The Wenxin Discount: Why Baidu's ERNIE Scores International Brands Lower Than Any Other Chinese AI Model (2026)

Across the 39 international brands tracked in the hubGEO database, one pattern is more consistent than any category ranking or platform battle: Baidu's Wenxin (ERNIE) is the toughest room in Chinese AI. Wenxin hands out the lowest visibility score of any of the six tracked models for 27 of 39 brands — 69% of the index. Its average brand score of 64.5 sits nearly six points below DeepSeek's 70.4, the most generous grader in the set. For a brand marketer benchmarking China AI visibility on a single model, that gap is large enough to change the conclusion from "we have a problem" to "we're fine" — or the reverse.

But the more interesting finding is the exception cluster. Four brands — Sony, Shiseido, Lexus, and Volvo — don't just escape the Wenxin discount. They score higher on Wenxin than on any other Chinese AI model. Understanding why reveals something structural about how ERNIE was trained, and what kind of brand equity it rewards.

The Numbers: Six Models, One Systematic Gap

Here are the average visibility scores (0–100) across all 39 tracked international brands, by model:

The spread between DeepSeek and Wenxin — 5.9 points on average — is not random noise. Wenxin scores below a brand's own six-model average by 3.0 points on average, and it is the single lowest scorer for 27 brands. DeepSeek, by contrast, is the highest scorer for 16 brands, more than any other model.

The discount is broad-based. It hits luxury (Chanel: 87 on Wenxin vs. a 90.2 average), sportswear (Nike: 86 vs. 89.5), beauty (Lancôme: 86 vs. 89.8), and even the strongest brand in the entire index (Apple: 92 on Wenxin vs. 96 on Doubao and 97 on DeepSeek). Weaker brands feel it proportionally harder: Drunk Elephant scores just 10 on Wenxin against an already-low 13.8 average, and Lululemon drops to 30 against a 36.8 average.

Why ERNIE Grades Harder: A Training Corpus From a Different Era

Wenxin's knowledge of brands is shaped by what Baidu can see — and Baidu's index skews heavily toward its own content ecosystem: Baidu Baike (the dominant Chinese encyclopedia), Baijiahao (Baidu's publisher platform), Baidu Zhidao (Q&A), and the open Chinese web that Baidu has crawled for two decades.

What Baidu largely cannot see is the content universe where most international consumer brands have concentrated their China marketing spend since roughly 2019: Xiaohongshu notes, Douyin videos, and WeChat private-domain content. These platforms are walled gardens with limited or no exposure to Baidu's crawler. A brand whose Chinese-language footprint is social-first — built on KOL seeding, livestream commerce, and short video — is nearly invisible to the corpus that trained ERNIE.

Compare that with Doubao, whose parent ByteDance owns Douyin, or Qwen, whose parent Alibaba sits on Tmall's transaction and review data. Those models have native pipelines into exactly the content international brands have been producing. Wenxin does not. The result is a systematic discount against brands whose China presence is young, social-led, or e-commerce-led — which describes most international brands that entered or re-invested in China in the past five to seven years.

This matters more than Baidu's declining chatbot market share suggests. According to QuestMobile's Q1 2026 data, Doubao leads AI-native apps with roughly 340 million MAU, followed by Qwen at ~170 million and DeepSeek at ~130 million — and Doubao overtook Baidu in user scale for the first time in Q1 2026. But ERNIE's answers are not confined to a chatbot app. They are embedded directly into Baidu Search results, which still processes a large share of China's search queries, particularly from older and lower-tier-city users — the "silver + sinking market" demographic that QuestMobile identifies as the fastest-growing segment of Chinese AI adoption. A Wenxin blind spot is not a minor-model problem; it is a Baidu Search problem.

The Four Exceptions: Sony, Shiseido, Lexus, Volvo

Now the inversion. Four brands score higher on Wenxin than on any other model:

What do these four have in common? None of them is a social-commerce native. All four built their Chinese brand equity in the 2000s and early 2010s — the era when Chinese brand discourse lived on exactly the platforms Baidu indexes best. Sony's decades of product discussion accumulated on tech forums and Baidu Tieba. Lexus and Volvo generated years of owner reviews, comparison threads, and safety discussions on open-web auto verticals and Baidu Zhidao. Shiseido entered China in 1981 and has forty years of open-web editorial coverage, counter-brand history, and Baike depth that no Xiaohongshu-era beauty brand can match.

In other words: Wenxin is a seniority test. It rewards brands whose Chinese-language authority content predates the walled-garden era, and discounts brands whose equity lives inside platforms Baidu can't crawl. It is no coincidence that the exception list is dominated by Japanese and European brands with conservative, durable, forum-era reputations — and that the biggest Wenxin discounts hit Instagram-era prestige beauty brands like Drunk Elephant (10), Charlotte Tilbury (15), and NARS (22).

What This Means for Cross-Model Benchmarking

Three practical distortions follow from the Wenxin discount:

Single-model audits mislead. A brand that checks its China AI visibility only on DeepSeek (a common choice, since it is the most internationally famous Chinese model) will see a number that runs ~6 points hot relative to Wenxin. A brand that tests only on Wenxin will conclude things are worse than they are on the platforms where most Chinese AI queries actually happen.

The discount also compounds at the bottom of the index. For brands scoring above 85, the Wenxin gap is an accounting footnote — Apple at 92 instead of 97 is still ubiquitous. For mid-tier brands, the same relative discount can push them below the visibility threshold where AI models stop mentioning them in recommendation-style answers altogether. MAC Cosmetics at 38 on Wenxin versus 52 on Qwen is the difference between occasionally appearing in makeup recommendations and effectively not existing in them.

The discount is a proxy for open-web deficit. Because Wenxin's corpus is the closest thing to "the open Chinese web," a large gap between a brand's Wenxin score and its Doubao/Qwen scores is a diagnostic signal: the brand's Chinese content strategy is over-concentrated in walled gardens. Tesla shows the widest cross-model spread in the entire index — 20 points, from 78 on DeepSeek down to 58 on Wenxin — consistent with a brand whose Chinese discourse is dense on social platforms and news cycles but whose open-web authority layer is thin relative to its fame.

Wenxin strength predicts durability. The four exception brands hold visibility built on content that doesn't decay when an algorithm changes or a KOL contract ends. Baike entries, Zhidao answers, and forum archives persist and keep feeding every future ERNIE training run.

Takeaway for Brand Marketers

First, benchmark on all six models, and treat the Wenxin-vs-rest gap as a standing KPI. If your Wenxin score trails your six-model average by more than the index-wide 3 points, you have an open-web content deficit, not just a "Baidu problem."

Second, close the deficit with crawlable Mandarin authority content: a complete and current Baidu Baike entry, editorial coverage on open-web media that Baidu indexes (36Kr, industry verticals, auto/beauty portals), and structured Q&A-style content that mirrors how Zhidao-era answers were written. This is unglamorous compared to Xiaohongshu seeding, but it is the only content pathway ERNIE reliably sees.

Third, don't deprioritize Wenxin because Doubao won the MAU war. ERNIE's distribution runs through Baidu Search, and its user base skews toward the older, lower-tier-city demographics that are the fastest-growing cohort of Chinese AI users — and often the highest-value customers for auto, appliance, and premium retail brands.

Fourth, study the Shiseido playbook. The Wenxin premium enjoyed by forty-year China veterans can't be replicated overnight, but its mechanism — durable, open-web, third-party-validated content — can be started today. Every quarter of delay is another ERNIE training cycle that learns China's brand landscape without you.

Related: Explore full brand scores across all six Chinese AI models