What ERNIE Actually Cites: 81.7% Baidu-Owned, 41.3% Baijiahao

For most of 2026, advice about Chinese AI search has been platform-agnostic: publish authoritative content, make it verifiable, get into the sources the models trust. That advice is correct but incomplete. A GEO vendor's teardown of ERNIE (文心一言), Baidu's consumer assistant, published in early September, put numbers on something practitioners have suspected: ERNIE cites Baidu properties far more heavily than any other Chinese assistant cites its own ecosystem.

The figures: 81.7% of ERNIE's citations come from Baidu-owned or Baidu-controlled content, and Baijiahao (百家号) alone accounts for 41.3%. A separate third-party evaluation found that Baidu-ecosystem content is cited roughly 40% more often in ERNIE than in competing assistants.

For overseas brands, this is the most concrete GEO finding of the quarter — because it turns "get into Chinese AI answers" into three named properties you can either access or not.

📊 The numbers, and where they come from

Two independent measurements point the same direction:

• 81.7% of ERNIE's cited sources are Baidu-ecosystem content, with Baijiahao at 41.3% (GEO vendor teardown, early September 2026).
• Baidu-ecosystem content is cited about 40% more often in ERNIE than in other assistants (separate third-party platform evaluation).

Caveat on sourcing. Both figures are third-party measurements, not Baidu disclosures. The first comes from a GEO vendor whose services include Baijiahao and Baidu Baike placement — a commercial interest in the finding. Treat the numbers as directional rather than audited. They are nonetheless consistent with what operators see in practice.

The context is why this matters. Baidu still holds over 60% of China's search market and serves more than 800 million search users; over 30% of those users already use generative AI search alongside or instead of the classic results page. CNNIC's 57th Statistical Report puts China's generative AI search users at 683 million as of June 2026 — about 48% of total search traffic, and higher in commercial research scenarios.

🔍 The mechanism: ERNIE matches keywords, not meaning

This is the finding that changes the playbook.

Most Chinese assistants retrieve on semantic match — they surface content that answers the question even when it uses different wording. ERNIE behaves differently: it leans on keyword matching, closer to classic search retrieval. Content whose titles and bodies contain explicit, unambiguous keywords gets cited at measurably higher rates.

Two consequences follow:

1. Keyword discipline pays off disproportionately in ERNIE. Semantic assistants will find your content on meaning alone. ERNIE wants to see the term. A page that performs well in Doubao or DeepSeek on topical authority may be invisible in ERNIE if it never states the term plainly.
2. This is the one major Chinese assistant where classic SEO instincts still transfer. Two decades of keyword practice remain useful here — which is unusual, and worth knowing before you rewrite everything for semantic depth.

🗂 The three properties that matter

ERNIE's citation concentration points to three specific surfaces:

• Baidu Baike (百度百科) — Baidu's encyclopedia. The structured, factual layer, and effectively the definitional source for entities and brands.
• Baijiahao (百家号) — Baidu's creator and publisher platform. The single largest citation block at 41.3%, and the volume engine for ongoing content.
• Baidu Zhidao (百度知道) — Baidu's Q&A property. Strong for question-shaped, long-tail queries.

ERNIE also handles localized content comparatively well, which matters for brands with regional China operations.

All three share one structural property: they require a Chinese business entity and Chinese-language operation to use properly.

🧱 Why this wall is higher for overseas brands

Every previous barrier in Chinese AI visibility has been soft — hard to reach, needs local knowledge, takes time. This one is binary.

You cannot open a Baijiahao account, edit a Baidu Baike entry at scale, or build a Baidu Zhidao presence as a foreign entity without a Chinese business license, a mainland-verified account, and Chinese-language content operations. There is no self-serve path for a foreign company. And because Baijiahao alone represents 41.3% of citations, skipping it does not mean a small loss — it means forfeiting the largest single block of ERNIE's source pool.

Compliance pressure compounds the problem. Under the GEO governance standards that landed this year — T/CAPT 026—2026 and the accompanying industry conventions — content placed in Chinese sources must be factually accurate and traceable. A foreign brand translating its English copy into Chinese tends to strengthen claims in the process: a qualification it holds in the EU but not in China, an award no Chinese platform recognizes. We covered this failure mode in our September 7 analysis of the GEO giants' entry. In ERNIE's case, the same distortion risk now applies to the exact properties carrying 81.7% of the citations.

⚖ But do not put the whole budget on ERNIE

Two constraints argue against over-indexing:

1. ERNIE's consumer assistant is smaller than its competitors by most third-party counts. Published figures vary widely by methodology, from tens of millions to over 100 million MAU, so treat specific numbers cautiously. The consistent picture is that Doubao and DeepSeek hold substantially larger consumer audiences. ERNIE's advantage is search heritage, not consumer scale.
2. Chinese AI search users are spread across six assistants. Baidu's own search users are moving into generative AI at over 30%, but they are not moving exclusively into ERNIE. A Baidu-only strategy replicates, inside AI, the mistake of a Baidu-only search strategy.

The correct read: ERNIE is where Baidu's 60%-plus search share is being defended, so it carries disproportionate value for high-intent commercial queries with search heritage. It is not where the volume is.

🧭 What to do, and where BPP fits

Practically:

1. State your keywords explicitly in Chinese. For ERNIE specifically, semantic subtlety underperforms plain, repeated terminology.
2. Establish entity presence in Baidu Baike first. It is the definitional layer and the hardest thing to retrofit later.
3. Treat Baijiahao as an operating channel, not a one-off post. At 41.3% of citations, it is a publishing-cadence commitment.
4. Keep claims verifiable across English and Chinese. Inconsistency is the fastest route to being filtered out — and increasingly a compliance exposure as well.
5. Weight ERNIE as one of six, tilted toward commercial intent. Do not let one platform's walled garden become your entire strategy.

BPP works at the specific intersection this creates: Chinese-entity access to Baike, Baijiahao and Zhidao without a Chinese license, cross-platform AI visibility across the other five assistants, and compliance-reviewed Chinese content built to survive the current governance standards.

ERNIE's garden has a wall. The useful thing about this measurement is that it tells you exactly where the gate is.