AI Search Hinders E-commerce Efficiency | Generated by AI
Question: Is JD.com’s AI search feature (DeepSeek-powered) in their mobile app actually worse for users than traditional keyword search? Would users be better off using a dedicated chat AI for research, then switching to JD.com with conventional keyword queries?
Answer:
Yes, and this is a well-understood product design failure pattern. Here’s why JD.com’s AI search is suboptimal and why your instinct is correct:
Why AI Search in E-commerce Apps Fails
Misaligned objective functions. JD.com’s AI search isn’t optimized for your best answer — it’s optimized for GMV (gross merchandise value). The DeepSeek model is fine-tuned or prompted to surface products that convert, not to give you neutral, accurate product research. Every AI answer is a funnel, not an oracle.
Context window ≠ product catalog intelligence. A general LLM like DeepSeek doesn’t have deep structured knowledge of SKU-level specs, real user reviews at scale, or cross-category comparison reasoning baked in. It’s essentially doing RAG over JD’s product data, which is noisy and merchant-biased.
Hallucination risk on specs. If you ask “which graphics card has the best price-performance under 3000 CNY,” the AI might confidently give you outdated benchmark comparisons or wrong TDP numbers. Traditional search returns the actual product pages where you verify specs yourself.
The UI/UX forcing function. AI search answers are a dead end — you read the answer, maybe click one recommended product. Traditional keyword search gives you a results grid where your eye can scan 20 products, sort by price, filter by brand, check ratings count. Information density is higher with keyword search for shopping tasks.
The Right Mental Model: Two-Phase Research
Your described workflow is actually optimal:
Phase 1: Research (ChatGPT / Claude / Perplexity)
→ "Compare RTX 4070 vs RX 7800 XT for ML inference workloads"
→ "Best mechanical keyboard switches for programming under 500 CNY"
→ Get structured, unbiased analysis
Phase 2: Execute (JD.com with keywords)
→ Search "RTX 4070 Ti ASUS" — already know exactly what you want
→ Sort by price + filter by JD self-operated (京东自营)
→ Check ratings count and recent reviews manually
This separation works because research is a generative task (LLMs are good at this) and shopping execution is a retrieval + filtering task (structured search is good at this).
Why JD Won’t Fix This
JD’s incentive is to keep you inside the app and reduce comparison shopping. AI search that gives you a confident answer discourages you from checking Taobao or PDD. The AI feature is partly a retention mechanism, not a pure user value play.
DeepSeek integration is also likely a cost signal — cheap inference, fast deployment, and a marketing story (“AI-powered”). Product quality for the user is secondary to the narrative.
The General Principle
AI search is good when: the query is ambiguous and needs interpretation, the corpus is well-structured and curated (e.g., enterprise knowledge bases), and there’s no conflict of interest between the platform and the user.
AI search is bad when: the platform has a financial stake in what you find, the underlying data is merchant-generated (biased), and you need to compare across many options rather than describe one.
JD.com’s case hits all three “bad” conditions. Your two-app workflow is the correct response to a poorly aligned product.