AI Splits Retail: Discovery to Agents, Ownership Stays with Retailers

Introduction: The Dual-Layer Structure of AI Retail

Intelligent-agent e-commerce is quietly dividing retail into two layers: discovery — what to buy, which is cheaper, whose recommendation to trust — handed to AI agents (Amazon Rufus, Walmart Sparky, Taobao AI Search, Shopify AI Recommendations); and ownership — checkout, payment, membership, after-sales, purchase history, pricing and inventory — kept firmly by the retailer. Adobe Analytics shows that in March 2026, AI-driven traffic to U.S. retail sites converted at a rate 42% higher than non-AI channels, whereas a year earlier it was 38% lower . Shopify reports AI-referred shoppers convert nearly 50% higher on product pages with a 14% higher average order value. The shift from -38% to +42% in one year occurred without a model-generation change; the key is that upstream recommendation quality and downstream product data both improved simultaneously.

Pain Point: Retailers' Dilemma — Want AI Traffic, Fear Losing Members and Data

NVIDIA's third annual survey shows 91% of retailers use or evaluate AI, and 90% plan to increase AI budgets in 2026 . China's AI smart-retail penetration reached 41.7% in 2025 , market size ~$48.7B, growth 38.2% leading globally. Yet retailers simultaneously crave AI's incremental traffic, higher conversion, and new acquisition channels while guarding their membership relationships, historical purchase/preference data, and recommendation/pricing data. The core tension: traffic belongs to others; members belong to us. An agent may bring a shopper via ChatGPT, but checkout, membership ownership, and repeat purchase all reside in the retailer's own systems.

Core Concept: Dual-Layer Structure — Discovery Layer vs. Ownership Layer

Discovery Layer (given to agents): natural-language search, cross-platform price comparison, personalized recommendation, content seeding, cross-platform shopping guides. AI helps users "find" but cannot see transactions, touch membership or history. Adobe: AI traffic conversion flipped from -38% to +42% in one year.

Ownership Layer (retailer keeps): checkout & payment, membership & accounts, after-sales & fulfillment, historical purchases & preferences, pricing & inventory, repeat-purchase operations — the "lifeblood" of retail, never outsourced.

Key players illustrate the split:

Walmart Sparky: 35% higher average order value, half of app users tried it; also integrates with ChatGPT and bets on open protocol UCP — opens discovery but Sparky and Marty serve Walmart's own retail media and checkout.

Amazon Rufus: 300M+ users by early 2026, users convert ~60% higher, assisted sales grew ~$12B in a year; yet technically blocks ChatGPT and Gemini from scraping product data — boost conversion, but don't feed my catalog to competitors' agents.

Domestic (Taobao/Tmall, JD): Taobao's 2025 Double-11 AI Universal Search solved ~50M consumer needs, generated ~2M AI shopping lists; JD's JoyStreamer digital human costs 1/10 of human hosts, serves 40K+ brands, GMV 700M+ — but transactions and membership stay on-platform.

The iron rule: open discovery layer to AI for traffic; guard ownership layer with machine-readable product data — but never hand over the keys.

Path: Testing Retail with the "AI+Industry Five Questions"

The series uses a consistent framework — Data Source, Decision Authority, Accountability, Measurement Unit, Boundaries — applied to retail:

① Data Source: Retailers natively possess structured product catalogs (title, category, price, stock, SKU, reviews, rich media). The challenge is fragmentation across ERP/OMS/membership systems and varying machine-readability. Alibaba Mama's "Wanxiang Lab" uses multimodal AI to turn main images, detail pages, reviews into searchable content, lifting product conversion 37% and user dwell 2.1x — essentially translating "human shelves" into "machine-readable" format.

② Decision Authority: Discovery and comparison delegated to agents; checkout/payment/membership stay with retailer. Visa survey: 69% of consumers willing to let agents handle evaluation, 62% checkout, 64% after-sales; yet checkout remains retailer's home turf. Amazon uses Rufus to lift conversion while blocking ChatGPT scraping — concrete embodiment of this line.

③ Accountability: If AI recommends wrongly (unsuitable product, leaks floor price), who is liable? Platforms typically require agents to "advise not buy"; refunds, counterfeits, false advertising remain merchant/platform responsibility. Tsinghua's Liang Zheng warns: prevent AI false advertising, prevent big-data price discrimination, protect consumer privacy — three regulatory red lines.

④ Measurement Unit: No unified financial yardstick, but multiple metrics: conversion rate (AI traffic +42% vs -38% year ago), average order value (Sparky +35%), repeat & membership (Ulta uses own data to link cross-year purchases & preferences), GMV (JD digital human 700M+).

⑤ Boundaries: Three lines. Discovery layer fully open; checkout/payment/membership/pricing are red lines. AI cannot execute transactions, cannot feed product data to competitor agents without thresholds. "Big-data price discrimination" and privacy are regulatory red lines (China 2025 Retail AI Graded Management Guidelines, EU AI Act retail provisions).

Benchmark Players: Quantitative Effects and Dual-Layer Actions

Amazon Rufus: 300M+ users, conversion +60%, assisted sales +$12B, heavy users 2.74x more likely to purchase. Action: open search/recommendation / guard product data. Source: Sensor Tower.

Walmart Sparky: AOV +35%, half of app users tried; integrated ChatGPT, bets on UCP. Action: open shopping dialogue / guard checkout & membership. Source: Digiday 2026-05.

Shopify: 2026 Q2 AI traffic orders up 3x YoY, AI recommendation conversion ~50% higher, AOV +14%. Action: open discovery/recommendation / guard store checkout. Source: Shopify earnings.

Taobao/Tmall Double-11: AI Universal Search solved ~50M needs, AI lists ~2M; customer-service refund time 8 min → 18 sec. Action: open shopping guide/service / guard transaction & membership. Source: The Paper / CNR 2025.

JD Digital Human: JoyStreamer cost 1/10, >80% of human host performance, serves 40K+ brands, GMV 700M+. Action: open live-stream explanation / guard transaction & after-sales. Source: JD 2025 Double-11.

Wumart Supermarket AI: AI assortment drove 5x+ traffic & sales; replenishment accuracy 95%+; clearance SKU listed in 5 minutes. Action: open assortment/replenishment / guard store membership. Source: People's Daily 2025-06.

Meitu Design Studio: Cross-border product image cost 1/5, efficiency +90%, conversion +25%. Action: open asset generation / guard brand store. Source: World Internet Conference 2025.

Signal: Same "use AI", but Amazon/Walmart open Rufus/Sparky as traffic portals while welding shut the product-data gate; Wumart/Meitu open assortment advice and asset generation while guarding stores/brands. Difference lies not in technology but in which layer they dare to release and which they must clutch .

Cold Reality: AI Recommends Once, Retailer Knows You for Years

1. "Shelf" Redefined — From Position to Readability

Old shelf: limited good positions, expensive bidding, traffic via ranking; AI couldn't crawl or understand. New shelf: structured product data retrievable by agents; real-time stock/price credibility = entry into candidate pool. Whoever makes data machine-readable enters the recommendation pool. Recommendation authority partially shifts to cross-platform agents. Condition: machine-readable product data — precisely why retailers release discovery but clamp ownership. Ulta Beauty: AI-referred customers convert double, yet insists checkout stays on its own site.

2. AI Recommends Once, Retailer Owns the Multi-Year Relationship

An agent may participate in only one purchase's discovery phase, but the retailer's membership system records cross-year purchases, returns, preferences, repeat buys. Accenture: 75% of global consumers trust AI agents more than best friends; 56% of loyal shoppers would switch brands on agent recommendation. Trust shifts to AI, but the party that owns the relationship remains the retailer. Traffic can be intercepted for a leg; relationship and data sedimentation stay forever in the retailer's layer.

3. Closer to Checkout, Closer to Regulatory Red Lines

Digital-human livestreaming and AI customer service cannot place orders or pay for users. "Big-data price discrimination" and privacy harvesting are explicitly banned (China Retail AI Graded Management Guidelines, EU AI Act retail rules). For SMEs, real opportunity lies not in "deploying digital humans" but in first making product data machine-readable and using AI to master discovery-layer tasks like assortment advice and customer service — exactly the path Wumart and Meitu demonstrate.

Counter-intuitive Hook: The most quotable finding isn't "digital human GMV 700M" but that AI referral conversion flipped from -38% to +42% in one year . Not because models got stronger, but because upstream recommendation quality and downstream product data improved together . The shelf's definition was rewritten — from "position" to "readability". When "can an agent read you" replaces "how good is your slot", retail's underlying competitive logic changes fundamentally.

Self-Checklist: What Retail AI Projects Must Ask First

One-sentence close: Don't ask "how many digital humans did we launch"; ask "is our product data machine-readable? Did we release discovery rights? Did we keep ownership rights?"

Is product data machine-readable? Are title, category, price, stock, SKU, reviews structured and retrievable by agents? If still relying on manual keyword stuffing or buying slot traffic, you haven't entered the "new shelf". Pass means: first break catalog into machine-readable structured fields.

Discovery released, ownership kept? AI shopping guides, digital humans, smart recommendations can open; but checkout, payment, membership, original product catalog must stay in-house. Any solution demanding "hand transaction execution to agent" is a red line, not a dividend.

Count hidden costs before counting visible traffic. Assortment advice, service response, asset generation — these discovery-layer steps are highest ROI for SMEs (Wumart replenishment 95%+ accuracy, Meitu cost 1/5). Don't start by overhauling the whole chain for flashy digital humans; first walk the machine-readable step.

Retail e-commerce is the third piece of the large-circulation module. Together with finance and logistics they form a "money, goods, people" triptych: Finance's AI value appears in financial statements; logistics' AI value appears in scheduling engines; retail's AI value appears in the dual-layer boundary line. The methodology transfers — any industry's AI landing must first answer "what to release, what to clutch". Taking this method to the next sector sharpens the answer — next episode leaves "selling goods" for "selling services, selling expertise": AI + Legal Compliance, where 150 pages of evidence face 5 pages of argument, and the true rewrite is the billing unit.