The Three-Baike Problem: Why Your Brand Facts Diverge Across Chinese AI Models in 2026

Here is a structural fact about Chinese AI that has no Western equivalent: the three companies that own China's largest AI assistants each also own their own encyclopedia. Baidu owns Baidu Baike and ERNIE. Tencent owns Sogou Baike and Yuanbao. ByteDance owns Baike.com — the platform formerly known as Hudong Baike — and Doubao.

In the West, every major model reaches for the same reference layer. An analysis of 30 million citations found ChatGPT cited Wikipedia at 47.9% of all citations, the single highest-cited domain in its responses. One encyclopedia, one shared set of brand facts, one version of who you are.

China has three, and they are owned by competitors.

That single ownership fact explains more about cross-model brand score variance than most GEO audits ever surface. When a brand scores 71 in Doubao and 44 in ERNIE, the instinct is to blame prompt phrasing or content volume. Increasingly, the real cause is simpler and more fixable: the three encyclopedias hold three different versions of the brand's entity record, and each model is drinking from the one its parent company owns.

The vertical integration nobody priced in

The scale asymmetry between these three platforms matters as much as their existence.

Baidu Baike announced in January 2026 that it had passed 30 million entries, contributed by more than 8.03 million users, and shipped AI-native features including "Dynamic Baike" and an AI knowledge graph layer. In February 2026 Baidu launched BaiduWiki, an international edition supporting English, French, Spanish, Russian and Japanese. That is not a dormant reference site — it is an actively expanding, AI-restructured knowledge base with a first-party pipeline into ERNIE.

Sogou Baike came to Tencent through the Sogou acquisition and now functions as a core training and retrieval set for Yuanbao, Tencent's assistant, which sits inside the WeChat ecosystem where roughly a billion users already are.

Baike.com is the one most international brands have never touched. ByteDance acquired it specifically to support its search ambitions, and it feeds Doubao — which, per QuestMobile's June 2026 rankings, holds 382 million monthly active users, more than double Qwen's 167 million and nearly triple DeepSeek's 130 million.

Read that last column against the MAU table and the problem becomes uncomfortable. The encyclopedia with the thinnest international brand coverage feeds the assistant with the largest audience in China. Brands that have spent years maintaining a clean Baidu Baike entry have optimized for the reference layer behind ERNIE, which is not where the users went.

What divergence actually looks like

Entity divergence is rarely dramatic. It is almost never a wrong company name. It is small, boring inconsistency that compounds:

• Founding year and headquarters. A brand lists its global HQ on one platform and its China JV registration city on another. When a model is asked "is this a domestic or foreign brand," it answers from whichever record it holds.
• Parent company and portfolio. Post-acquisition brand structures update on Baidu Baike within months and sit stale on the other two for years. Models then attribute the brand to a former owner, or fail to associate it with a portfolio that would have earned it a mention in category queries.
• Category language. The single most consequential field. A brand described as a "personal care company" in one entry and a "beauty and skincare brand" in another will surface in completely different category prompts. Chinese AI category retrieval is unforgiving about this — the model does not infer that these are the same thing.
• Product line naming. Chinese product names, transliterations, and official Chinese brand names that differ across the three platforms fragment the entity into what the model treats as separate, weaker entities.

This is why cross-model score gaps often look irrational. The brand is not being penalized. It is being described differently by three separate sources, and each model is reporting its own source faithfully.

Consensus as an authority signal

The deeper mechanic is that agreement between independent sources is itself a ranking signal. When a model cross-checks a claim and finds the same founding year, the same headquarters, the same category description across three encyclopedias owned by three rival companies, that convergence functions as verification. When it finds three different answers, the model does what a cautious system should do: it hedges, softens, or omits the brand from a recommendation set where confidence is required.

This reframes what an encyclopedia entry is for. Most brand teams still treat Baidu Baike as a brand-owned property — a place to publish a flattering description. In a retrieval-driven environment, its function is closer to a notary. Its value is not persuasion; it is corroboration. A neutral, verifiable, mutually consistent entity record is worth more to your AI visibility than a well-written one on a single platform.

Zhihu occupies a related position for a different reason. It is one of the few high-authority Chinese sources that no single model owner controls — both Tencent and Baidu are investors — which makes it a genuinely cross-model surface. Where the three Baike platforms fragment along ownership lines, Zhihu is one of the few places where a well-supported answer can influence DeepSeek, Qwen and ERNIE simultaneously.

Why brands miss this

Three reasons, in roughly this order.

First, the Western mental model. Teams who have maintained a Wikipedia entry assume the job is done once the Chinese equivalent exists. The equivalent is not singular.

Second, agency scope. Most China digital agencies were retained for Baidu SEO, and Baidu Baike was inside that scope. Sogou Baike and Baike.com were never in anyone's contract, because until ownership consolidated behind AI assistants they carried little search value.

Third, editorial friction. Chinese encyclopedia platforms require authoritative third-party citations — state media, established outlets, academic sources, official corporate announcements — and reject forums, blogs and self-published material. A brand with thin Chinese-language press coverage often cannot get an entry approved at all, which means the encyclopedia problem is downstream of a media problem. That is also why practitioners describe entity establishment in Chinese AI as a six-to-twelve-month program rather than a campaign: you frequently have to earn citable coverage before you can fix the record that depends on it.

Takeaway for Brand Marketers

Four things, in priority order.

1. Audit all three, not one. Pull your brand's entry from Baidu Baike, Sogou Baike and Baike.com and put the core fields side by side: official Chinese name, founding year, headquarters, parent company, category description, flagship product names. Most brands running this for the first time find at least two material inconsistencies. This costs an afternoon and is the highest-leverage GEO diagnostic available to a brand that has never done it.

2. Weight the fix by audience, not by familiarity. If your gaps are on Baike.com, that is the ByteDance record feeding a 382-million-MAU assistant. It should outrank a cosmetic improvement to a Baidu Baike entry that is already broadly correct — even though Baidu Baike is the platform your agency knows how to edit.

3. Standardize category language first. Before touching prose, fix how your brand's category is described, and make the three descriptions identical. Category phrasing determines which prompts you are eligible to appear in. Everything else is refinement.

4. Sequence media before encyclopedia. If entries are rejected for lack of citable sources, the encyclopedia is not your first problem. Industry media placement — 36Kr, Huxiu, Leiphone-tier outlets — does double duty here: it is directly citable by models, and it unlocks the encyclopedia edit you actually wanted. A 200-word brand mention in credible industry media carries more AI citation weight than a 2,000-word article on your own domain, and it is also the evidence your Baike submission requires.

The uncomfortable summary is that in China, the reference layer beneath AI is not neutral infrastructure. It is competitive infrastructure, owned by the same companies competing for your customers' attention. Brands that keep treating it as one channel will keep getting three different answers about who they are.

Related: track how these divergences show up as measurable score gaps across all six Chinese models on the GEO Hub brand rankings.

Sources: QuestMobile AI-native app MAU rankings (June 2026); Xinhua / Beijing News coverage of Baidu Baike's 30-million-entry milestone (January 2026); Search Engine Land, "How China's fragmented search ecosystem is reshaping SEO in 2026."