China’s 618 AI shopping war
Doubao, Qianwen, JD AI Gou, and Xiaohongshu are competing for the AI commerce layer this 618. Hands-on testing reveals a two-horse race. The deciding factor is not model quality.
Four platforms went live with AI commerce features in Q1-Q2 2026, converging on China’s second-largest shopping festival
Hands-on testing reveals two product philosophies: conversational discovery versus chat-wrapped shelf search
The decisive moat is content-to-commerce loop data, not model capability. Only Doubao has it closed today
JD AI Gou recommended products with features they do not have. The accuracy problem nobody is pricing in
Xiaohongshu is the only credible challenger, if it can convert content depth into transaction gravity
China’s 618 shopping festival began in 2004 as a sales event for JD.com. It has since expanded into a month-long industry-wide promotion and is now China’s second-largest annual commerce event after Double 11.
This year, a second contest is running in parallel to the GMV race. Four major platforms have deployed AI shopping assistants in the months leading into 618, each betting that the next structural shift in Chinese e-commerce runs through conversational AI.
The timing reflects two converging forces. Foundation model capability crossed a practical threshold in late 2025. China’s Interim Measures for Generative AI Services, effective August 2023, provided enough regulatory clarity for platforms to commit to agentic transaction features at scale.
The four entrants arrived in rapid succession. Alibaba’s Qianwen integrated directly with Taobao on May 11, enabling users to browse and purchase via AI dialogue inside the Qianwen app.
ByteDance’s Doubao is expected to launch paid commerce features in late June, according to 36Kr, with Douyin Mall subsidy integration planned for Q3 and full operational rollout in Q4.
JD launched a standalone app called JD AI Gou in December 2025 and opened it to broader testing ahead of 618. Xiaohongshu, through its Dots AI division and its Wenyiwen in-search assistant, has embedded AI dialogue directly into its main app search layer, building a path from conversation to product discovery to in-app purchase.
Which platform architecture can make AI commerce work at scale is the useful question here. The decisive variable is which platforms have closed the loop between user-generated product content and transaction data. Only 2 platforms in this field qualify.
Two philosophies, not four competitors
Testing conducted by Jingzhe Research Institute ahead of 618 sent identical product queries to Doubao, Qianwen, and JD AI Gou. The results exposed a fundamental divergence in product philosophy.
Doubao and Qianwen behave as discovery agents. Both platforms probe user intent through dialogue, narrow options progressively, surface product tradeoffs explicitly, and close each response with a follow-up question designed to refine the recommendation. Doubao flags product weaknesses alongside strengths. Qianwen offers a summary comparison table. Both treat the chat interface as a reasoning environment, not a search box.
JD AI Gou operates differently. A query for earphones under RMB 500 returned 15 product recommendations across 5 categories. The categories were defined by functional attributes: noise cancellation, battery life, audio quality, sports use, and value. JD offered no follow-up question. No tradeoffs surfaced. The interface prompts users to browse and filter, which is structurally identical to JD’s existing shelf-based search experience with a chat input layer placed on top.
The design language reveals the strategic intent. Doubao and Qianwen are attempting to shift where purchasing decisions are made, from the browse-and-filter shelf to the AI conversation. JD AI Gou is attempting to route existing JD search behaviour through a new interface without disrupting the underlying shelf logic.
The distinction matters because it determines what each platform needs to win. Conversational discovery requires deep, accurate, preference-sensitive knowledge about products. Shelf search augmented by chat requires SKU breadth and fast retrieval. These are different technical and data problems. They reward different platform architectures.
Xiaohongshu, tested separately through its Wenyiwen feature, behaves closer to the Doubao-Qianwen pattern. It has a distinct architectural advantage examined below.
The accuracy problem nobody is pricing in
What JD AI Gou’s recommendations contain matters more than how many it returns. That is the significant finding from comparative testing.
In a follow-up query for earphones with strong audio quality, JD AI Gou returned a category labelled “bone conduction audio.” 2 of the top-ranked products in that category do not use bone conduction technology.
Both are Edifier earclip models. Jingzhe Research Institute verified this with JD merchant customer service. Both products use air conduction, not bone conduction. The AI attributed a feature to a product that the product does not have.
A separate category surfaced the same Edifier Comfo Clip Q model in 2 distinct recommendation buckets simultaneously. 2 different store listings for the same Moto Buds Clip model appeared across separate categories in the same response.
Both findings are structural outputs of a system that prioritises recommendation breadth over recommendation accuracy. When the query input is treated as a keyword trigger rather than a reasoning prompt, the output optimises for coverage, not correctness.
The commercial risk is concrete. China’s Law on the Protection of Consumer Rights and Interests and the 2023 Interim Measures for Generative AI Services both carry obligations around accuracy of information provided to consumers.
A platform whose AI agent misrepresents product specifications to drive purchase decisions faces exposure on both fronts. Whether regulators have appetite to test this in 2026 is an open question. That the exposure exists is not.
The deeper risk is trust.
Conversational commerce requires consumers to delegate a portion of their purchasing judgment to an AI agent. That delegation is contingent on accuracy.
One high-profile failure is sufficient to collapse consumer confidence in the agent format. A bone conduction earphone that uses air conduction. A waterproof jacket that is not. JD’s current architecture creates that risk at scale.
Why the loop matters more than the model
Foundation model capability will reach parity across all four platforms within 12 months. Alibaba’s Qwen 3 series and ByteDance’s internal model investments are both at frontier quality for Chinese-language commerce tasks.
JD licenses third-party model infrastructure. Xiaohongshu’s Dots unit, announced in April 2026, is a first-party AI development effort at significant scale. The model gap is a timing advantage, not a structural one.
The structural advantage belongs to whichever platform owns a closed content-to-commerce loop: the flywheel in which content generated by real users about real products feeds AI training data, which improves recommendation accuracy, which drives more transactions, which generates more content.
This loop requires 2 inputs that are hard to acquire independently: authentic product experience content at scale, and transaction data that links that content to actual purchase outcomes.
Only Doubao has both inputs fully connected today. By monthly active users, Douyin is China’s largest short-video platform, according to QuestMobile. Its content corpus includes hundreds of millions of product reviews, demonstrations, and purchase-experience videos.
TikTok Shop, operating under ByteDance infrastructure, provides the transaction linkage. When a Doubao user asks for a product recommendation, the underlying model can draw on content generated by users who bought the product and returned to post their own assessment. The loop is closed.
Xiaohongshu has the content input. Its UGC corpus is among the deepest product review databases in Chinese consumer goods, with particular strength in beauty, fashion, home, and outdoor categories. The transaction data to close the loop is not yet there.
Xiaohongshu’s in-app commerce function has grown. Transaction volume remains a fraction of Doubao’s Douyin Mall pipeline. Xiaohongshu is the only credible challenger, provided it can convert content depth into transaction gravity within the next 12-18 months.
