Guide · Chinese AI Search

Chinese GEO: How Chinese AI Assistants Choose Their Sources

How Doubao, ERNIE, Qwen, DeepSeek, Yuanbao and Kimi decide what to cite — and what that means for brands that want to be found in Chinese AI answers.

Published: 2026-09-28
Updated: 2026-09-28

By Marcus Pentzek and Kun Tang, Jademond Digital. This guide draws on research and case studies the authors presented at the Shenzhen SEO Conference and Search SEOul in 2026. Figures marked as estimates are modelled from the sources named; all other figures are sourced.

The Short Answer

Chinese AI assistants find their answers on the web. What they cite depends on which sources are available, trusted and in Chinese.
Editorial and vertical media lead. Media and specialist publications make up 37% to 70% of cited source types, depending on assistant and industry.
Ecosystem bias is smaller than assumed. Across the industries studied, assistants drew on average only 1% to 37% of their citations from their own company's platforms.
Language and industry change the rules. Chinese sources dominate in some industries and not in others; B2B and consumer topics are cited from different kinds of sources.
Stay honest, stay relevant, don't spam. Lasting visibility comes from being present and accurate where your buyers and the assistants look — not from gaming the system.

Why Chinese GEO Matters Now

Chinese users are moving part of their information search to AI assistants. According to CNNIC, the number of generative AI users in China rose from 249 million in December 2024 to 602 million in December 2025 — an increase of 353 million in one year. Over the same period, the number of search engine users fell by about 96 million, to 782 million.

This is not a simple swap. Time spent online rose from 28.7 to 32.5 hours per week, and many other online activities grew or shrank at the same time. But the direction is clear: according to QuestMobile, answering questions is the most common reason people use AI in China, named by 76% of AI users.

For brands, both channels matter. Web search engines such as Baidu still reach several hundred million people and still send visitors to websites. AI assistants rarely send clicks — but they increasingly shape which brands buyers consider, compare and trust. Generative engine optimization (GEO) is the work of being found, described correctly and recommended in those answers.

Search Engine Users vs. Generative AI Users in China

Source: CNNIC Statistical Reports on China's Internet Development; QuestMobile.

Search engine users: 877.82 million (December 2024) → 782.06 million (December 2025): −95.76 million
Generative AI users: 249 million (December 2024) → 602 million (December 2025): +353 million
Average weekly hours online: 28.7 (December 2024) → 32.5 (December 2025)
Main use of AI: Answering questions: 76% of AI users (QuestMobile)

China's Scattered Search Landscape

In China, "search" does not mean one search engine. People search where the intent fits: on Baidu for practical and commercial questions, on Rednote for purchase research, on Douyin for discovery, on Bilibili for learning, on Zhihu for expert answers — and increasingly in AI assistants. A 2024 poll by Nanjing Marketing Group found that 44% of searches happen in classic web search engines, 42% on social, video and community platforms, and 12% in AI tools.

Dominant Search Intent by Platform

Based on the authors' analysis of platform data, QuestMobile and the Nanjing Marketing Group poll.

Baidu: Utility and real-life research — health, housing, cost of living, policy — plus commercial and B2B research
AI assistants: Answers and research; rarely send traffic, but increasingly shape decisions
Rednote: Information seeking, then purchase research; Rednote reports that more than 70% of its users search on the platform
Douyin: Entertainment-first discovery and edutainment
Bilibili: Learning and long-form content
Zhihu: Credentialed, in-depth expertise
WeChat: Private, trusted, closed-loop communication; search is used far less than the platform's size suggests

The Chinese AI Assistants and Their Ecosystems

Chinese AI Assistants by Monthly Active Users

Source: QuestMobile. Many people use more than one assistant, so the figures cannot be added up. The assistants of ByteDance, Baidu, Alibaba and Tencent together account for about 86% of the combined user base.

Doubao (ByteDance): About 382 million
Wenxin / ERNIE (Baidu): About 200 million
Qwen (Alibaba): About 167 million
DeepSeek: About 130 million
Yuanbao (Tencent): About 114 million
Kimi (Moonshot AI): About 9 million

Content Ecosystems Behind the Assistants

Each large technology company owns a universe of user-generated and editorial content. A common assumption is that each assistant mainly cites its owner's platforms — the research below tests that assumption.

Baidu (ERNIE / Wenxin): Baidu Zhidao, Baijiahao, Baidu Baike, Baidu Tieba, Baidu Wenku, and Zhihu (Baidu-invested)
ByteDance (Doubao): Douyin, Toutiao, Xigua Video, Dongchedi and other ByteDance platforms
Alibaba (Qwen): Taobao, Tmall, Xianyu, Youku, 1688, Amap, Fliggy, Quark and other Alibaba platforms
Tencent (Yuanbao): WeChat Official Accounts and Channels, QQ, Tencent Video, Sogou, and Zhihu (Tencent-invested)
DeepSeek and Kimi: No large consumer content ecosystem of their own; they rely mainly on the open web and other data sources

How an AI Answer Is Made: Model Knowledge vs. Grounding

Two Layers of Knowledge

Model knowledge: What the model learned during training. Influencing it is slow — it can take weeks, months or even years.
Grounding: The web sources an assistant retrieves while answering. These can change within days to weeks when the right content appears in the right sources.
What it means for planning: Short-term work targets the sources assistants retrieve today; long-term work builds a consistent, accurate brand presence that later model versions learn from. A sound GEO plan addresses both.

What Chinese AI Actually Cites: Four Assumptions Tested

Kun Tang's research "Crack the China GEO Code", presented at the Shenzhen SEO Conference 2026, tested four common assumptions about Chinese AI citations. The analysis covered six Chinese AI assistants — Wenxin (ERNIE), Doubao, Qwen, Yuanbao, DeepSeek and Kimi — across three industries: smartphones, electric vehicles and industrial machinery, with more than 40,000 citations in scope. Every cited URL was classified by owner, language and source type, and compared with authority and social signals.

Assumption 1: Chinese AI Favours Its Own Ecosystem

Finding: Only partly true. Averaged across the three industries, own-ecosystem sources accounted for between 1% and 37% of citations, depending on the assistant.
By industry: Smartphones: 2% to 56%. Electric vehicles: 0% to 41%. Industrial machinery: 0% to 26%.
Qwen (Alibaba): The biggest surprise: only 0.8% of citations came from Alibaba's own domains — less than from Baidu properties (8%).
Doubao (ByteDance): Relies heavily on ByteDance sources, yet also cites Tencent (3.9%), Alibaba (2.3%) and Baidu (1.6%) domains.
Wenxin (Baidu): Relies heavily on Baidu's own sources, yet also cites Tencent (5.5%) and Alibaba (1.6%) domains.
Yuanbao (Tencent): Relies heavily on Tencent sources, and also cites ByteDance (4.7%), Baidu (1%) and Alibaba (0.7%) domains.
DeepSeek: Cites mainly Tencent (4.7%) and Baidu (1.8%) properties among the large ecosystems, and ByteDance (0.6%).
Kimi: Relies least on the large ecosystems — only Alibaba domains appeared, at 1.5%.
Takeaway: Do not over-optimize for one company's platforms; being present on the major platforms and the open web is the safer strategy.

Assumption 2: Chinese AI Cites Mostly Chinese Sources

Finding: It depends on the industry.
Electric vehicles: More than 95% of cited sources were Chinese.
Smartphones: About 62% of cited sources were Chinese.
Industrial machinery: Split between Chinese and non-Chinese sources.
Takeaway: Localize into Chinese. Non-Chinese sources are largely invisible in some consumer industries, while B2B topics leave more room for international sources.

Assumption 3: SEO Authority Predicts AI Citations

Finding: Moderately, and not everywhere. Authority was measured with a third-party Baidu authority rating (Aizhan).
Smartphones: Modest positive correlation (r ≈ 0.35): authority matters, but it is not the only gate.
Electric vehicles: Stronger positive correlation (r ≈ 0.49).
Industrial machinery: No significant correlation: foreign manufacturers' websites were cited heavily despite low Chinese authority ratings.
Takeaway: Authority helps, especially in consumer industries — but in B2B, specific and technical content on your own website can be cited regardless of authority scores.

Assumption 4: Social Proof Drives AI Citations

Finding: Social signals and citations tend to move together, but public data is limited — the Bilibili sample, for example, contained only eight videos.
Takeaway: Social proof helps, but it is not the primary driver of citations.

Editorial and Vertical Media Win

Across assistants and industries, media and vertical or editorial publications accounted for 37% to 70% of all cited source types. That makes earned coverage in relevant Chinese media — digital PR — a cornerstone of GEO in China.

Which Sources Matter by Industry

Smartphones (B2C): Social media and user-generated content matter most for Wenxin and Doubao, at 46% to 57% of cited source types.
Electric vehicles (B2C): Social media and user-generated content remain important, but vertical and editorial media also carry great weight.
Industrial machinery (B2B): Specialist reference sources, industry directories and brand or corporate websites become more important sources.

"It's a citation game won by authority and language, not platform ownership." — Kun Tang, Shenzhen SEO Conference 2026

The Chinese GEO Playbook

Three Moves to Start With

1. Audit your platform mix: Check where your brand is present today — your website, media coverage, and platforms such as Zhihu, Baijiahao, Rednote or industry directories — and where the assistants actually look for your topic.
2. Map your industry's logic: Find out which source types the assistants cite in your industry: user content, vertical media, directories or corporate websites.
3. Build a GEO scorecard: Define a set of Chinese-language prompts that reflect real buyer questions, run them in the assistants that matter, and track the results over time.

The Ongoing Work

Status quo and source analysis: Measure how your brand and competitors appear, and identify which sources each assistant draws on for your topic.
Owned sources: A Chinese website and brand information that are accurate, specific and easy to quote — corporate sites are a secondary citation layer, and a primary one in B2B.
Digital PR: Earned coverage in neutral editorial and vertical media.
Platform presence: Genuine, useful content on the platforms your buyers use — without fake accounts or seeded answers.
Reference sources: Correct entries in reference sources such as Baidu Baike, following each platform's editorial rules.
Search: Baidu SEO and GEO support each other; web results feed what several assistants retrieve. See the step-by-step guide to Baidu SEO.

Measuring Chinese AI Visibility

What a GEO Scorecard Tracks

Prompt set: Chinese-language questions buyers actually ask, grouped by topic and stage — kept stable so results are comparable over time
Assistants: The assistants that matter for your audience, for example Doubao, ERNIE, Qwen, DeepSeek, Yuanbao and Kimi
Mention and recommendation: Whether your brand is mentioned or recommended, and in what position
Accuracy: Whether facts about your company and products are correct
Sentiment: How your brand is described
Cited sources: Which websites and platforms the answer relies on
Method: Dates, number of runs and settings — AI answers vary between runs, so a single test proves little

What Not to Do

Because grounding sources can be influenced quickly, some providers flood them with manipulative content — mass-produced AI articles on self-media platforms, seeded answers under fake identities, orchestrated reviews. This is sometimes called AI poisoning. Platforms remove such content, and it puts the brand's credibility at risk.

"Stay honest, stay relevant, do not spam!" — Marcus Pentzek, Search SEOul 2026. The red flags in How to Choose a Chinese GEO Agency describe these practices in detail.

Examples from Practice

Three Case Examples

Anonymized examples from Jademond Digital's client work, presented at Search SEOul and the Shenzhen SEO Conference 2026.

Education support services (international student sector): The client chose to focus on digital PR only, so ByteDance platforms were deliberately left out. Visibility was built through editorial sources.
Industrial measurement and quality-control instruments (global leader): SEO work was already in place before GEO-specific services started in 2026, giving the brand a foundation of authoritative Chinese content.
Industrial signage solutions (global manufacturer): Visibility was tracked for the client and its competitors over the same period. Only the client gained significant AI visibility — evidence that visibility does not simply grow for every brand over time.

GEO and Baidu SEO: Not Either-Or

Web search still reaches several hundred million users in China and still drives website traffic, because users need to visit a website to act on what they find. AI assistants seldom send traffic, but may be a key driver of conversion. Baidu and the AI assistants cover much of the same classic search intent — which is why the strongest approach plans both together, sharing research, content and measurement.

Frequently Asked Questions

What is Chinese GEO?: Chinese GEO (generative engine optimization) is the work of making a brand findable, correctly described and recommended in the answers of Chinese AI assistants such as Doubao, ERNIE, Qwen, DeepSeek, Yuanbao and Kimi — mainly by being present and accurate in the sources these assistants rely on.
How many people in China use generative AI?: According to CNNIC, China had 602 million generative AI users in December 2025, up from 249 million a year earlier. Over the same period, search engine users fell by about 96 million to 782 million.
Do Chinese AI assistants only cite their own company's platforms?: No. Research presented at the Shenzhen SEO Conference 2026 found that, averaged across three industries, own-ecosystem sources made up between 1% and 37% of citations. Alibaba's Qwen cited Alibaba domains in only 0.8% of cases.
Which sources do Chinese AI assistants cite most?: Media and vertical or editorial publications accounted for 37% to 70% of cited source types in the research. User-generated content matters more in consumer categories such as smartphones, while directories, reference sources and corporate websites matter more in industrial B2B topics.
Do Chinese AI assistants cite English-language sources?: It depends on the industry. In the research, more than 95% of cited sources for electric vehicles were Chinese and about 62% for smartphones, while industrial machinery was split between Chinese and non-Chinese sources.
How long does it take to become visible in Chinese AI answers?: Changes to the sources assistants retrieve when answering can show effects within days to weeks. Influencing what a model learned in training takes much longer — from weeks to years.
Does domain authority help with Chinese AI citations?: Moderately in consumer industries: the research found correlations of about 0.35 for smartphones and 0.49 for electric vehicles, but no significant correlation for industrial machinery.
Is GEO replacing Baidu SEO?: No. Web search still reaches several hundred million users in China and drives website traffic, while AI assistants rarely send clicks. The two work best together.