What has been added to the app
JD.com said that following the release of version 16.0 of its app on September 9, 2026, the Dongdong AI assistant became available to shoppers at different stages of a purchase. According to the company's description, the assistant interprets broad or vague requests and connects user needs with products and services within the JD.com ecosystem. [1 · JD.com]
How video context is used
Users can communicate with Dongdong through voice and video. For example, when choosing furniture or appliances, they can show the available space, the surroundings or the problem they want to solve. The assistant uses this context to refine recommendations. These capabilities are described in the corporate JDDiscovery overview; it provides no independent metrics on recommendation accuracy or sales impact. [1 · JD.com]
Editorial analysis
This interface changes the starting point of search: a shopper can begin with a household task without even knowing the name of the product they need. For commerce, this could shorten the path to a suitable assortment. At the same time, the quality of the result depends on how accurately the system accounts for the visible space, product specifications and the constraints of the request itself. [1 · JD.com]
Sources
- JD.com — highlights from JDDiscovery 2026 — JD.com blog, September 9, 2026. The company itself describes the capabilities.
- Dietvorst, Simmons and Massey (2015) — algorithm aversion after errors — Original forecasting experiments; the mechanism is applied to the shopping assistant as a hypothesis.
- Dietvorst, Simmons and Massey (2018) — the ability to adjust an algorithm — Original experimental research; not an evaluation of Dongdong.
Expert commentary
I see Dongdong as an attempt to shift competition from displaying products to understanding the shopper's task. A request for a suitable washing machine includes constraints that a person may not know how to express through filter names. If the assistant identifies them correctly, it will reduce the cost of finding a solution. But usefulness should be judged by the suitability of the delivered product, not by the persuasiveness of the dialogue or the number of options suggested in the conversation. [1 · JD.com]
For manufacturers and sellers, this increases the potential value of accurate specifications, dimensions and compatibility information. In my scenario, a well-described product gains an advantage because the assistant can more easily match it to a particular space and task. But the effect depends on JD.com's recommendation rules. They cannot be assumed in advance to be neutral or purely advertising-driven: the company needs to disclose the basis for recommendations to make that assessment testable. [1 · JD.com]
The psychological risk is explained by Dietvorst, Simmons and Massey's research on algorithm aversion after observing errors. In forecasting experiments, participants could reject an algorithm even when it outperformed a human on average. For Dongdong, this suggests a hypothesis: one conspicuous error involving dimensions could cause disproportionate damage to trust. The transfer is not automatic—forecasting and shopping differ—so shoppers' reactions need to be measured separately. [2 · Dietvorst, Simmons and Massey (2015)]
Another study by the same researchers shows that being able to make small adjustments to an output can increase willingness to use an imperfect algorithm. My recommendation is therefore to let people explicitly set their budget, dimensions and exclusions, revise recommendations and return to conventional search. User control could become part of the quality of the relationship here. This is a design principle derived from experiments, not a claim that everything listed has already been implemented in Dongdong. [3 · Dietvorst, Simmons and Massey (2018)]
For society, voice and video can lower the barrier for people who find it difficult to formulate technical requests. At the same time, filming a home environment requires a considered choice about what information to share. I would regard a clear explanation of processing purposes and the ability to get help without unnecessary data as signs of maturity. A broader international effect will depend on language accessibility and the quality of local assortments; one Chinese launch does not yet establish the model's universality. [1 · JD.com]
Success should be tested through an experiment with comparable groups of shoppers. Useful metrics include time to a suitable choice, returns due to incompatibility, post-purchase inquiries and repeat use after several orders. Higher conversion accompanied by more mistaken purchases would be a troubling result. The optimistic scenario is that the assistant improves decisions and strengthens trust; the negative one is that it moves people to payment faster while leaving customers with the hidden cost of correcting their choices. [1 · JD.com] [2 · Dietvorst, Simmons and Massey (2015)]