Doubao and Qwen Answered the Same Shopping Query. The Brand Lists Barely Overlapped.

On September 14, 2026, a reporter at Beike Finance ran a small experiment that should worry anyone managing brand visibility in China. She typed the same request into two of China's largest AI assistants — coffee capsules averaging about 3 yuan each — and compared what came back.

Doubao returned four brands, sorted by flavor and price, with a buying summary and product links. Its top pick was priced at roughly 2.9 yuan per capsule. Qwen returned two items: one coffee product and one alternative it judged better value, both with links.

The recommended brands had almost no overlap. The price bands the two models worked with didn't match either.

This is not a ranking gap. Two models disagreeing on ordering is normal and measurable. This is a catalog gap — the two assistants were reading from different product universes entirely, and neither one was reading the open Chinese web. For international brands tracking a single "China AI visibility score," that distinction is the difference between a number that means something and a number that means nothing.

September 2026: every major platform shipped a shopping agent

The past three weeks compressed what looked like a two-year roadmap into a single news cycle.

On September 9, at its JDD conference, JD.com launched App version 16.0 with Dongdong, a native shopping AI agent supporting text and voice product search, video-based lookalike search, 3D in-room product placement, and integrated savings and after-sales assistants. Around the same time at CIFTIS, Alibaba shipped five AI applications, including a Taobao shopping assistant built on Qwen covering virtual try-on, discovery-to-purchase, and coupon aggregation.

They were not alone. Taobao/Tmall, Douyin, JD, Pinduoduo, Meituan and Kuaishou have all now launched or refreshed AI shopping capabilities. Pinduoduo began greyscale-testing AI search in March 2026 and launched in May. Meituan upgraded its "Xiaotuan" assistant to 2.0 in July, adding ordering, ride-hailing and reservations. Kuaishou wired add-to-cart directly into the chat window, removing the product page from the flow.

The usage numbers are no longer trivial. JD disclosed that nearly 80 million users relied on its Jingyan AI shopping guide in Q1 2026, up more than 200% year over year. Alibaba reported that its Spring Festival Qwen giveaway campaign drove over 120 million AI-completed orders in six days. Industry data released at CIFTIS 2026 showed one platform logging 9.816 million AI-related service orders, up 157% year over year.

Qwen's integration with Taobao, completed May 11, 2026, gives it access to a catalog of roughly 4 billion SKUs plus two decades of shopping behavior data — supporting cross-condition filtering, fuzzy intent inference, and end-to-end ordering, payment and after-sales.

That 4 billion SKU figure is the whole story. Qwen answering a shopping question is not Qwen answering a general knowledge question. It is a retrieval operation over Taobao's product graph.

Two stacks, one brand name

The measurement problem is structural, and it splits Chinese AI brand visibility into two surfaces that most brand teams currently treat as one.

A brand can score 100 on the left column and be structurally absent from the right. If you have no JD flagship store, Dongdong cannot recommend you — not because your content is weak, but because you are not in the index it queries. If your Tmall listings have thin attribute data, Qwen's cross-condition filtering ("averaging about 3 yuan each") cannot match you, because the constraint the user expressed maps to a structured field you left empty.

This reframes what a GEO input even is. On the general-model surface, the input is content that models read. On the commerce-agent surface, the input is data your merchandising team enters into a platform back office — title structure, specification completeness, unit pricing, category assignment, SKU-level variant data. No amount of PR moves that number.

Why Doubao and Qwen disagreed

ByteDance and Alibaba built toward the same destination from opposite directions, and the divergence in the coffee-capsule test falls out of that.

Alibaba's approach is a closed loop: the user opens Qwen, Qwen reaches into Taobao's catalog, and the order, payment and after-sales all stay inside Alibaba. Answer quality is bounded by Taobao's product data.

ByteDance's approach runs through content and traffic. Doubao's AI commerce function — internally nicknamed "Doubao helps you choose" — surfaces product cards in conversation and links out to Douyin Mall, covering electronics, home, maternity and beauty. Douyin is separately testing "AI Workshop," which turns AI applications themselves into feed content delivered as interest cards.

JD split the difference: externally partnering with Tencent Yuanbao (Yuanbao interprets the need, JD supplies the goods, logistics and service), internally fusing the interaction layer. JD Retail's innovation product lead described the goal as a two-way loop between GUI — the shelf — and CUI — agent dialogue.

Three different retrieval graphs, three different merchant populations, three different data schemas. Asked the same question, they were always going to return different brands. What's notable is how little they overlapped — suggesting the divergence is driven less by model judgment than by what each one could see.

The neutrality question nobody has answered

Senior internet analyst Yang Huaiyu framed the shift precisely: AI changes the allocation of decision cost. Filtering, comparing and trading off used to be the user's work; the agent now absorbs it, and the platform's role changes from displayer of goods to interpreter of demand.

His warning matters more for brands than the launches do. The hard part of AI shopping, he argues, is not technology but the incentive structure. A decade of shelf e-commerce rests on the premise that traffic can be bought — merchant ad spend is a core platform revenue line. If the agent simply hands over a conclusion and users stop browsing pages, both the merchant's incentive to buy traffic and the platform's ad model take a hit. Shelves and chat windows will coexist for a long time, and, in his view, recommendation neutrality may determine how far AI shopping ultimately goes.

Read that as a forecast about your costs. Every one of these agents sits on top of an advertising business. The current organic window inside platform commerce agents is a transition artifact, not a settled state — the same pattern we flagged when Doubao's ad system was slated for Q4 2026.

Note also which categories are genuinely in play. Yang expects standardized commodities to keep following price-comparison logic on the shelf, with AI intervening most in categories with complex requirements, many parameters, and high decision cost. If you sell a simple, price-transparent product, agent visibility is a smaller lever than it looks. If you sell something people research — appliances, skincare regimens, gear, anything with a spec sheet — it is a large one.

A footnote worth tracking: models on the shelf

One more September signal points the same direction. On September 2, Zhipu opened the first LLM flagship store on Tmall, selling exactly one category of product: model API quota. On September 3, Tmall launched an "AI Space Station" token recharge center collecting subscription products from Alibaba Cloud, Zhipu, Kimi, MiniMax and DeepSeek onto a single page. Within 48 hours of Zhipu's store opening, AI token subscription transactions across Taobao and Tmall grew over 160%. JD launched a comparable "AI Software Recharge Hall" at nearly the same moment.

The models are now SKUs in the same catalogs their agents query. The boundary between "AI platform" and "retail platform" has stopped being a useful distinction in China, and any GEO model that assumes the two are separate layers is already describing an older market.

Takeaway for Brand Marketers

Four concrete moves:

1. Split your visibility measurement in two. Track general-model scores (Doubao, DeepSeek, Kimi, Qwen, Wenxin, Hunyuan) and platform-agent presence separately. They are driven by different inputs, and averaging them produces a number that hides the failure. Run the same commerce-intent prompt across both surfaces monthly and record the overlap rate — low overlap means your content strategy is not reaching your commerce surfaces.

2. Audit catalog data as a GEO asset. Pull your top 20 SKUs on each platform you sell on and check attribute-field completeness, unit pricing, spec structure and category assignment. Constraints like "averaging 3 yuan each" only resolve against structured fields. Empty fields are invisibility, and this work belongs to merchandising, not to content.

3. Map your platform coverage against agent coverage. List which of Taobao/Tmall, JD, Douyin Mall, Pinduoduo, Meituan and Kuaishou you have an active store on. Each absence is an agent that structurally cannot recommend you, regardless of your brand score.

4. Budget for the window closing. Organic placement inside platform agents is not durable. Treat 2026–2027 as the period to build catalog data quality — the input that keeps working after paid placement arrives — rather than chasing agent rankings that a future ad auction will reprice.

The single most useful diagnostic is also the cheapest: ask two different Chinese AI surfaces the same buying question about your category, and count how many brands appear on both lists. If the answer is close to zero, you do not have one visibility problem. You have several.

Related: see how your brand scores across China's six major AI models on our brand rankings page, and our earlier analysis of China's AI shopping guide conversion gap and operational signals in AI commerce loops.

Sources: Beike Finance / The Beijing News (September 15, 2026) head-to-head Doubao vs. Qwen shopping test and platform survey; JD JDD Conference disclosures (September 9, 2026); Alibaba CIFTIS 2026 announcements; Tmall AI Space Station disclosures (September 2–3, 2026).