JD.com
JD.com runs CogSearch, a multi-agent AI search built on small Qwen3 models
JD.com's researchers describe CogSearch, an AI search mode on JD.com that helps shoppers decide on complex purchases, built on open-weight Qwen3 models. In a two-week test on 10% of live traffic, JD reports a 5% cut in decision cost and a 0.41% rise in conversion.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
5%
Reduction in decision cost (average searches plus clicks per transaction) in a two-week online A/B test on 10% of live traffic; the paper does not say whether this is relative or absolute1
0.41%
Average increase in user conversion rate (completed orders per unique visitor) in the same A/B test; the paper does not say whether this is relative or absolute1
over 30%
Increase in user conversion rate for decision-heavy intents such as product interpretation, gift selection and functional comparison, same test; relative or absolute not stated1
What it took
Cost, team size and timeline were not disclosed in the sources.
The problem
Online store search is built to retrieve and rank products. Shoppers with a harder question, such as which gift to pick or how two products differ, end up running many searches and clicks. JD.com's team wanted search to help shoppers decide, not only find.
What they built
CogSearch is a group of cooperating AI agents inside JD.com's search. A Planner breaks the shopper's need into tasks, an Executor gathers product and web information, a Guider asks follow-up questions to narrow the choice, and a Decider produces a structured purchase recommendation. Shoppers reach it through an opt-in "AI Search" entry.
How they did it
The Planner and Guider run on Qwen3-4B and the Decider on Qwen3-8B. Both are open-weight models, meaning their files are published so a company can adapt and run them. JD trained these small models by distillation, which means teaching them to imitate a larger model, here DeepSeek-R1. To meet speed limits it also used 8-bit compression of the models, caching and parallel scheduling. The paper does not report cost or team size.
Results
In a two-week online A/B test on 10% of live traffic, shoppers with access to CogSearch needed 5% fewer searches and clicks per transaction (decision cost). Their overall conversion rate rose on average by 0.41%. For decision-heavy intents, conversion rose by over 30%. The paper does not say whether these are relative or absolute changes. Average latency was 1.3s, against 800ms for traditional search. The paper says CogSearch has since been fully deployed.
Lessons for your project
Aim an AI mode at the queries where shoppers struggle. The reported lift was concentrated in decision-heavy intents, not across all traffic.
Measure effort as well as sales. JD tracked decision cost, the searches and clicks needed per purchase, next to conversion.
Use small models where speed matters. JD picked 4B and 8B-parameter models to meet latency limits, and still ended up slower than ordinary search.
Be careful reading opt-in tests. Users chose to use the new mode, which can skew the comparison.
Caveats
- The figures come from a paper by JD.com's own staff and are not independently confirmed.
- The paper does not say whether the 5%, 0.41% and 'over 30%' changes are relative or absolute. It also gives no confidence intervals or test dates.
- Access was through an opt-in 'AI Search' entry, so users who tried it may differ from those who did not. The paper does not discuss this.
- CogSearch is slower than ordinary search: 1.3s against 800ms on average, softened with streamed output.
- The paper gives no deployment date. It says CogSearch 'has since been fully deployed'. The deployed date shown is the paper's submission date of 12 March 2026 (the announcement date); the actual go-live is earlier or unknown.
- Training cost, team size and rollout scale beyond the 10% test are not stated.
Sources
- CogSearch: A Cognitive-Aligned Multi-Agent Framework for Proactive Decision Support in E-Commerce Search
arXiv (authors from JD.com) · 2026-03-12 · Research paper · arxiv.org
Summarized in our own words from the sources above. Spotted an error? See corrections.