Audience-Weighting Barely Moves Your China AI Brand Score. The Consensus Gap Does. (2026)

We re-ran the numbers on the GEO Hub panel of 39 international brands across six Chinese AI models, this time weighting each model by its real user base instead of treating them equally. The result was almost boring: the weighted and unweighted rankings correlate at 0.9994, and no brand moves more than 2.8 points. The result that matters is a different one. Brands averaging 80+ get nearly the same verdict from all six models (a 6.5-point spread on average), while mid-tier brands face a 12.2-point spread. Visibility in China's AI search is not just a level problem. It is a consensus problem, and consensus is something only strong brands have.

What We Measured

GEO Hub's /brands panel scores each brand from 0 to 100 on Doubao, Kimi, DeepSeek, Qwen, Wenxin and Hunyuan, based on 12 category-relevant queries per brand (last refreshed June 4, 2026). The panel covers 39 brands in six categories: auto, beauty, electronics, food, luxury and sportswear. It is a small, curated panel, so read the findings below as directional patterns rather than market-wide statistics.

For the audience weights we used QuestMobile's H1 2026 AI-native app ranking, which Sina Finance reported on August 5, 2026 using June data: Doubao 382 million monthly active users, Qwen 167.18 million, DeepSeek 129.82 million, Yuanbao 49.84 million, and Kimi 7.29 million. We mapped Yuanbao's audience to the Hunyuan score, since Yuanbao runs on Tencent's models. Wenxin did not appear in that top-10 list, so it received zero weight. This is a simplification, but it is the most defensible one available from public data.

Finding 1: Weighting by Audience Changes Almost Nothing

Doubao alone holds about 52% of the weighted audience (382 of roughly 736 million combined MAU across the five ranked apps). Kimi holds under 1%. Intuitively, that should reshuffle brand rankings dramatically. It does not.

Across the 39 brands, the audience-weighted score differs from the simple six-model average by between -0.84 and +2.79 points. The biggest upward moves were Tesla (68.5 to 71.3), Porsche (58.8 to 61.3) and SK-II (76.0 to 78.4), all brands that score slightly better on Doubao and Qwen than on Kimi and Wenxin. The only brands that slip are Volvo (-0.84), Lexus (-0.67) and Shiseido (-0.50), and none of them changes position in its category.

What this means: the panel's six models are broadly aligned in how they rank brands. Whether you obsess over Doubao's 382 million users or treat all six equally, you will land on the same ordering. Brand-versus-brand rank is stable. The variance lives somewhere else.

Finding 2: The Consensus Gap

Here is where the data gets interesting. We calculated the spread (highest model score minus lowest model score) for every brand, then grouped brands by average score.

The correlation between a brand's average score and its spread is -0.52. The strongest brands are the most consistent: KFC, McDonald's and Starbucks all show only a 4-point spread across six models, and Apple shows 5. In the middle of the pack, Tesla's spread is 20 points (58 on Wenxin, 78 on DeepSeek) and Porsche's is 18 (50 on Wenxin, 68 on DeepSeek).

The mechanism is plausible and worth spelling out. Chinese AI models draw on different retrieval stacks and source mixes, so a brand with thin or inconsistent presence gets a different answer depending on which retrieval path fires. A brand with saturated presence across Baidu Baike, e-commerce listings, review platforms and news gets pulled in by every path, so the models converge. For a mid-tier brand, which model your customer happens to open is effectively a coin flip with a 12-point swing.

Finding 3: Wenxin Is the Floor, Qwen and DeepSeek Are the Ceiling

Who gives the highest and lowest score to each brand is highly lopsided in this panel:

• Wenxin issued the lowest score for 27 of 39 brands (69%), and Kimi the lowest for 11 more.
• Qwen or DeepSeek issued the highest score for 33 of 39 brands (85%). Qwen led for 17 brands, DeepSeek for 16.
• Doubao was the top scorer for only two brands.

Average scores by model reinforce this: DeepSeek 70.4, Qwen 69.3, Doubao 68.7, Hunyuan 67.1, Kimi 65.4, Wenxin 64.5. A 6-point band between the most and least generous model is small on paper, but it is consistent enough that it shows up in nearly every brand's profile. We covered the Wenxin side of this in our earlier analysis of the Wenxin discount. The new angle here is that Wenxin's strictness hits mid-tier brands hardest, because strong brands have enough evidence to clear any model's bar.

This also reframes the Doubao priority. Doubao has the largest audience, and it is not the most generous scorer. In the weighted view, it counts for more than half the audience but typically sits near the middle of the pack on score. Brands should treat Doubao as the volume channel to fix, not the easiest one to win.

Finding 4: Category Dispersion Is Not Category Strength

Category averages tell a clean story. Food leads at 79.0 average (with a tight 5.5-point average spread), followed by luxury (75.7), electronics (69.4), auto (68.4), sportswear (66.0) and beauty (55.8). But the gap between the best and worst brand within each category is the more actionable number: beauty has a 76-point gap between Lancôme (89.8) and Drunk Elephant (13.8), auto 49.8 points, luxury 48.5, and food 54.0.

Beauty is the extreme case: the same category contains both the highest-consensus brands (Lancôme, Estée Lauder, L'Oréal at 84.5 to 89.8) and the lowest-scoring ones (Drunk Elephant, Charlotte Tilbury, NARS at 13.8 to 27.2). If you are a challenger beauty brand, the category's average is irrelevant; what matters is that the models have made a hard binary judgement about which brands belong in the answer.

Why the Market Context Raises the Stakes

The user base is also not static. The TMTPost analysis of China's AI market, published June 17, 2026, notes that Doubao, Qwen, DeepSeek and Yuanbao together exceed 710 million (non-deduplicated) monthly users and take close to 90% of the top-10 AI app share. It also reports DeepSeek's MAU down 24% year over year to 127 million even as per-user frequency rose, and that Doubao's new paid tiers (68, 200 and 500 yuan a month) coincided with roughly 6.1 million fewer active users in a single month. Audience mix is moving, which means the model-by-model weights we used above will shift again. A consensus-built brand is protected from that churn. A brand whose visibility depends on one model's generosity is not.

Takeaway for Brand Marketers

1. Stop optimising for a single model score. Because weighted and unweighted rankings are nearly identical (r = 0.9994), you gain little by tuning to one platform's audience. Track the spread, not just the average. A spread above 12 points is a signal that your presence is uneven across retrieval sources.
2. Fix the weakest model first if you are mid-tier. For brands in the 50 to 79 band, the lowest-scoring model, usually Wenxin or Kimi, is where the gap is largest. Closing a 12-point gap there usually costs less than adding 5 points on a model where you already score well.
3. Build the shared evidence base. Consensus comes from presence on the sources every model uses: encyclopedia entries, major e-commerce listings, review and shopping-guide sites, and authoritative news. This is the same base that makes McDonald's and Apple look uniform.
4. Use the spread as a monitoring metric. If your average is flat but your spread widens, one model has changed how it sees you. Treat that as an early warning before the average moves.
5. Do not mistake audience size for priority. Doubao's 52% share makes it the volume channel, but it is not where brands win easiest. Qwen and DeepSeek are the most generous scorers in this panel and the two likeliest to reward quick fixes.

Caveat: this analysis rests on 39 brands and 12 queries per brand, and the audience weights come from third-party estimates. We would expect the direction of the findings to hold on a larger panel, but the exact numbers will move.

Related: Browse all brand scores on /brands, The Six-Platform Brand Score Gap, The Wenxin Discount.