China's Next Brand Visibility Battle Isn't Inside a Chatbot — It's Inside the Phone (2026)

By the end of 2025, StepFun's foundation models were already running natively on roughly 60% of China's top smartphone brands, pre-installed on more than 42 million devices and answering close to 20 million queries a day. In July 2026, the Beijing-based model maker stopped being a backend supplier and became a consumer brand in its own right: it launched STEPX, billed as the world's first "AI-native" agent phone brand, alongside its own agent operating system and a flagship device, STEPX Neo. IDC now forecasts that China will ship 147 million next-generation AI agent phones in 2026 alone.

For brand marketers who have spent 2025 and early 2026 learning to optimize for Doubao, DeepSeek, Kimi and Qwen, this is a different kind of visibility problem — and most brand GEO programs aren't built for it yet.

StepFun (阶跃星辰) is not a fringe player experimenting with hardware. Founded in 2023 by Jiang Daxin, a former Microsoft global vice president, the company built a full Step-series model family spanning language, multimodal and reasoning tasks before pivoting toward device-native distribution. That pivot is the part brand teams should pay attention to: StepFun didn't wait to win the model-quality race against DeepSeek or Qwen. It won distribution first, by embedding directly into phone OEMs' software stacks, and only then launched a consumer-facing brand around it.

Two Battlegrounds, Not One

Most of the China AI-visibility conversation so far has focused on chat apps: Doubao's monthly active users surged past 380 million in mid-2026 (QuestMobile), Qwen sits around 166-170 million, DeepSeek around 127-130 million, and Kimi has fallen out of the top tier entirely after peaking as a breakout star in 2025. Brands have responded by building content and structured data aimed at getting cited when a user asks one of these apps a question.

StepFun's STEPX represents a structurally different layer: an agent embedded at the operating-system level of the phone itself, not a chat app a user has to open. Instead of a brand hoping to be recommended when a consumer types a query, the phone's own agent decides what to surface, what to compare, and increasingly, what to do — book, order, or purchase — often without ever showing the user a ranked list of options.

From "What Gets Cited" to "What Gets Done"

The practical difference matters more than the technical one. When a chat app cites a brand, the user still sees the brand name, compares it against alternatives, and clicks through. That is a recognizable, if AI-mediated, version of search — brands can measure share of citation, track competitors, and optimize content the way they've optimized for GEO all year.

An agent phone changes the transaction. If STEPX's agent is asked to "find me a good air purifier under 2,000 yuan" or "book a flight to Chengdu next Friday," the ideal outcome for the platform is that the agent simply completes the task — comparing merchant integrations, price feeds, and inventory APIs the user never sees. The brand that wins isn't necessarily the one with the best content; it's the one the agent's backend can transact with cleanly. This is closer to how a delivery app's default restaurant works than how a search engine's top result works.

That shift is already visible in how StepFun and its OEM partners talk about the product: the launch narrative emphasizes an "agent system" that handles multi-step tasks end-to-end, not a smarter way to answer questions. For a brand, invisibility in this layer doesn't look like a missing citation — it looks like a sale routed to a competitor with a live inventory API without the brand ever coming up for consideration.

StepFun Isn't Alone — Honor Already Shipped the Preview

Skeptics should note this isn't a single company's launch-day hype. Honor, one of China's largest phone OEMs, has already integrated Doubao directly into agent-capable handsets, letting the phone's assistant handle multi-app tasks — booking a ride, filling out a form, comparing prices across apps — without the user manually switching between them. StepFun's STEPX push is the same distribution logic taken further: instead of one OEM partnering with one chat app, StepFun is positioning its own models as the default agent layer across a wide swath of OEMs simultaneously, then adding its own hardware on top. Two of China's most consequential AI players are converging on the same conclusion from different directions — the phone OS, not the app icon, is where the next round of AI-mediated brand discovery will actually happen.

Why This Lands Harder in China Than Elsewhere

Three factors make this shift more consequential in the China market specifically:

Distribution is pre-installed, not opt-in. Western AI agent experiments (browser agents, OS copilots) still largely require a user to install or activate something. StepFun's models are already the default underlying layer on 60% of top-brand devices before STEPX even launched — the agent phone push builds on infrastructure that already reaches tens of millions of people passively.

The phone, not the browser, is the primary computing surface. China's mobile-first internet means the operating system layer carries far more weight relative to browser-based search than it does in the US or Europe, where desktop and browser extensions still mediate a meaningful share of discovery.

IDC's 147 million-unit forecast implies scale within a single year. That is not a niche early-adopter category — it is a meaningful fraction of China's roughly 1.1 billion smartphone user base transacting through agent-mediated defaults by the end of 2026, layered on top of (not replacing) the existing Doubao/DeepSeek/Qwen chat competition.

What Brands Are Getting Wrong Right Now

Most international brands' China AI-visibility work in 2026 has focused on the app layer: structured product pages, FAQ schema, entity clarity for Doubao and DeepSeek, and monitoring citation share the way you'd monitor a search ranking. That work isn't wasted — chat-app usage isn't going anywhere, and StepFun itself still needs to be understood as one more model to track for citation behavior.

But it is incomplete for the agent-OS layer, because the failure mode is silent. A brand can have excellent citation performance in Doubao and still lose the transaction entirely when a StepFun-powered agent phone completes a purchase through a merchant API the brand was never integrated into. There is no missed citation to notice — there is simply a sale that never surfaced the brand as an option.

There's also a measurement gap most Western marketing teams haven't confronted yet. GEO reporting built around "did we get cited, and how prominently" assumes a visible answer surface — a chat transcript, a generated summary, a list of recommendations a human can screenshot. Agent-completed tasks often produce no such artifact for the brand to audit from the outside. A brand's China team can watch its Doubao citation share climb every week while having no visibility at all into whether StepFun- or Honor-powered agents are quietly routing purchase volume elsewhere. That blind spot is structural, not a reporting failure — it will require brands to work directly with agent-OS platforms or their merchant integration partners to get any signal at all, the same way retailers had to instrument in-app purchase APIs once mobile commerce moved past the mobile web.

Takeaway for Brand Marketers

Two things should change in how China GEO programs are scoped for the rest of 2026. First, treat "agent OS reach" as a separate tracking category from "chat app citation share" — StepFun's 42-million-device installed base and IDC's 147-million-unit 2026 shipment forecast mean this layer is already too large to file under "emerging" and revisit next year. Second, start asking a different diagnostic question than the one that drives most GEO content work: not "would an AI cite us if asked," but "could an AI transact with us without asking anyone." That means auditing whether product, pricing, and availability data is exposed in a form an agent's backend can actually execute against — not just a form a language model can describe in an answer. Brands that treat this purely as a content problem will keep winning citations in Doubao while quietly losing completed transactions to whichever competitor plugged into the agent layer first.

Related: See how international brands rank across China's major AI models on hubGEO's brand tracker.