Differences in data structure, algorithmic goals, and organizational methods.

$PDD(PDD.US) $Alibaba(BABA.US) $JD.com(JD.US) 
 

It all started a few days ago when I bought a keyboard.

Out of habit for research, I searched for the same Logitech keyboard on Pinduoduo, Taobao, and JD.com apps and compared the prices. Then something strange happened: during subsequent use, the recommendations from the three apps showed different results.

Pinduoduo: Showed many related products, including keyboard wrist rests, cable organizers, desktop speakers, computer stands, desk mats, charging stations, and heaters.

Taobao: Showed keyboards and related products, including keyboards from other brands, keyboard wrist rests, and keyboard cleaners.

JD.com: Showed more keyboards—other Logitech series, JD’s own brand, Huawei, Lenovo, Acer, Stalker… with relatively similar types and prices.

The most obvious difference is that Pinduoduo understands the "scenario" of computer work better and recommends related products accordingly; JD.com is like a straight guy—after seeing a keyboard, it just keeps pushing more keyboards; Taobao is somewhere in between, but the recommended keyboards vary greatly in price and type, and the related products aren’t as diverse.

The result was that I ended up buying a bunch of small items on Pinduoduo, and my intuitive feeling was:

Pinduoduo really knows what I need.

Why does this difference occur?

If calling an AI model could analyze user needs, clearly the recommendations wouldn’t differ fundamentally. The current differences indicate that this isn’t just a technical issue.

After discussing with friends, I attribute this difference in recommendation accuracy to fundamental differences in the three companies’ data structures, algorithmic goals, and organizational methods.

1. Data structure is the path to understanding user needs

First, let’s understand a concept: data tracking ("埋点").

Data tracking refers to embedding statistical code in an app’s code ("埋点") to track user behavior (e.g., button clicks, page dwell time, feature usage) and app performance (e.g., crashes, loading failures). The core purpose is to collect data to support product optimization and operational decisions. Simply put, it’s like installing "invisible sensors" in the app—every user action is recorded. This is the necessary path to understanding user needs.

The longer the user behavior chain, the more business iterations, and the higher the activity density, the more complex the tracking becomes. The resulting problem is excessive data volume, which introduces noise in understanding user needs. This is the root cause of the differences in how the three e-commerce platforms understand user needs.

Taobao has one of the most complex UIs in China, with the highest volume of tracking points, but this also makes recommendations extremely difficult.

First, there are too many business lines: Taobao main site, Tmall, flash sales, Taobao Live, content ("逛逛"), maternal and child products, apparel, international, second-hand, event pages (changing daily). Each line has its own logic and requires independent tracking.

Second, there are too many events: Double 11, Double 12, 38 Shopping Festival, 618, Content Festival, Appliance Festival, merchant promotions… The volume of page changes for Taobao’s annual activities may exceed the total of the other two platforms combined. Each iteration requires adjustments to tracking.

Third, the first two business attributes create the problem of blurring user intent. Users might come in: just browsing, watching live streams, checking outfits, opening and closing the app… These "low-intent behaviors" make intent recognition difficult.

In a 2019 interview, Jack Ma once said: Every night, 17 million people browse Taobao without buying anything. No one knows what they’re doing—they’re just casually browsing.

If you don’t even know the intent behind each user click, how can you make accurate recommendations?

JD.com has another problem: too little data, and user intent is too clear.

JD.com is a strong-intent shopping scenario. Before opening the app, users often already know what they want to buy: an iPhone, an air conditioner, diapers, dog food. The user journey is simple: search, browse, buy—done. The volume of tracking points is also smaller. Although JD.com has many business lines and events, user intent is clear. The problem is insufficient information, leaving little room for algorithms to work.

As a result, JD.com’s recommendations tend to be more essential, practical, and budget-friendly, but the coverage is narrow. When searching for a keyboard, it won’t recommend wrist rests or desk mats…

Pinduoduo has fewer tracking points, but its "purchase-driven recommendations" are more accurate.

Pinduoduo has fewer business lines and even fewer events. There’s no special Double 11, 618, or 38 Shopping Festival—every day is the lowest price. This makes Pinduoduo an ideal scenario for recommendation systems:

Strong conversion focus → Low content noise → Low UI noise → High-frequency, low-price purchasing behavior

Why does Pinduoduo have fewer tracking points but still deliver strong recommendations? Because Pinduoduo captures: what you bought, how much you bought, your price range, your purchase frequency, your price sensitivity, your response to promotions (bargaining, flash sales)… These actions more finely depict user needs.

Pinduoduo’s data is like a diary, with each page recording your complete purchase journey: from discovering discounts, to hesitating, to bargaining, group buying, giving up, returning… Each step is tightly connected. It’s not as noisy as Taobao or as thin as JD.com—it "sees you" completely. This is why Pinduoduo’s recommendations always hit the mark.

2. Algorithmic goals determine business direction

Pinduoduo’s recommendation goal is very simple: orders, orders, and more orders.

Pinduoduo’s algorithmic KPI is straightforward: drive higher purchase rates tonight, increase repeat purchases within N days, enhance users’ perception of savings per order, and push the cheapest, most attractive items to you.

It doesn’t pursue interests, content, or retention time—it only cares whether recommendations lead to transactions. This makes Pinduoduo’s recommendations "pure" and "efficient."

Taobao, with its many business lines and events, has more recommendation goals.

Taobao’s algorithm isn’t just about "accurate recommendations." It also considers: retaining users, growing users, stimulating interest, increasing exposure, balancing the merchant ecosystem, and ad revenue.

Multi-goal recommendations easily lead to: complex logic, conflicting strategies across departments, and impure accuracy (content interest ≠ purchase interest), inevitably reducing recommendation precision.

JD.com’s recommendation goal is: improving conversion efficiency in strong-intent purchases.

JD.com’s algorithm focuses on: whether you place an order, your preferences among price, logistics, and brand, and your purchase frequency. Recommendations are more about "assisting transactions" than "stimulating demand," so the ceiling for recommendations isn’t high.

3. Organizational methods are the root cause

Pinduoduo has only one department in the entire company: orders.

Pinduoduo’s organizational method is extreme: flat, small teams, fast iterations. Models built this week are evaluated for order conversion next week—this is a typical "combat squad structure." All algorithm, product, and operations teams are unified under one goal: driving purchases, enhancing the perception of affordability, and improving traffic efficiency. All strategies revolve around instant conversion.

The result is that while other platforms are still coordinating, Pinduoduo has already updated the same algorithm path 10 times.

Taobao, from content teams to live streaming, search, merchant operations, and recommendations, has different goals across teams.

Different goals create conflicts, and coordinating across departments is costly. Additionally, data from different departments is siloed in their own systems, making flow less smooth. Algorithm teams may not even have access to all data. However, Taobao has seen improvements in organizational coordination this year.

JD.com is a large-middle-platform organization, centered on "authentic products + self-operated + service."

JD.com’s organizational structure is stable, emphasizing: self-operated supply chains, after-sales experience, repurchase, search quality, and experience orientation. This complements JD.com’s business model, with the core being "authentic products + self-operated + service." About half of its products are self-operated, which means innovation moves much slower.

When we say Pinduoduo’s recommendations are more accurate, it essentially reflects:

Pinduoduo’s business model and its recommendation system were designed as one.

To summarize the differences in the three platforms’ recommendation systems in one sentence:

JD.com: I know what you want to buy now.

Taobao: I know what interests you (but you might not buy it).

Pinduoduo: I know what else you’ll buy.