Search vs Recommendation vs Advertising: Concepts, Differences, and System Architectures

Hello everyone, I am Kaiyuan. Although we have previously shared many articles about search and recommendation advertising, we lacked a systematic framework to connect them, so this series uses the business flow as the core to organize the topics.

Traffic and Monetization are the two key factors for internet companies: first they invest to acquire users (traffic), then they seek ways to monetize. Search, recommendation, and advertising are standard services that generate most of the traffic and revenue.

Data : massive and noisy user logs; extracting valuable signals is challenging.

Algorithm : balancing innovation and deployment; online latency constraints require trade‑offs between effectiveness and efficiency.

Architecture : classic three‑stage pipeline (recall → coarse ranking → fine ranking) limited by performance.

Search VS Recommendation VS Advertising

Although these three share many similarities, their details differ enough to require separate teams.

Differences in problem definition

Search : users actively input a query, expressing a clear intent; the engine retrieves the top‑K most relevant results.

Recommendation : no explicit query; the system infers user interests from profiles and behavior to proactively present likely consumable items.

Advertising : involves three parties—advertiser, platform, and user—and aims to match information to people to generate revenue.

Differences in business goals

Search : primary goal is relevance (the "correct answer"), followed by CTR/CVR/GMV; personalization is increasingly important.

Recommendation : goals vary by domain—watch time for video, CTR for news, order value for e‑commerce—aligned with user engagement metrics.

Advertising : uniformly targets accurate prediction of CTR and CVR.

Differences in algorithmic models

The models are largely "generic"; a model that works well for one scenario can be adapted to another with minor tweaks, while differences mainly lie in samples and features.

Web Search VS E‑commerce Search

Search can be divided into many scenarios; here we compare web search (e.g., Baidu, Sogou) with e‑commerce search (e.g., Taobao, JD).

Data differences : Scale – web search handles trillion‑level pages, while product search deals with billions of SKUs. Source – web search crawls the whole internet; product search uses internal business databases. Structure – web search processes largely unstructured data; product search works with high‑quality structured data.

Optimization target differences : web search focuses on relevance and freshness; e‑commerce search optimizes CTR, CVR, GMV on top of relevance.

Personalization differences : web search offers limited personalization; e‑commerce search requires strong personalization to match products to users' purchasing power.

E‑commerce Search Systems

Below are high‑level overviews of several major e‑commerce search architectures.

Taobao Search

From Alibaba KDD'21 paper "Embedding‑based Product Retrieval in Taobao Search": Query → Recall → Rank.

JD.com Search

Based on JD's sharing on semantic retrieval and product ranking, the pipeline includes: Query understanding (correction, rewriting, expansion, segmentation). Recall stage (retrieve candidate products from the catalog). Ranking stage (score candidates with many features to select the best to display).

Meituan Search

Meituan focuses on O2O services (group buying, travel, medical, etc.) where location is extremely important. The system consists of data, recall, ranking, and presentation layers.

58.com Search

Development History

The evolution of search, recommendation, and advertising follows a typical internet project lifecycle:

Project launch – get the pipeline running.

Automation – transition from manual to automated processes.

Intelligence – apply AI techniques to further improve effectiveness.

Early stages relied on small data and manual classification; web portals dominated. The 2000s saw rapid growth of search (Google, Baidu) and social platforms (Tencent, Facebook). Since 2012, exploding data volumes and user expectations have driven specialized, intelligent solutions across verticals such as e‑commerce, travel, and real‑estate.

For more details, see the reference list below.

References

[1] "Ranking Better vs Estimating Better vs Searching More" – Zhihu discussion. [2] "Semantic Retrieval and Product Ranking in E‑commerce Search" – Zhihu. [3] "Towards Personalized and Semantic Retrieval: An End‑to‑End Solution for E‑commerce Search via Embedding Learning" – arXiv. [4] Same as [1]. [5] "Multi‑Business Modeling in Meituan Search Ranking" – Meituan Tech. [6] "BERT in 58.com Search" – 6AIQ. [7] Duplicate of [2]. [8] "How to Build a Good E‑commerce Search Engine?" – InfoQ. [9] "Search, Recommendation, Advertising" – GitBook.