Two Minutes to Win: China's AI Sessions Are Short, and Brand Shortlists Are Set Inside Them (2026)

QuestMobile's H1 2026 data contains one number that most brand dashboards ignore: the average Chinese AI-native app user opens their assistant 92.7 times a month and spends 183 minutes in total. Divide one by the other and the typical session lasts roughly 2 minutes. In a two-minute exchange, a model has time to produce one answer, maybe one follow-up. For international brands, that means the shortlist is usually built in a single response, not a long research journey.

This article works through what the engagement data implies for brand visibility across Doubao, Qwen and DeepSeek, and where the exceptions are. Note: the per-session average is our own calculation from QuestMobile's published totals; the interpretation is our analysis, not QuestMobile's.

The numbers behind the two-minute session

QuestMobile's 2026 H1 AI application report (published 14 July 2026 and summarized by IT之家) gives the following picture of the market:

Time spent is up 40% year over year, but 92.7 sessions a month is about three openings a day. The assistant has become a utility checked in passing, much like a search box or a weather app, rather than a destination people settle into. That matters because a utility session is a decision-compression device: the user wants one usable answer fast.

Compare that with the report's separate AI-plugin and terminal-assistant categories (644 million and 755 million monthly actives respectively). The wider AI surface is even more embedded in other tasks, which pushes sessions shorter still. Our earlier piece on preinstalled phone assistants covers that layer.

Deep sessions are a minority, and they are not evenly distributed

The same report breaks out users who spend 10 minutes or more in a session:

Two things stand out. First, even on the "stickiest" model, roughly seven in ten users do not reach ten minutes. Second, absolute scale reverses the ranking. If the percentage applies to each app's monthly actives (an assumption the report summary does not spell out), Doubao's 27.5% of 382 million is roughly 105 million deep-session users, while DeepSeek's 30.0% of 130 million is roughly 39 million. Doubao has about 2.7 times as many long-session users as DeepSeek, despite the lower rate.

For brand marketers this splits the funnel in an uncomfortable way. DeepSeek and Kimi produce the highest depth per user; Doubao produces the highest depth in total and the highest breadth. A brand that appears on the first screen of a Doubao answer reaches more deliberating shoppers than one that wins a long DeepSeek reasoning chain.

What a two-minute session does to a brand shortlist

Based on how these assistants typically respond, a short session has three consequences.

1. The first answer is the shortlist. A query like "which sunscreen is best for sensitive skin" returns three to five named products. A user who has 2 minutes will pick from that list, ask one clarifier, and leave. Brands outside the list rarely get a second chance in the same session. Our analysis of reasoning-model query fan-out showed the same elimination dynamic from the supply side.

2. Entity clarity beats content volume. In a short session the model cannot afford ambiguity. Brands with a consistent Chinese name, a clear category association and agreeing reference pages (Baidu Baike, Zhihu, official site) are cheap for a model to include. Brands with three translated names or conflicting descriptions cost reasoning effort and tend to be skipped. See the three-Baike problem and the alias gap.

3. Follow-up questions are where depth models matter. The 26–30% of users who go past ten minutes are disproportionately doing comparison work: spec checks, price comparisons, "A versus B" questions. This is where DeepSeek and Kimi quotes of detailed, citable sources become important, and where thin brand pages get exposed.

Where each model fits in the brand journey

We read the data as three roles rather than one leaderboard:

A 36Kr hands-on test of five domestic assistants published in May 2026 reached a compatible conclusion: Kimi was rated strongest for daily question answering and sourcing, Doubao was recommended for general reliability, and DeepSeek led on document-heavy and writing tasks, while all five showed "different degrees of hallucinations" when retrieving tutorials. Hallucination risk is another reason short sessions are dangerous: a user who accepts the first answer rarely checks it. Our price hallucination analysis explains the brand-fact side of that risk.

Why the average hides a measurement problem

Most brand tracking still runs a fixed prompt set once and reports a score. That mirrors the long-session, high-intent user. It underweights the short-session majority, whose prompts are vaguer, more conversational and often voice-driven. A brand can score well on carefully written test prompts and still be absent from the casual, two-minute queries that make up most volume.

Practical implication: when you review your brand's presence in the GEO Hub brand rankings, look at it by prompt type, not as a single average. Separate "discovery" prompts (category, no brand named) from "evaluation" prompts (brand named, comparison). A strong evaluation score paired with a weak discovery score means you are being researched after you are already considered, but not placed on the shortlist.

A worked example: how the same budget splits differently

Consider a hypothetical international skincare brand with a fixed GEO content budget. Under a pure MAU approach, spend follows audience size: roughly 56% to Doubao, 25% to Qwen and 19% to DeepSeek (382, 167 and 130 million out of a combined 679 million). Under a pure depth approach, using the deep-session user estimates above, DeepSeek's weight rises sharply because it converts a higher share of its audience into long, comparison-oriented sessions.

Neither split is right on its own. The more useful frame asks which stage of the journey each model owns for your category. For a low-consideration product such as a facial cleanser, most decisions are made in the short session, so Doubao and Qwen visibility on category prompts should dominate the budget. For a high-consideration product such as a premium device, a skincare device or a supplement regimen, the long-session users on DeepSeek and Kimi carry more of the conversion weight, and authoritative long-form sources matter more. The point is that one brand, two product lines, and two different optimal allocations are normal.

This is why a single blended "China AI score" can mislead. The score tells you how you rank on average; it does not tell you whether you rank where your category's buyers actually decide.

Caveats

• The two-minute figure is an average of totals. Real sessions are skewed: many are seconds long, and a minority run past ten minutes, so the mean does not describe a typical single session.
• QuestMobile counts AI-native apps. Usage inside WeChat, Alipay, phone assistants and shopping apps is measured separately and may behave differently.
• The deep-session user counts above are our estimates and rest on the assumption stated earlier.
• Doubao's and Qwen's growth rates come from different base periods and product launches; Qwen's 5,792.9% points to a very small prior-year base, so it says more about the starting point than about steady-state momentum.

Takeaway for Brand Marketers

1. Optimise for the first answer. Audit your brand against short, vague category prompts in Doubao, Qwen and DeepSeek, not only brand-named queries. If you are absent from the first response, you are largely absent from the session.
2. Fix entity consistency first. One Chinese brand name, one category description, matching facts across Baike, Zhihu and your official pages. This is the cheapest lever in a short session.
3. Match content to the model's role. Quotable, compact facts for Doubao; structured commerce data for Qwen; long-form, citable comparison content for DeepSeek and Kimi.
4. Split your reporting. Track discovery and evaluation prompt sets separately. Weight them by likely volume, not by how easy they are to write.
5. Budget by reach times depth. Doubao's scale and DeepSeek's depth are both real; allocating by MAU alone or by engagement rate alone will misdirect spend.

Related: GEO Hub brand rankings, Doubao vs DeepSeek budget analysis, China AI H1 2026 MAU vs engagement.

Sources: QuestMobile 2026 H1 AI Application Market Insight (via IT之家); 36Kr five-assistant comparison, May 2026.