The Retailer's Own Model: JD's Jingyan Handled 80 Million Shoppers Last Quarter, and No GEO Dashboard Queried It Once

In the first quarter of 2026, nearly 80 million people used Jingyan (京言) — JD.com's in-app AI shopping guide — to help decide what to buy. That is up more than 200% year over year, according to figures JD disclosed at its September 9 JDD conference. For scale: DeepSeek's consumer app, the model every Western GEO vendor benchmarks against, reported roughly 130 million monthly active users in the same period.

Jingyan is not a general-purpose assistant. It does not crawl the open web. It runs on JD's in-house Yanxi (言犀) model, answers only from JD's own catalog, and lives inside a shopping app rather than a chatbot. Which is precisely why no brand visibility platform — ours included, until we started building for it — has ever sent it a query.

That gap matters more this month than it did last month. Between September 2 and September 15, every major Chinese commerce platform either launched or upgraded an AI shopping guide. The surface where Chinese consumers actually ask "which brand should I buy" is no longer six general models. It is six general models plus six retailer-native ones, and the second group is scoring your brand on completely different evidence.

What launched in two weeks

The pace here is the story. From reporting by Beijing News' Bewell Finance (贝壳财经) on September 15:

Alibaba's route is the closed loop: Qwen connected fully to Taobao on May 11, 2026, giving the assistant access to a 4-billion-SKU catalog and two decades of shopping behavior data. During Spring Festival, Qwen's "3 billion yuan free-order" campaign drove more than 120 million AI-completed orders in six days.

JD is running two legs at once. Externally, it feeds products, logistics and after-sales into Tencent's Yuanbao. Internally, Wang Yuanyuan, who heads innovative products at JD Retail, describes the goal as a two-way closed loop between GUI (the shelf) and CUI (the agent dialogue) — the conversation and the product grid enhancing each other rather than one replacing the other.

Why the retailer models score your brand differently

A general model answering "best coffee capsules under 3 yuan each" reasons over web text: reviews on SMZDM, Xiaohongshu notes, Zhihu threads, media roundups, encyclopedia entries. That is the evidence layer GEO work has been optimizing for.

A retailer-native model reasons over something else entirely:

• Catalog completeness — whether your SKUs are listed with structured attributes the model can filter on. Pinduoduo's AI search explicitly parses version, specification and price band from natural language. A product missing those fields is not badly ranked; it is unfilterable.
• Price history and current price position — the model can see your actual price, not a number scraped from a page months ago. This quietly fixes the price-hallucination problem that plagues general models, and it means an uncompetitive price now removes you from consideration silently, with no answer text explaining why.
• Merchant operational data — fulfillment speed, return rate, dispute rate, service scores. None of this is content you can write.
• First-party review corpus — the platform's own reviews, weighted by verified purchase.

The Beijing News reporters ran a clean test on September 14: the same query — coffee capsules averaging about 3 yuan — put to Doubao and Qwen on the same phone. Both parsed the constraint correctly. Both folded in budget, category and use case. But the price ranges they returned diverged, and the recommended brands barely overlapped at all. Doubao led with a brand at roughly 2.9 yuan per capsule, attaching price, features and a purchase link; Qwen returned one coffee pick plus a "better value" alternative in a different category.

That is two general-purpose assistants disagreeing. Add retailer-native models — each constrained to its own catalog, each weighting its own operational data — and the divergence widens further. A brand can plausibly be the top recommendation on Doubao, absent from Jingyan because three SKUs lack structured attributes, and deprioritized by Pinduoduo's AI search on price band alone.

The demand side is not waiting

Skepticism about AI shopping guides has been reasonable — conversion has lagged the hype. But the adoption curve is steepening. Industry data disclosed at CIFTIS 2026 put one platform's AI-related service orders at 9.816 million, up 157% year over year. JD's own AI shopping assistant crossed 150 million annual active users in 2025, with penetration above 20% of its base.

There is a second signal worth noting, because it is an inversion: AI is now a product sold on these platforms. Zhipu opened the first LLM flagship store on Tmall on September 2, selling nothing but model-call credits. Tmall launched an "AI Space Station" token top-up hub on September 3, consolidating Aliyun, Zhipu, Kimi, MiniMax and DeepSeek subscriptions onto one page; JD opened a parallel "AI Software Top-Up Hall." Within 48 hours of Zhipu's store opening, Taobao and Tmall AI subscription GMV rose more than 160% month over month.

When model vendors compete for shelf space on the same platforms whose models rank brands, the commerce layer has become the distribution layer for AI itself.

The unresolved conflict

Senior internet analyst Yang Huaiyu frames the shift as a reallocation of decision cost. Consumers used to do the filtering, comparing and trading-off themselves; the AI now absorbs that step, and the platform's role changes from displayer of goods to interpreter of demand.

He is careful about scope: shelves are not disappearing. Standardized commodities will keep following price-comparison logic. AI intervenes hardest in categories with complex requirements, many parameters and high decision cost — appliances, home, maternal and infant, beauty, electronics.

And the hard problem is structural, not technical. Shelf e-commerce has run for a decade on purchasable traffic; merchant ad spend is a core platform revenue line. If the AI simply hands over a conclusion, users stop browsing pages, and both merchant incentive to buy traffic and the platform's ad model take damage. Yang expects shelves and dialogue boxes to coexist for a long time, with platforms balancing two logics — and he argues recommendation neutrality may decide how far AI shopping ultimately goes.

For brands, that unresolved conflict is a window. Paid placement is not yet in these AI answers. The ranking is still earned.

Takeaway for Brand Marketers

Three concrete moves, in order of cost.

1. Treat platform product data as GEO infrastructure, not merchandising admin. For every SKU on JD, Taobao, Pinduoduo and Douyin Mall, audit whether structured attributes are complete: specification, version, size, material, compatibility, certification, usage scenario. Retailer-native models filter on these fields before any language reasoning happens. An incomplete listing is not a ranking penalty — it is an elimination. This is usually owned by an e-commerce ops team that has never heard the phrase "AI visibility," which is exactly why it goes unfixed.

2. Add retailer surfaces to your measurement set, and score them separately. A general-model score and a retailer-model score are not comparable numbers and should never be averaged. Run your top 20 category queries manually inside JD's Dongdong, Taobao's assistant, Pinduoduo's AI search and Doubao's Help You Choose, monthly. Record whether you appear, at what position, and against which competitors. Manual is fine — the point is that the number exists at all. Most brands have zero observations here.

3. Recognize which lever actually moves a retailer model. Content marketing does not. Fulfillment speed, return rate, dispute resolution, review volume on verified purchases and price competitiveness do. If your China GEO plan is entirely a content plan, it addresses roughly half the surfaces where Chinese consumers now ask which brand to buy — and the half that is growing more slowly.

The brands that will look inexplicably dominant in Chinese AI search in 2027 are not the ones publishing the most. They are the ones whose operational data reads well to a model that has never visited their website.

Related: see current brand scores across China's six major AI models on the GEO Hub brand index, and our earlier analysis of operational signals in China's AI commerce loops.