JD's Agentic Advertising Shift: From Probability Prediction to Logical Reasoning with AI Agents

Core Thesis: Decision-Makers Have Changed

The advertising chain's decision-makers are shifting from humans to a "human + AI Agent" hybrid model across three links: advertisers (delegating placement decisions to agents), consumers (moving from list browsing to conversational need expression), and platforms (requiring specialized decision-making that understands business goals, calls real tools, and responds in tens of milliseconds). A production-ready advertising agent must simultaneously: understand goals, execute real actions, maintain state, evolve via feedback, and be governably trusted.

New Standard: From Next Token to Next Business Outcome

Traditional advertising algorithms optimize for next-token prediction (click-through, conversion probability). The new standard demands alignment with the next business outcome — explaining why to increase budget or pull back, and how to roll back if wrong.

Delivery Brain: From Probability Prediction to Logical Reasoning

JD's "Delivery Brain" advertising decision reasoning model comprises three layers:

Base layer : fuses platform ad data, merchant placement behavior/effects, and external industry trends.

Model layer : general LLM base infused with vertical advertising knowledge, learning placement decision understanding.

Capability layer : outputs pre-placement strategy suggestions, intelligent diagnosis, and full autonomous托管.

The reasoning loop operates in three steps:

Dynamic perception : real-time view of traffic quality, competitive landscape, goal progress — woven into a dynamic competition graph.

Reasoning decision : identifies traffic gaps, yields personalized executable recommendations (budget, bid), drives advertiser adoption.

Feedback loop : decision trajectories, reasoning chains, reflection, and attribution flow back into training; the model grows more domain-aware over time.

Online metrics: advertiser adoption rate +15.4% , advertiser ROI +8.7% .

B2B Practice: Jing Xiaotong Advertising Agent

Merchants define WHAT (business goals), not HOW . Example goal: "This new product completes cold start in 30 days, guards ROI, finds scaling room." The agent runtime then:

Builds a business world model (product, store, audience, account, industry context).

Planner decomposes the goal — generates hypotheses, schedules dynamic tasks, sets success/exit conditions.

Layers strategy, audience, creative, placement, analysis actions; each has tools, skills, workflows plugged into ad APIs.

Online decision combines rules, LLM, prediction models, Bandit, causal experiments, and Memory — not pure LLM.

Before vs. After comparison:

Before : dozens of daily manual checks across product lines, 5+ steps to download creative data, Excel pivot tables for daily reports.

After : "Check all plan consumption" — 5 seconds; "Check 7-day creative" — 5 seconds, near 100% accuracy; "Generate yesterday's report" — 2 seconds.

Case study — Shan Ze brand : During a major promotion, full托管 captured the 2–5 AM manual blind spot. Night placement conversion efficiency +30% . Daytime: competitor budget shifts, traffic cost changes, search term spikes — windows of minutes. Agent monitors 7×24, converts fragmented opportunities into growth.

Safety guardrails (7 points) : permission isolation, risk grading, execution simulation, key approval, anomaly circuit breaker, full audit, one-click rollback. Advertisers hand over the goal but keep the red lines on the steering wheel.

C2C Agentic Advertising: Conversational Shopping Agents

~70% of user queries are vague ("good and cheap", "gift for elder") not specific SKUs. Traditional keyword ads fail here — poor match, high bounce, low conversion. Solution: multi-turn conversational shopping agent that translates vague needs into structured product profiles. Commercialization must sit under experience constraints — if chat feels broken, any monetization damages trust.

JD APP now runs hybrid dual-drive : traditional search/recommendation model + AI model.

Think Then Recommend (TTR) — WWW 2026

Traditional conversational recommendation suffers two flaws: only looks at current short query, ignores long-term purchase history; black-box — user doesn't know why, advertiser can't adjust.

TTR splits into three agents:

Basic intent extraction : places multi-turn context and current query together to judge immediate need.

Fine-grained intent generation : completes with historical behavior, falls back to popularity for cold start, grows "I want shoes" into structured intent with brand, category, price tier.

Preference alignment recommendation : vector recall → semantic alignment → rerank → returns results with reasons.

RIGER — SIGIR 2026

Injects reasoned intent into generative ad recall. Maps concepts like "Spring Festival, gift, specialty" to Intent SID; online fuses intent SID with behavior SID — explores latent demand while preserving existing preference.

Results: user understanding capability +227% ; online click +1.6% , consumption +1.3% , strict cold-start new items +4.08% ; serving 1.9 billion users . For advertisers: previously invisible demands become recallable; cold-start items gain more exposure.

Agentic Underlying Architecture & Compute

Open System Architecture: Professionalism & Openness

Not just another API set — a complete intelligent agent runtime enabling external-agent ↔ platform-agent collaboration (A2A). Three-layer open design:

Client access layer : A2A gateway — auth, rate limiting, routing; external agents plug into ad platform.

Dedicated runtime : KA (key account) customers get isolated runtime with version management and rollback for business stability.

Capability open layer : core ad capabilities encapsulated as Skills and A2A toolsets with strict authorization boundaries.

Compute Infrastructure: General + Intelligent Super-Node System

Agents need persistent memory and fast multi-turn reasoning — core pain point is ultra-low latency, KVCache, massive cache. JD introduced a general + intelligent compute super-node system achieving:

Cross-node communication latency reduced 90%

Cache pool expanded to 300 TB+

Supports thousand-node, hundred-TB KVCache shared cache

Three core technologies:

URMA : cross-node high-speed transport, solves large KVCache object cross-machine transfer bottleneck.

OS heterogeneous pooled memory management : connects HBM, local memory, SSD multi-level cache.

Super-node communication : million-level RPC, zero-serialization copy.

In plain terms: reaction fast enough, brain large enough, throughput strong enough. Without this compute base, large-scale ad agents remain demos.

Full-Stack Optimization: Tens of Milliseconds End-to-End

Unlike chatbots, ad decisions happen in milliseconds — a few tens of milliseconds late and the bidding window closes. The agent must do LLM inference, vector retrieval, and heavy business logic. Performance comes from algorithm–compiler–OS–chip full-stack co-tuning . JD collaborated with domestic and international teams: from chip and communication up through OS customization, AI compiler customization, down to retrieval algorithms optimized at 0.5 ms granularity. Result: top-layer ad agent decision speed reaches tens of milliseconds .

Outlook

Agentic advertising is not simply LLM + ads. It is a complete new paradigm spanning B2B merchant empowerment, C2C consumer understanding, and the compute capability that underpins both. JD expects to advance this paradigm together with merchants.