The 755M Blind Spot: China's Preinstalled Phone Assistants Are Bigger Than Every AI App You Measure
Every brand visibility dashboard for China — including ours — starts from the same list: Doubao, DeepSeek, Qwen, Kimi, Wenxin, Hunyuan. Six downloaded apps. That list covers 499 million monthly active users, the figure QuestMobile reported for China's native AI app market at H1 2026.
There is a larger surface that almost nobody scores. Preinstalled handset assistants — Huawei's Xiaoyi, Xiaomi's Super Xiaoai, Honor's YOYO, vivo's BlueLM assistant — reach 755 million MAU, according to the same QuestMobile data set. That is 51% more reach than the entire standalone AI app category combined, and roughly double the 382.3M MAU that Doubao, the category leader, posted in June 2026.
If your China GEO program tracks six apps, you are measuring the smaller half of the market.
The reach-versus-intensity split
The reason this surface gets ignored is that engagement metrics make it look weak. Standalone AI apps are opened 92.7 times per month per user; system assistants, 51.4 times. Standalone apps command roughly 18 times as much total usage time.
That gap is real, and it is why the "Doubao is the market" framing survives. But it measures the wrong thing for brand discovery.
| Standalone AI apps | Preinstalled assistants | |
|---|---|---|
| MAU | 499M | 755M |
| Opens per user/month | 92.7 | 51.4 |
| Relative usage time | ~18× | 1× |
| Typical session | Long, exploratory | Short, task-triggered |
Long exploratory sessions are where users write essays, debug code, and plan trips. Short task-triggered sessions are where users ask "which brand of X should I buy" on the way to a decision. Time-on-app rewards the first behavior. Brand recommendation lives disproportionately in the second.
Put differently: 755 million people, opening a query surface 51 times a month, is over 38 billion assistant sessions per month in China. Even if brand-relevant queries are a low single-digit percentage of that volume, the absolute number dwarfs what most brands are optimizing for.
Why it's an independent surface, not a mirror
The common objection: OEM assistants are thin wrappers over the same foundation models, so a good Doubao score transfers. In 2025 that was mostly true. In 2026 it stopped being true, for three reasons.
1. The routing layer is proprietary. Each OEM decides which model handles which query class. A shopping query, a local-services query, and a general knowledge query may hit three different backends on the same phone. Your brand's score in one model tells you nothing about which queries that model actually receives.
2. Third-party agents now sit at system level. In August 2026, Alipay's Abao agent was reported to be integrating into Huawei Xiaoyi, vivo BlueLM, and Honor YOYO — giving Alibaba an agent presence across major handset brands simultaneously. When a commerce agent is wired into the OS assistant, product and brand queries can be intercepted before they ever reach a general model. The answer a user gets is shaped by Alipay's merchant graph, not by the citation web that determines your Doubao score.
3. Regulatory approval created a distinct product category. Seven on-device AI services cleared Chinese regulatory approval, with Apple, Huawei, and Xiaomi among the first licensed. On-device inference means part of the answer is generated from a locally resident model with its own knowledge cutoff and its own tool-calling permissions — a genuinely different retrieval path from the cloud API your monitoring tool queries.
Xiaomi's Super Xiaoai reports 74.4% penetration among its own users. That is distribution no downloaded app can buy.
The vertical-app detour brands should skip
The more common recommendation circulating among China marketing agencies right now is that brands should chase vertical AI apps — education, design, shopping assistants — because their growth rates look spectacular. Doubao Aixue (ByteDance's education app) grew 142.4% year over year. Tencent's Yuanbao grew 100.9%.
The growth rates are real. The absolute numbers are not competitive: no vertical AI app cleared 20M MAU in the H1 2026 rankings, despite three-digit growth. A category where the largest player is under 20 million users is not where a multi-market brand should allocate its second tranche of GEO budget when a 755-million-user surface is unmeasured.
Meanwhile Kimi contracted 42.7% year over year — a reminder that standalone AI app rankings reshuffle fast enough that building a program around any single app's current position is fragile. Handset preinstallation is the opposite: it changes at the pace of device replacement cycles, which in China runs roughly 30 months.
What actually moves the needle on this surface
The citation mechanics differ from downloaded apps. A July 2026 study of citation source distribution found Doubao pulls about 60% of citations from ByteDance properties (Toutiao 35%, Douyin 25%), ERNIE about 55% from the Baidu ecosystem (Baijiahao 35%, Baidu Baike 20%), DeepSeek leaning on CSDN (20%) and Zhihu (15%), and Kimi favoring Zhihu (18%).
OEM assistants route across several of these at once, plus their own ecosystem sources — Huawei's Petal Search index, Xiaomi's app and IoT graph, the handset maker's own shopping channel. Three practical implications:
Entity consistency matters more than any single-platform play. When a query might be answered by three different backends depending on routing, the only durable defense is that your brand name, category, positioning, and factual attributes are identical across Baidu Baike, Toutiao, Zhihu, and the OEM's own product database. Inconsistency that costs you a few points in one app can cost you the entire answer here, because the routing layer resolves ambiguity by picking whichever entity record it trusts.
Merchant-side data is now GEO data. With commerce agents at system level, the fields you control in Tmall, JD, Douyin Mall, and Alipay's mini-program merchant backend feed answers directly. Price, availability, official-store status, and category tags are no longer just e-commerce hygiene — they are AI recommendation inputs.
Test on-device, not just via API. Every monitoring product in this space, ours included, queries cloud APIs. Nothing in an API response tells you what a Honor phone says out loud. Quarterly manual testing on two or three of the top handset brands is currently the only way to see this surface at all, and it takes an afternoon.
How to run the handset audit without a budget line
The audit is cheap enough that there is no procurement argument against it. What it requires is discipline about what gets recorded.
Use three devices covering the largest installed bases — a Huawei running HarmonyOS 5.0 or later, a Xiaomi, and an Honor. Ask the same ten queries on each, in Chinese, spoken rather than typed, because voice invocation is the native path for these assistants and can route differently from text entry. Five queries should be category-level ("which brand of X is best for Y"), three should be brand-level ("is brand Z any good"), and two should be factual ("how much does product P cost", "does brand Z have a store in Chengdu").
Record four things per answer: whether your brand appeared, in what position, which competitors appeared alongside it, and — most important — whether the assistant handed the query to a visible sub-agent or shopping card rather than answering in prose. That last field is the one nothing else in your stack captures. A query that gets intercepted by a commerce agent is not a model-visibility problem at all; it is a merchant-data problem, and the fix lives in a different team's backlog.
Two failure patterns show up repeatedly in early testing. The first is the stale entity: an assistant returns a brand description matching a positioning the company retired two years ago, because an OEM knowledge graph refreshes far more slowly than a cloud model. The second is the wrong-market answer: a global brand's international product line surfaces instead of its China SKU range, which is an entity-disambiguation failure rather than a ranking failure. Neither is visible in an API-based score.
Log the results in the same sheet as your model scores, with a clear column marking them as manually collected. The point is not precision — it is having a second reading that can contradict the first.
Takeaway for Brand Marketers
The honest position: this surface is under-measured industry-wide, and any vendor claiming a comprehensive preinstalled-assistant score today is overstating what is technically available. That does not make it safe to ignore — it makes it the largest uninstrumented risk in a China GEO program.
Three concrete moves for Q4 2026 planning:
- Do not reallocate budget from the six core models yet. They remain the measurable, optimizable core, and the entity work you do for them is the same work that protects you on handsets.
- Add a manual handset audit to your quarterly cadence. Ten brand and category queries, spoken to Xiaoyi, Super Xiaoai, and YOYO, logged verbatim. If the answers diverge sharply from your Doubao and Qwen scores, you have found a routing problem worth escalating.
- Audit merchant-backend fields as GEO assets. Assign an owner to product data accuracy across Tmall, JD, Douyin Mall, and Alipay. In an agent-mediated purchase, a stale price field is a lost recommendation.
The brands that win China AI visibility in 2027 will not be the ones with the highest Doubao score. They will be the ones whose entity data is consistent enough that it survives being routed to a backend nobody predicted.
Related: see how the six core models score international brands on hubGEO's brand rankings, and our earlier analysis of China's AI query routing layer.