China's AI Fourth-Place Cliff: Why Three Models Decide Your Brand Visibility in 2026

QuestMobile's June 2026 ranking of China's AI-native apps contains one number brand marketers should read twice: the fourth-ranked app, Tencent's Yuanbao, has 49.84 million monthly active users. The third-ranked app, DeepSeek, has 129.82 million. That is a 2.6x drop in a single rank. Below Yuanbao the curve collapses further: Ant Afu at 28.97 million, Doubao Aixue at 19.44 million, and Kimi at just 7.29 million. For international brands deciding where to spend limited China GEO budget, the market has a very clear shape: three models at scale, then a cliff.

This analysis uses the QuestMobile figures, published in its H1 2026 AI application insight report and reported by IT之家 and HK01, to work out what the distribution means for brand visibility work. The ratios and shares below are our own calculations from those published figures.

The Numbers: A Power Law, Not a Market of Equals

Two things stand out. First, Doubao alone has roughly 2.3 times Qwen's audience, which QuestMobile itself highlights as evidence of market-leading dominance. Second, the top three apps together account for about 679 million MAU, or roughly 87% of the combined MAU of the seven ranked apps that have published figures (about 785 million). Doubao by itself is about 49% of that pool.

Growth rates sharpen the picture. QuestMobile reports Doubao up 172.1% year over year, and Qwen up an extraordinary 5,792.9% from a very small base. Qwen's surge matters less as a headline than as a warning: any brand benchmark taken before 2026 was measured against a Qwen that barely existed as a consumer product.

Why the Sum Overstates the Audience, and Why That Matters

A careful reader will notice a problem. QuestMobile's May 2026 data puts total native AI app MAU at 499 million, up 85.4% year over year. Yet the individual June app figures for the top seven already sum to roughly 785 million. The gap is not an error; it is overlap. Chinese AI users are multi-homing, running Doubao for creative and daily tasks, DeepSeek for reasoning-heavy questions, and Qwen for shopping-adjacent queries inside the Alibaba ecosystem. (The two figures are from adjacent months, so treat the ratio as indicative, not exact.)

Our rough reading: the average native AI user touches at least 1.5 apps. For brands this has an uncomfortable implication. A buyer who asks the same category question in two or three models can receive two or three different shortlists. Being strong in one model and absent in the other two is not a 33% presence; it can mean the user sees a competitor in most of their comparisons. This is the logic behind the cross-platform score gaps we track on the hubGEO /brands pages, where the same brand frequently scores very differently across models.

Depth of Use: The Quiet Counter-Signal

Audience size is not the only dimension. QuestMobile also measured the share of users who spend more than ten minutes per session: DeepSeek 30.0% (up 7.8% year over year), Doubao 27.5% (up 2.2%), Kimi 26.1% (up 8.4%). Across all native apps, users average 92.7 sessions and 183 minutes per month, with session duration up 40% year over year.

Read together with the MAU table, this explains why the cliff is less brutal than it looks for some categories. Kimi's audience is 52 times smaller than Doubao's, yet a quarter of its users stay in long sessions, which is consistent with the research-heavy, higher-consideration usage we have discussed before. A small, deep audience can matter for B2B, financial services, or premium durable goods. For mass consumer categories such as beauty, food and apparel, it rarely offsets the reach gap.

What the Cliff Means for Budget Allocation

We would frame the allocation problem in three tiers.

Tier 1 — Must-win (Doubao, Qwen, DeepSeek). These three hold the audience. Within them the signals differ: Doubao rewards content that flows through the ByteDance ecosystem, Qwen draws on Alibaba commerce signals, and DeepSeek leans on authoritative web sources. A brand with a single content strategy applied uniformly will under-perform in at least one of them. The practical minimum is a documented plan per model.

Tier 2 — Monitor and selectively invest (Yuanbao). At 49.84 million MAU, Yuanbao is small next to the top three but roughly 6.8 times Kimi's size, and it sits inside Tencent's WeChat distribution. If your category depends on WeChat-led discovery, such as mini-program commerce or official-account content, Yuanbao deserves more weight than its MAU suggests.

Tier 3 — Vertical and niche (Ant Afu, Doubao Aixue, Kimi). Ant Afu (financial and tax questions) and Doubao Aixue (education) are vertical tools. They matter enormously if you sell insurance, wealth products, or educational services, and almost not at all otherwise. Kimi's 7.29 million MAU makes it a specialist channel, not a pillar.

Three Mistakes the Cliff Exposes

1. Equal-weighting six models. Many measurement dashboards, ours included, track six or more models for completeness. Completeness is not a budget strategy. A one-point gain in a model with 7 million users is worth far less than the same gain in a model with 382 million.
2. Treating MAU as the only weight. Sessions longer than ten minutes and category fit change the math. Weight by relevance to your buyers, not by headline audience.
3. Benchmarking against stale data. With Qwen up nearly 58-fold in a year, any competitive score set from 2025 needs to be re-run before it informs a 2027 plan.

How to Test the Cliff in Your Own Category

The cleanest way to turn this analysis into action is a simple three-step audit. First, write ten buyer-style questions for your category in Chinese, the way a real shopper would ask them: comparisons, "which is better" questions, price-bracket questions, and trust questions. Second, run each question in Doubao, Qwen and DeepSeek on the same day, and log whether your brand is named, in what position, and from which cited source. Third, repeat the exercise for one Tier 2 or Tier 3 model chosen for buyer fit.

Most international brands that run this audit find the same pattern: a strong showing in one model, often the one where their existing PR or e-commerce footprint is deepest, and near-invisibility in the others. That asymmetry is the practical meaning of the cliff. Because the top three models together reach about 679 million monthly users in QuestMobile's data, a gap in any one of them is a gap in front of well over 100 million people. By contrast, a gap in Kimi, with 7.29 million users, is a rounding error for a mass-market brand and a meaningful loss only for specialist categories.

It is also worth watching the ratios over time rather than the absolute scores. If Doubao's lead over Qwen narrows from 2.3x toward 1.5x, the budget split between the two should move with it. Treat the 2.3x figure as a snapshot, not a constant.

Takeaway for Brand Marketers

• Build the plan around three models. Put at least 80% of China AI visibility effort into Doubao, Qwen and DeepSeek, and write a distinct content and source plan for each.
• Test for overlap, not averages. Run the same ten category queries across all three models and record who appears in each. The brands that appear in all three are your true share of voice; the ones that appear in one are exposed.
• Choose a Tier 2 or 3 model on purpose. Add Yuanbao if WeChat is central to your funnel, Ant Afu or Doubao Aixue if you sell in finance or education, and Kimi if your buyers are researchers or professionals. Do it because of buyer fit, not for completeness.
• Re-baseline quarterly. The ranking shifted sharply within a year; assume it will again.

Related: see your brand's per-model scores on the hubGEO /brands page, and our earlier analyses of the Qwen growth rebaseline and the MAU-versus-engagement budget trap.

Source: QuestMobile H1 2026 AI Application Market Insight Report (figures via IT之家 and HK01). Ratios, shares and overlap estimates are hubGEO calculations.