For most of the search era, marketers operated on a comfortable assumption: search engines were largely interchangeable. Optimise for Google and Bing followed. Optimise for Baidu and 360 followed. The mechanism was the same. Match the query, surface the page. That assumption broke when AI replaced search. Generative engines don't surface a page; they write an answer. And what they reach for to write that answer is different on every engine.
The practical consequence: a brand can dominate one Chinese AI's recommendation set and be effectively absent from another's, because they're simply pulling from different shelves of the library. The three engines pull from substantially different source pools. Top-15 source overlap between any two of them is just 20–30%. The differences are large, stable, and consistent enough to plan around.
What "Chinese AI source overlap" means
Chinese AI source overlap is the share of websites that any two engines (DeepSeek, Qwen, Doubao) both name when asked which sources they relied on for a brand recommendation. We measure it as a top-15 Jaccard score per category. In our 2026 panel, the overlap sits at 20–30% — meaning 70–80% of each engine's source pool is unique to that engine.
The Universal Five Mainland Chinese sources every China-facing brand needs
Before the divergence, the agreement. Despite drawing from very different source pools, the three engines do agree on five Mainland-Chinese platforms, and they showed up in every single test cell we ran. They are the floor. No China-facing GEO strategy clears the bar without presence on all five.
Zhihu (知乎) · SMZDM (什么值得买) · Xiaohongshu (小红书) · Bilibili · Dianping (大众点评, services only)
The most concentrated of the five is SMZDM (什么值得买, "What's Worth Buying"): cited in the top 3 of every engine, with mention rates of 91% / 56% / 72% across DeepSeek, Qwen and Doubao. SMZDM is a Mainland-Chinese deal-and-review aggregator that publishes long-form, structured, ranked product comparisons. That format — long-form, structured, expert-style consumer reviews — is exactly what these engines reach for when answering Mainland-Chinese consumers, regardless of which company built the engine. If you have resources to optimise for one platform in Mainland China, this is it.
The Universal Five define the Mainland-China minimum-viable AI footprint. Get these in place first, then layer on engine-specific work.
DeepSeek, Qwen, Doubao personalities — how each Chinese AI picks sources differently
The remaining 70–80% of each engine's source pool is where the divergence lives. If the AIs were people, they wouldn't be three colleagues reading from the same handbook. They'd be three different kinds of advisor, each with their own taste in evidence, each reaching for different shelves of the library.
DeepSeek (深度求索): the academic friend
The friend who reads a lot, treats Reddit as legitimate research, and prefers a long-form Zhihu answer over a quick recommendation. Defaults to community Q&A and explainer content.
- Zhihu (知乎): cited in 100% of responses (every single one)
- SMZDM (什么值得买): 91%
- Xiaohongshu (小红书): 89%
- Bilibili long-form videos: 74%
- Reddit: 63% (the highest-cited Western source)
- Wikipedia at 21% and YouTube at 20% (next two Western sources, both meaningfully present, especially for FMCG and global products)
Reasons like a thoughtful prosumer · Reaches for community wisdom · Mildly Western-flavoured by Chinese-AI standards
Qwen (通义千问): the credentialed bureaucrat
Qwen still leans on the same mainstream consumer platforms as the other two: Zhihu 61%, SMZDM 56%, Bilibili 29%, Xiaohongshu 27%. What makes it distinctive is what it uniquely adds on top: government registries and professional-body sources the other two engines never meaningfully cite.
- For toothpaste: cites the Chinese Stomatological Association (中华口腔医学会) and the NMPA, China's drug regulator (国家药品监督管理局)
- For dental clinics: the National Physician Registry and DXY (丁香医生, China's leading professional medical platform)
- For services: the National Enterprise Credit Registry (国家企业信用信息公示系统) for vendor verification
- For appliances: the China Household Appliance Research Institute (中国家用电器研究院)
These institutional sources never appear meaningfully on DeepSeek. They are a Qwen signature, additive to the consumer-platform core rather than replacing it.
Layers clinical evidence and regulatory backing on top of mainstream consumer platforms · The most "credentialed" voice of the three
Doubao (豆包): the savvy shopper
The friend who sends you Tmall links with "look at the comments." Defaults to commerce + lifestyle KOL.
- SMZDM (什么值得买): 72% (its #1 source, even ahead of Zhihu)
- Xiaohongshu (小红书): 64%
- Zhihu (知乎): 59%
- Bilibili: 47%
- Dianping (大众点评): appears for services
- Niche vertical forums: iCard Forum (我爱卡论坛, credit cards), Qichacha / Tianyancha (企查查 / 天眼查, B2B verification), Bilibili Digital channels (B站数码) for consumer tech
Thinks like a shopper, not a scholar · Reaches for hyper-niche communities · The most Mainland-shopper-friendly voice of the three
How Chinese AI brand recommendations shift across tech, FMCG, and services
The same three AIs behave differently when the question category changes. The personalities aren't fixed; they amplify or quiet down depending on what's being asked.
| Category | How the AIs behave |
|---|---|
| Tech products (smartwatches, vacuums, headphones) | All three lean on SMZDM + Bilibili tech reviews + Zhihu. The academic, the shopper, and the bureaucrat largely agree here. |
| FMCG (toothpaste, shampoo) | Qwen and Doubao bring in regulatory and professional sources like the Chinese Stomatological Association, NMPA, and the DXY medical platform. DeepSeek goes to Wikipedia for ingredient explanations. The bureaucrat dominates this vertical. |
| Services (dental, banking, fitness, renovation) | All three suddenly cite Dianping (大众点评) for the first time. Qwen specifically leans on government registries for vendor verification. DeepSeek surfaces niche-community forums such as FlyertEa (飞客茶馆) for credit cards. |
Which Chinese AI should I optimise for first? It depends on category and audience. A toothpaste brand needs a different source-presence strategy than a robot-vacuum brand, even when targeting the same AI.
The Doubao Mainland-content bias and the Qwen institutional-source bias — replicated across two runs
Beyond the Universal Five and the three personalities, each engine carries a directional tilt that matters for where you spend GEO budget. Three to plan around:
- DeepSeek tilts hardest into community Q&A depth. Zhihu is cited in 100% of responses (every single one). Reddit at 63% is the highest-cited Western source. The brands DeepSeek surfaces tend to have long-form prosumer review presence.
- Qwen tilts toward institutional and regulatory backing. Industry bodies, regulators and national registries appear meaningfully on Qwen and essentially never on DeepSeek. The brands Qwen surfaces tend to carry credentialed third-party validation.
- Doubao tilts toward Mainland-Chinese consumer content. This is the most-replicated of the three tilts. It shows up on both metrics we measured (brand recall and source attribution), with two independent runs each.
The Doubao Mainland-content gap, sized two different ways:
Brand recall: share of top-8 brands that are Mainland-Chinese
| Panel | DeepSeek | Qwen | Doubao | Doubao gap |
|---|---|---|---|---|
| Panel 1 (5 niches) | 35% | 35% | 45% | +10pp vs both |
| Panel 2 (5 different niches) | 35% | 30% | 45% | +10pp vs DS, +15pp vs Qwen |
Source attribution: share of cited sources that are China-domiciled
| Panel | DeepSeek | Qwen | Doubao | Doubao gap |
|---|---|---|---|---|
| Panel 1 (10 categories) | 72% | 85% | 89% | +4pp vs Qwen, +17pp vs DS |
| Panel 2 (same prompts, retested next day) | 70% | 83% | 87% | +4pp vs Qwen, +17pp vs DS |
Two metrics, two independent runs each. Doubao led on every single one, with the same gap size in each replication.
If your Mainland-Chinese audience uses Doubao primarily, your visibility starts measurably higher (about +10pp on brand recall) than the DeepSeek / Qwen baseline — but only if your off-site footprint matches Doubao's preferred substrate (SMZDM, Xiaohongshu, Tmall depth, niche forums).
What this means for your China AI visibility strategy
Each row below is a highest-priority intervention hypothesis to test, not a guaranteed-causal recommendation. Pair any of them with a before/after measurement at T+12 weeks.
| If your audience uses… | Top GEO targets, in priority order |
|---|---|
| DeepSeek (developers, Cursor users, Tencent Yuanbao) | 1. Zhihu (知乎), the highest-priority target (cited in 100% of answers) 2. SMZDM (什么值得买) 3. Xiaohongshu (小红书) 4. Bilibili 5. Reddit (highest Western source at 63%; Wikipedia 21% and YouTube 20% also meaningfully present) |
| Qwen / Tongyi (Alibaba ecosystem) | 1. Industry & regulatory bodies (Chinese Stomatological Association, NMPA for FMCG; National Physician Registry for services) 2. Government registries (National Enterprise Credit Registry) 3. Zhihu + SMZDM |
| Doubao (ByteDance ecosystem, the largest Mainland consumer-AI platform) | 1. SMZDM (top source at 72%) 2. Xiaohongshu 3. Zhihu 4. Bilibili 5. Niche forums per vertical (iCard Forum for finance, Qichacha for B2B) |
| Universal Mainland-China baseline | Zhihu + SMZDM + Xiaohongshu + Bilibili + Dianping (services). These five appear in every test we ran. |
Common GEO advice vs what our data showed
| Common advice | What we found |
|---|---|
| "Add comprehensive schema markup to your website" | Schema density did not predict AI mention rate. The most heavily marked-up control we audited was among the lowest-mentioned. Schema is not sufficient on its own. |
| "Convert pages to FAQ format" | The academic study our research extends reports a −5.7% absorption effect for Q&A-only content. Treat as a flag against pure FAQ formats, not a ban on Q&A elements inside richer pages. |
| "Make sure your homepage is fast and fetchable" | One of our top-mentioned brands had a thin on-site GEO signal during our audit yet still reached a 92% mention rate. Fetchability is not the whole brand-recall story, but it still matters for chat-with-Search browsing surfaces. Keep it. |
| "Optimise your own website first" | For Mainland-Chinese consumer queries, the off-site footprint (community / KOL / e-commerce / institutional) is what the AIs actually pull from. Owned-site work is the foundation that feeds those third-party sources: necessary but not sufficient. |
The single most important shift: treat owned-site optimisation as the foundation, not the whole strategy. The leverage is in the third-party sources the AI cites, and that's where Mainland-China GEO budgets are currently underweighted.
Where Eastbound comes in
Eastbound runs the same source-attribution and brand-recall measurement on your specific category and audience-AI mix. Audit-grade output, Mainland-China consumer-voice prompts, all three engines covered.
If your team needs to know which third-party sources are doing the work for your category — and which AI your audience actually uses — run the free China AI visibility audit on your domain or book an intro call.
Methodology
- Sample. 1,830 API calls across 20 Mainland-Chinese consumer categories, 27 April – 2 May 2026. Two brand-recall panels (5 niches each) and two source-attribution panels (10 categories each, with the second a byte-identical retest of the first).
- Engines. DeepSeek (
deepseek-chat); Qwen-Plus on DashScope international (dashscope-intl.aliyuncs.com); Doubao (seed-2-0-lite-260228, BytePlus ModelArk international, Lite tier). All API-mode, default decoding (temperature 0.7). Chat-with-Search-ON browsing surface is a separate study. - What we measured. Source-attribution mention rate per (prompt × rep) cell; brand recall in top-8; share of Mainland-Chinese sources / brands. This is descriptive measurement of LLM behavior; it is not a causal claim about training data, retrieval, or human conversion.
- What we did not measure. Sales, conversion, attributable revenue. ChatGPT / Claude / Gemini / Perplexity / ERNIE / Yuanbao — not in this panel. Hong Kong, Taiwan, English-speaking, B2B audiences — out of scope; all prompts in Mainland register only.
- Reliability. Source-attribution panels passed standard test-retest thresholds: ICC(2,1) ≥ 0.97 across all three engines; Cohen's κ on top-5 source membership = 1.00; all p < 1e-4. Brand-recall pattern replicated across two non-overlapping niche panels. Caveat: Doubao long-tail source-ranking is less stable than top-5 (top-15 κ = 0.46) — treat long-tail Doubao findings with that caveat.
- Limitations. We cannot guarantee stability across silent vendor model updates; we recommend re-testing every 6 months.
Per-record JSONs and replication scripts available on request.