ByteDance's Doubao Phone Assistant shipped September 16 on the RMB 5,999 nubia NaviX Ultra. It compares prices across every commerce app and completes the purchase — a surface no China brand visibility score can sample.

The Execution Query: China's First AI Agent Phone Just Made Brand Recommendation Optional

On September 16, 2026, ZTE's nubia sub-brand put the NaviX Ultra on sale at RMB 5,999 (RMB 5,499 after national subsidy), with every configuration starting at 512GB. It is marketed as the world's first AI agent phone, co-developed with ByteDance, and it ships with the consumer release of Doubao Phone Assistant — which ByteDance announced two days earlier, on September 14.

The specification that matters to brand marketers is not the Snapdragon flagship silicon or the 7,000mAh battery. It is this: a user can say "compare prices across every platform and order me that dress," and the assistant opens multiple commerce apps, runs the search, compares prices, applies available coupons, and completes the purchase. ByteDance puts end-to-end task success above 80%.

That one sentence quietly retires the question every China AI visibility dashboard was built to answer.

Two queries, two different selection functions

Almost every brand visibility score in the China market — including the six-model scoring on hubGEO's own /brands index — is built on a recommendation query. You ask a model "what are the best value projectors," "which luxury watch brands hold value," "recommend a hotel in Shanghai," and you count whether the brand appears, in what position, and on what evidence.

The agent phone runs an execution query. The user does not ask which brand is best. The user states an outcome and delegates the comparison. The agent then performs the shortlist itself — inside Taobao, JD, Douyin's storefront, Meituan — using live price, live stock, and live coupon eligibility, and returns not a list of brands but a completed transaction.

	Recommendation query	Execution query
User input	"Which brand is best for X?"	"Get me the best X, order it"
Where it runs	Chat interface (Doubao, DeepSeek, Qwen app)	On-device, across third-party commerce apps
What gets compared	Brands, as described in training data and cited sources	SKUs, as listed right now in each channel
What wins	Consensus in authoritative Chinese content	Price, stock, coupon stackability, channel presence
Leaves a transcript a GEO tool can sample	Yes	No
Covered by your current brand score	Yes	No

These are not two phrasings of the same question. They are two different selection functions with two different inputs. A brand can score 100 on the first and be structurally unreachable by the second.

The design decision buried in the launch

The most consequential detail about this device was not in the launch keynote. As reported by LatePost and summarized by IT Home ahead of the September 16 launch, nubia and ByteDance changed how the assistant cooperates with Alibaba's and Tencent's super-apps: the shipping product does not read the user's screen and simulate human taps — the GUI-agent approach that most Western agent demos rely on.

That choice reroutes the entire brand-visibility question.

A screen-reading GUI agent is, in effect, brand-neutral: if a merchant is visible to a human eye inside an app, it is visible to the agent. Whatever the agent can see, it can compare. Distribution and content quality determine outcomes, roughly the way they do for a human shopper.

An agent that operates through negotiated integrations instead has a finite, commercially determined reachable universe. The set of platforms it can transact in, and the set of product attributes it can compare on, is the outcome of contracts between ByteDance, ZTE, and each super-app operator. No amount of Xiaohongshu content, Baike editing, or press syndication moves a brand into a channel the agent cannot call.

For international brands, this is a familiar Chinese market pattern arriving in a new layer. It is the Tmall flagship-store question — where do you actually have a transactable presence — applied to the operating system.

Why no visibility tool will see this happening

Every China brand-visibility product on the market today, ours included, samples chat surfaces. It sends prompts to a model endpoint or an app, reads the answer, scores the brand mentions, and logs the cited sources. That methodology works because a chat answer is a text artifact you can capture.

An execution query produces no such artifact. The comparison happens on-device, inside apps the measurement tool has no access to, and it terminates in an order confirmation rather than a paragraph. There is no answer text to parse, no citation list to attribute, and no API a third party can poll.

This is the same class of blind spot we described in the coverage of China's 755 million preinstalled phone assistants — but worse in one specific way. Preinstalled OEM assistants were at least mostly answering questions. This one is completing purchases. The unmeasured surface has moved from the top of the funnel to the bottom of it.

For scale context: Doubao's chat app is reported at roughly 382.3 million monthly active users in the latest CNNIC-referenced industry data, against Qwen at about 167.2 million, DeepSeek at about 129.8 million, and Tencent Yuanbao at about 49.8 million, inside a generative AI user base that has passed 560 million. Your brand score covers the first number. It covers none of the agent layer.

The RMB 5,999 filter is a feature, not a limitation

The obvious objection is volume. One flagship phone at RMB 5,999 is not 382 million users, and it will not be this year.

That objection misreads who is in the first cohort. A RMB 5,999 device with 512GB minimum storage filters for exactly the consumer segment that international premium brands spend the most to reach: urban, high-income, early-adopting, and disproportionately responsible for discretionary category spend in beauty, travel, consumer electronics, and small luxury.

It also misreads the roadmap. Chinese tech media are already calling the NaviX Ultra "Doubao Phone gen 2," which means ByteDance is iterating a product line, not running an experiment. The assistant itself is a separate software product with its own release track — the consumer version that shipped September 14 is not locked to one handset.

The pattern to expect is the one ByteDance has run before: hardware partner first, capability proven, then the assistant widens across devices while the chat app keeps the reach. Brands that wait for the agent layer to be big enough to measure will be optimizing for it after the integration decisions have already been made.

What actually determines whether an agent can buy your brand

If the execution query is the surface that ends in revenue, the inputs that matter are not the ones GEO vendors currently sell:

1. Transactable channel coverage. The agent can only buy where it can transact. A brand with strong Chinese AI content scores but no Douyin storefront, no JD self-operated listing, or a dormant Tmall flagship is invisible at the moment of execution regardless of its recommendation score.

2. Structured attribute hygiene. Price, stock status, SKU naming, spec fields, and coupon stackability are the comparison inputs. Inconsistent SKU naming across channels — a persistent problem for international brands running separate cross-border and domestic listings — makes the same product look like three different products to a price-comparison agent, and usually loses all three.

3. Entity and alias resolution at the SKU level. We have written about the alias gap for brand names. At the execution layer it recurs one level down: if consumers say a nickname for a product line and your listings only carry the official model code, the agent's search inside each app fails before comparison begins.

4. Understanding what Doubao's ad system actually buys. ByteDance's standardized ad system is expected to open fully in Q4 2026 through Ocean Engine, with the primary format a brand recommendation card at the bottom of a conversation. That is inventory on the chat surface. It does not place a brand inside an execution query, and brand teams should not let a media plan imply otherwise.

5. Measuring the gap deliberately. Because the agent surface cannot be sampled remotely, the only honest method available today is manual: run a defined set of execution prompts on the device for your own category, record which channels the agent reached, which SKUs it surfaced, and where your brand dropped out. It is unglamorous, it does not scale, and it is currently the only primary evidence that exists.

Takeaway for brand marketers

Treat this as a second scoreboard, not a replacement for the first. Recommendation visibility still governs consideration — it is what decides whether a consumer arrives at the execution query already holding your brand name. But from September 2026, it is no longer the whole board.

Three concrete moves for Q4 planning. First, audit transactable coverage separately from content coverage, and stop reporting them in the same number; a brand can hold a 100 in one column and a zero in the other. Second, run a SKU-consistency pass across every Chinese channel before spending anything further on citation-building, because the agent compares listings and not narratives. Third, when a GEO vendor quotes you a guaranteed improvement in AI brand score, ask which surface the guarantee covers — the chat answer, or the purchase.

The gap between "China's AI recommends our brand" and "China's AI can buy our brand" opened on September 16. Most brand dashboards will not notice for another year.