohu and NetEase, with PubMed and government sites for authoritative content. Qwen’s versions also differed widely in what they cited.

Observed in practice · observed 2026-06 · 艾昆纬（IQVIA）· 上海零假设信息科技 (unpublished report, not publicly available; readers can't check it)

This is the main reason to date every observation. A source map from early 2025 described a different DeepSeek.

Unmanaged pages get cited

In the Hefei study, DeepSeek cited complaint and consumer-rights platforms 75 times, 23.9% of all its citations, and also often cited business registry lookups. An unresolved complaint can end up in an AI’s description of a business.

Observed in practice · observed 2026-08 · 腾讯新闻（报道合肥图灵象普拉提行业GEO白皮书）

AI search inside apps

WeChat’s AI search mostly drew on 5–10 official-account posts plus 1–3 pages from Tencent’s ecosystem. Its sources were fairly consistent for the same question at similar times. Xiaohongshu’s AI search summarised users’ notes, with no outside sources observed. Douyin’s AI search gave no clear sources for general topics. Its answers depended strongly on the user’s IP and account.

Observed in practice · observed 2026-06 · 艾昆纬（IQVIA）· 上海零假设信息科技 (unpublished report, not publicly available; readers can't check it)

What to do with this

• Test your own category and city. The studies agree on the pattern (own-group content first, little overlap). They don’t agree on specific sites, which vary by category and city. Map sources by sampling answers in the apps your customers use (Measuring AI visibility by sampling prompts).
• Don’t spread one article set everywhere. Content that one assistant cites may be invisible to another.
• Look after third-party pages about you. Complaint sites, registries and review platforms are read as sources.
• Re-check every six to nine months. Each finding here carries its observation date. The wiki flags claims older than nine months.
• Paying to plant content in these sources so models recommend you is what regulators call AI data poisoning (GEO regulation in China).