How to get cited on Chinese AI.

Why third-party citation is the central work

Published GEO research finds brands cited by third parties are referenced roughly 6.5× more often than brands cited only on their own domain. In our 540-call panel across DeepSeek, Qwen and Doubao, Mainland-CN third-party platforms accounted for the dominant fraction of source mentions: 72.3% on DeepSeek, 85.0% on Qwen, 88.6% on Doubao. The on-site work — robots.txt, llms.txt, sitemap, content design — is necessary baseline, but it does not, on its own, get a brand into Chinese AI answers. The off-site substrate does.

Tier-order matters. Investing in community presence (Tier 3) before the brand has minimum-viable canonical authority (Tier 1) is wasted: community mentions without a credible site behind them get treated as low-quality signal. The right sequence is Foundation → Endorsement → Community → Amplification, with Tiers 1 and 2 running before Tier 3 work begins to compound.

Tier 1 — Foundation: your own canonical surface

Before any external work compounds, the brand's own domain must function as a clean canonical anchor for everything else. The infrastructure layer is shared with general GEO and is the one-hour setup that should not be skipped:

• Granular five-bucket robots.txt — allow retrieval and user-triggered crawlers (OAI-SearchBot, PerplexityBot, Claude-SearchBot, ChatGPT-User, Claude-User), choose your training-bot policy (GPTBot, ClaudeBot, CCBot, Bytespider). See AI crawler readiness.
• /llms.txt at site root with one-line description and link list of canonical entry points; /llms-full.txt for sites with depth.
• Sitemap submission to Google Search Console and Bing Webmaster Tools. The Bing AI Performance panel is the only free tool that counts AI citations directly.
• Markdown alternates on top-10 priority pages — <link rel="alternate" type="text/markdown" href="/page.md">.
• IndexNow — API key file at site root, ping on every URL change.
• 1,000–3,000-word reference pages for each top product, category and methodology question. Encyclopedia-style explainer pages outperform news pages by ~3× per citation in published samples.

The Tier-1 deliverable is not "good marketing site." It is "site that an AI engine can fetch, parse and quote without ambiguity, with canonical answers to the questions a buyer might ask." The methodology page exists to give the engine somewhere to anchor when summarising the brand; the per-product reference pages exist to give the engine quotable chunks. See methodology for what this looks like.

Tier 2 — Endorsement: Mainland Chinese tier-1 outlets

Once Tier 1 is at minimum-viable, third-party endorsement work begins. This is the layer where outlet trust transfers to brand: a named-byline article in a Tier-1 Mainland-CN business or vertical publication carries source weight that the brand's own site cannot generate alone. Priority outlets, in rough order:

The work at this tier is conventional PR — securing earned coverage, contributed bylines, expert commentary, response to category trend pieces. The difference from Western PR is that the AI-citation downstream effect is materially higher than the human-readership effect. A 36氪 piece may be read by a few thousand people but cited by Chinese AI engines for months across thousands of consumer prompts.

Tier 3 — Community: where the 6.5× third-party leverage lives

The dominant source class in our DeepSeek, Qwen and Doubao panels — and the layer where the published 6.5× third-party leverage signal is realised. Mainland community platforms in priority order:

知乎 (Zhihu)

Long-form Q&A platform — China's closest analogue to Reddit + Quora + Stack Exchange. Surfaces consistently across DeepSeek, Qwen and Doubao for considered-purchase categories. In our handbag panel, Zhihu surfaced in 97% of responses. The work: build named-expert presence (a single individual representing the brand, not anonymous accounts), answer category questions substantively (not promotionally), accumulate upvotes and follower count over months. See Zhihu AI visibility insight.

小红书 (Xiaohongshu / RED)

Lifestyle, B2C, beauty, FMCG, travel. Strong on Doubao and DeepSeek for consumer-product categories; weaker for B2B. The work: KOL collaboration, branded-account posting cadence, hashtag participation in category trends. See Xiaohongshu AI visibility insight.

SMZDM (什么值得买)

Deal-and-comparison aggregator. Dominant for commerce and price-comparison prompts on Doubao especially — but the weight collapses at ultra-luxury price tiers (our handbag panel showed DeepSeek SMZDM weight dropped from 100% at aspirational tiers to 33% at ultra-luxury). For everyday-luxury and below, this is a high-leverage surface. For Hermès/Chanel-tier, the substitute stack is The Purse Forum, Vogue Business CN, auction-house archives. See SMZDM AI visibility insight.

V2EX, 百度贴吧 (Baidu Tieba)

Secondary community surfaces. V2EX is technical-developer-leaning; Tieba is broader consumer with fragmented quality. Useful supplements but not where to start.

Tier 3.5 — Encyclopedia: the highest single-page leverage

In our 5-niche probe (125 calls, single LLM), off-site encyclopedic presence (Wikipedia EN/ZH or Wikidata) was the strongest predictor of brand mention rate among the signals we tested for DeepSeek. On-site schema density was uncorrelated and mildly inverted in the sample — controls (lower-mention brands) averaged more schema types than winners. Descriptive correlation, not causal lift; n=125 calls; not generalised to Qwen, Doubao, ChatGPT, Perplexity. But consistent enough to flag encyclopedic anchoring as the highest-priority intervention test for any DeepSeek-focused engagement.

百度百科 (Baidu Baike)

Mainland-CN encyclopedia. Survives across LLM training cycles; cited by every Chinese AI engine. Earning a Baike entry requires Mainland-CN editor approval and citations to verifiable Mainland sources. Western brands without Tier-2 Mainland press coverage typically cannot succeed here yet — Tier 2 is a prerequisite.

Wikipedia (EN and ZH) + Wikidata

The highest-trust signal short of academic citation. Wikipedia surfaced in 21% of DeepSeek responses across our luxury-handbag prompt panel — the highest-mentioned Western source in that sample. Earning a Wikipedia entry requires a credible third-party source trail (notability-by-citation), so this is a downstream effect of Tier-2 work rather than something to pursue in isolation. Wikidata is more accessible: structured-data entry that requires fewer notability hurdles. See Wikipedia AI visibility insight.

Tier 4 — Amplification: capturing attention windows

The atomized-distribution layer. These platforms capture attention long-form misses and feed transcripts and discussion back into the citation graph. Run after Tier 1–3 are stable.

• WeChat 公众号 — well-indexed by Tencent's Yuanbao; useful for B2B distribution and B2C content marketing.
• Bilibili — long-form video; transcripts surface in DeepSeek and Doubao corpus; growing fast for product-demonstration categories.
• Douyin (TikTok CN) — short-form video; high reach but lower per-citation weight; useful for awareness more than for AI surfacing.
• 微博 (Weibo) — social commentary; secondary AI-surfacing weight but useful for breaking news and category trend participation.

Technical and academic surfaces (where applicable)

For tech-adjacent brands, two parallel ladders compound on top of the consumer ladder:

• Technical: GitHub, CSDN, 掘金 (Juejin). High weight in DeepSeek for developer-tools and OSS-adjacent categories.
• Academic: CNKI, 万方, arXiv (for global research-heavy brands). First-class trust signal for AI engines — particularly Qwen, where institutional sources dominate.

The mesh that compounds

Cross-tier reinforcement is the differentiator. Every Tier-2 piece, every Tier-3 community post and every Tier-4 amplification asset should link or refer back to a Tier-1 anchor on the brand's own domain. Every Tier-1 anchor should reference one or two Tier-2 or Tier-3 surfaces as evidence. This creates the citation mesh that AI engines repeatedly traverse — and which, in published research, drives the difference between brands that surface in 5% of prompts and brands that surface in 80%+.