# GEO in China: Doubao, Yuanbao, Qwen, ERNIE, DeepSeek and Kimi

> Chinese assistants each lean on their own group's content, and few sources are shared between them. Test in the real apps, publish in each ecosystem your customers use, keep the search engines able to crawl you, and look after third-party pages. Everything here is observed, not documented, so re-test often.

- URL: https://geo.casa/best-practices/china-platforms/
- Guide · Observed in practice
- Reviewed: 2026-10-28

This guide is for marketing and content teams whose customers ask Chinese AI assistants: Doubao, Yuanbao, Qwen, ERNIE (Wenxin), DeepSeek and Kimi. It gives the order of work: measure, map each platform's sources, publish where each one reads, and keep the search engines able to crawl your site. Expect a few days to set up measurement, then steady monthly work. The Chinese platforms publish almost nothing about how they pick sources. Most of what follows is dated industry observation, so treat it as a starting map to re-test.

## The short version

1. Measure in the consumer apps your customers use, not only through APIs.
2. Map which ecosystem each assistant draws on for your category and city.
3. Publish in each of those ecosystems instead of copying one article everywhere.
4. Keep Baidu, 360, Sogou and Toutiao able to crawl and verify your site.
5. Write comparison and list-style content that answers buying questions.
6. Look after third-party pages about you, including complaint and registry sites.
7. Stay inside the rules on AI data poisoning and fake content.

## 1. Measure in the real apps

Start by sampling answers, because you can't see inside these systems. The same model can give opposite results in its API and its app.

> **Observed in practice:**
> A 2026 study of Pilates studios in Hefei ran the same questions through vendor APIs and through consumer apps. DeepSeek recommended the studio in 69% of API runs but 1% of app runs. Yuanbao was the reverse: 3% by API and 99% in the app.
>
> Sources: [腾讯新闻（报道合肥图灵象普拉提行业GEO白皮书）: 行业首份GEO白皮书：AI正在重写'附近哪家好'的答案](https://news.qq.com/rain/a/20260909A0EGBH00)

> **Observed in practice:**
> In September 2025 Miaozhen asked DeepSeek, Doubao and Yuanbao 2,025 times in all: one typical question per category, five categories, 135 runs per model. It concluded that results reach basic stability only at 30 or more runs per prompt, and low-visibility brands need 60 or more.
>
> Sources: [中国商务广告协会AI营销应用工作委员会 · 秒针营销科学院: 2025年中国GEO行业发展报告](https://g4-img-all.oss-cn-beijing.aliyuncs.com/video/50108560/01503a0d1eaeb4429ee3c0014738e9ea.pdf)

- Build prompts from real customer questions, mostly category questions without your brand name.
- Ask on phones, logged out or on a fresh account, from your customers' city.
- Report each platform separately. Method: [Measuring AI visibility by sampling prompts](https://geo.casa/measurement/ai-visibility-measurement/).

## 2. Map each platform's ecosystem

Each assistant leans on content from its own corporate group. Knowing which group owns which assistant tells you where to look first.

> **Observed in practice:**
> A pharma study by IQVIA and a partner (three cases, 300 prompts each) found Doubao 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.
>
> Sources: 艾昆纬（IQVIA）· 上海零假设信息科技: 穿透医药GEO：医药品牌的内容可见性重构 (unpublished report, not publicly available; readers can't check it)

> **Observed in practice:**
> In the Hefei Pilates study (consumer apps, 3,469 citations), Doubao cited local classifieds, Douyin and Toutiao most. Yuanbao cited Toutiao and Tencent's portal most, and Qwen cited Dianping most. The report concludes that one set of articles for all models doesn't work.
>
> Sources: 合肥图灵象人工智能科技有限公司: 2026普拉提行业首份GEO白皮书 (unpublished report, not publicly available; readers can't check it)

> **Observed in practice:**
> DeepSeek narrowed sharply to academic and government sites in the second half of 2025. From the second quarter of 2026 it widened again to large portals such as Sohu and NetEase.
>
> Sources: 艾昆纬（IQVIA）· 上海零假设信息科技: 穿透医药GEO：医药品牌的内容可见性重构 (unpublished report, not publicly available; readers can't check it)

| Assistant | Group | Where to look first |
|---|---|---|
| Doubao | ByteDance | Toutiao, Douyin, Kuaidong Baike |
| Yuanbao | Tencent | WeChat official accounts, Tencent News |
| ERNIE (Wenxin) | Baidu | Baijiahao, Baidu Q&A sites |
| Qwen | Alibaba | Dianping for local questions |
| DeepSeek | — | Portals, authoritative sites; changes often |
| Kimi | — | Zhihu, Bilibili, portals |

The exact sites depend on category and city. Full detail and dates: [Which sources Chinese AI assistants cite](https://geo.casa/ai-search/china-ai-platform-sources/).

## 3. Publish in each ecosystem

One article syndicated everywhere tends to reach only some assistants. The studies found little overlap between platforms.

> **Observed in practice:**
> 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.
>
> Sources: 增长黑盒 · 百分点科技: 2025中国GEO趋势与品牌增长策略报告 (unpublished report, not publicly available; readers can't check it)

> **Observed in practice:**
> 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.
>
> Sources: GDMS全球数字营销峰会 · 秒针营销科学院: 2026品牌GEO行动指南 (unpublished report, not publicly available; readers can't check it)

Practical split:

- Pick the two or three assistants that matter most to your customers.
- Run an account in each matching ecosystem (a Toutiao account, a WeChat official account, Baijiahao, a Zhihu presence).
- Adapt the same facts to each channel's format. Keep prices, specs and names identical everywhere.
- Keep your own site as the canonical source of those facts.

> In China you don't optimise for "AI". You optimise for ByteDance, Tencent, Baidu and Alibaba, one ecosystem at a time.

## 4. Keep the search engines able to crawl you

The assistants don't document their crawlers, but the search engines in the same groups do. Your own site still needs to be crawlable by them. No public document from these engines yet says how their AI products choose pages.

> **Officially documented:**
> Baiduspider follows robots.txt. Blocking `Baiduspider` removes a site from Baidu search and from other search results Baidu provides. The FAQ names no separate token for Baidu's AI products.
>
> Sources: [Baidu: FAQs of Baiduspider](https://help.baidu.com/question?prod_id=99&class=0&id=3001)

> **Officially documented:**
> Baidu's sitemap protocol allows 50,000 URLs per file but requires each file to be under 10MB, also before gzip compression.
>
> Sources: [Baidu: 百度Sitemap协议](https://help.baidu.com/question?prod_id=99&class=0&id=3032)

> **Officially documented:**
> 360 Search supports `user-agent`, `allow` and `disallow` with the tokens `360Spider`, `360Spider-Image` and `360Spider-Video`.
>
> Sources: [360 Search: Robots协议支持](https://www.so.com/help/help_3_2.html)

> **Officially documented:**
> Sogou's crawler supports robots.txt, but a changed file can take several weeks to take effect.
>
> Sources: [Sogou: 站长指南](https://www.sogou.com/docs/help/webmasters.htm)

> **Officially documented:**
> Toutiao Search names its crawler `Bytespider` and published IP ranges in 2019. The page does not mention robots.txt.
>
> Sources: [Toutiao Search: 头条搜索spider介绍](https://www.toutiao.com/i6759451726496940557/)

> **Officially documented:**
> 360 Search says its crawler does not yet support reverse DNS checks. Match the IP against the ranges it publishes instead.
>
> Sources: [360 Search: 360蜘蛛IP](https://www.so.com/help/spider_ip.html)

Checklist for your site:

| Engine | Do this |
|---|---|
| Baidu | Allow Baiduspider; submit sitemaps under 10MB; verify by reverse DNS |
| 360 | Allow 360Spider; verify by published IP list |
| Sogou | Allow Sogou's crawler; allow weeks for robots changes |
| Toutiao | Don't block Bytespider by accident at the firewall; treat its 2019 IP list as incomplete |

Details: [robots.txt](https://geo.casa/crawling-indexing/robots-txt/), [XML sitemaps](https://geo.casa/crawling-indexing/xml-sitemaps/), [Verify AI crawler traffic in server logs](https://geo.casa/implementation/verify-ai-crawlers/). Main text must be in the server HTML ([JavaScript rendering and AI crawlers](https://geo.casa/implementation/javascript-rendering-ai-crawlers/)).

## 5. Write content that answers buying questions

Cited content in China leans heavily towards lists and comparisons.

> **Observed in practice:**
> DataStory asked 18–22 questions per category on six Chinese AI platforms (web and app, 12 endpoints), three times each, across nine consumer 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.
>
> Sources: 数说故事: 2026消费品GEO白皮书 (unpublished report, not publicly available; readers can't check it)

- Publish honest comparison pages and buying guides that name the criteria.
- State specs, prices, service areas and certifications as plain facts.
- Writing style: [For content editors: writing articles AI answers can quote](https://geo.casa/best-practices/content-editor/).

## 6. Look after third-party pages

Assistants read pages you don't control. A forgotten complaint can become part of an answer.

> **Observed in practice:**
> In the Hefei study, DeepSeek cited complaint and consumer-rights platforms 75 times, 23.9% of all its citations. It also often cited business registry lookups.
>
> Sources: [腾讯新闻（报道合肥图灵象普拉提行业GEO白皮书）: 行业首份GEO白皮书：AI正在重写'附近哪家好'的答案](https://news.qq.com/rain/a/20260909A0EGBH00)

- Search complaint sites, Dianping, registries and Baike for your name each quarter.
- Resolve complaints in public, and correct registry and encyclopedia data at the source.
- More on earned coverage: [PR and brand teams: how press, encyclopedias, Q&A sites and reviews shape AI answers](https://geo.casa/best-practices/pr-earned-media/). On wrong answers: [When AI gets your brand wrong: find it, fix it, prevent it](https://geo.casa/best-practices/wrong-ai-answers/).

> The page that hurts you most is usually one you never wrote and forgot to check.

## 7. Stay inside the rules

> **Officially documented:**
> The CAC's 2026 campaign on AI misuse lists AI data poisoning as a target, including forging authoritative data and using GEO for malicious marketing. Its second stage targets fake-account networks that mass-produce near-identical posts.
>
> Sources: [新华网（中央网信办通知）: 中央网信办部署开展“清朗·整治AI应用乱象”专项行动](https://www.news.cn/politics/20260430/5a335918ff184d098c169c3dc0401508/c.html)

> **Observed in practice:**
> CCTV's 3·15 show reported a test in which a GEO tool published more than ten promotional articles about a fitness band that did not exist. Two major AI models then recommended it near the top.
>
> Sources: [第一财经: 3·15揭露GEO“洗脑”：AI大模型被投毒问题日趋严重](https://www.yicai.com/news/103087422.html)

The second claim shows the tactic works for a while. The first shows why it is now a regulatory risk. Background: [GEO regulation in China](https://geo.casa/foundations/geo-regulation-china/).

## What not to do

- **Don't measure only through APIs.** API and app results can be opposite (step 1).
- **Don't treat old source maps as current.** DeepSeek changed its sourcing within a year (step 2).
- **Don't mass-publish near-identical "recommendation" articles.** That is what the Qinglang campaign targets (step 7).
- **Don't buy services that promise first place in an assistant.** Nothing public shows a reliable way to guarantee it ([Claim: seeding brand mentions across the web boosts AI visibility](https://geo.casa/ai-search/inauthentic-mentions/), [Before you hire a GEO agency](https://geo.casa/best-practices/choose-agency/)).
- **Don't block Bytespider or Baiduspider to stop AI use.** No documented token separates their AI use from search (step 4).

## How you'll know it's working

- Re-run the same prompt set monthly in each app, 30 or more runs per prompt ([Measuring AI visibility by sampling prompts](https://geo.casa/measurement/ai-visibility-measurement/)).
- Track per platform: mention rate, position of first mention, and which of your channels are cited.
- Check that the facts assistants state about you match your site.
- Re-map sources every six to nine months. Reporting to management: [For marketing leads: measuring GEO and reporting it upward](https://geo.casa/best-practices/measure-report/).

## Evidence at a glance

| Recommendation | Evidence |
|---|---|
| Measure in apps, 30+ runs | Observed |
| Each assistant favours its own group | Observed |
| Publish per ecosystem, not one copy everywhere | Observed |
| Crawler access for Baidu, 360, Sogou, Toutiao | Official |
| Lists and comparison content | Observed |
| Watch complaint and registry pages | Observed |
| Avoid mass-produced promotional content | Official (regulation) |

## Sources

- [腾讯新闻（报道合肥图灵象普拉提行业GEO白皮书）: 行业首份GEO白皮书：AI正在重写'附近哪家好'的答案](https://news.qq.com/rain/a/20260909A0EGBH00)
- [中国商务广告协会AI营销应用工作委员会 · 秒针营销科学院: 2025年中国GEO行业发展报告](https://g4-img-all.oss-cn-beijing.aliyuncs.com/video/50108560/01503a0d1eaeb4429ee3c0014738e9ea.pdf)
- 艾昆纬（IQVIA）· 上海零假设信息科技: 穿透医药GEO：医药品牌的内容可见性重构 (unpublished report, not publicly available; readers can't check it)
- 合肥图灵象人工智能科技有限公司: 2026普拉提行业首份GEO白皮书 (unpublished report, not publicly available; readers can't check it)
- 增长黑盒 · 百分点科技: 2025中国GEO趋势与品牌增长策略报告 (unpublished report, not publicly available; readers can't check it)
- GDMS全球数字营销峰会 · 秒针营销科学院: 2026品牌GEO行动指南 (unpublished report, not publicly available; readers can't check it)
- [Baidu: FAQs of Baiduspider](https://help.baidu.com/question?prod_id=99&class=0&id=3001)
- [Baidu: 百度Sitemap协议](https://help.baidu.com/question?prod_id=99&class=0&id=3032)
- [360 Search: Robots协议支持](https://www.so.com/help/help_3_2.html)
- [Sogou: 站长指南](https://www.sogou.com/docs/help/webmasters.htm)
- [Toutiao Search: 头条搜索spider介绍](https://www.toutiao.com/i6759451726496940557/)
- [360 Search: 360蜘蛛IP](https://www.so.com/help/spider_ip.html)
- 数说故事: 2026消费品GEO白皮书 (unpublished report, not publicly available; readers can't check it)
- [新华网（中央网信办通知）: 中央网信办部署开展“清朗·整治AI应用乱象”专项行动](https://www.news.cn/politics/20260430/5a335918ff184d098c169c3dc0401508/c.html)
- [第一财经: 3·15揭露GEO“洗脑”：AI大模型被投毒问题日趋严重](https://www.yicai.com/news/103087422.html)