Open a streaming app and the rows of shows are different from your friend’s. Behind them, a recommendation system has looked at what you watched, skipped and searched, compared that with millions of other viewers, and ranked the catalog for you. Online stores, music apps, social feeds, news sites and job boards run the same machinery, and for many of those businesses it drives a large share of engagement and sales.
Large systems usually work in stages, as described in Google’s overview and YouTube’s published architecture. Candidate generation quickly narrows millions of items to a few hundred plausible ones, often by finding items whose embeddings sit near the user’s. A ranking model then scores those candidates in detail, and a final pass re-ranks for freshness, variety and business rules. The two classic signals are content-based (items similar to what you liked) and collaborative filtering (people like you also liked this). Success is measured with offline ranking metrics such as mean reciprocal rank and live experiments.
Recommenders are among the most widely deployed AI systems, so they raise policy questions: feedback loops that narrow what people see, optimizing for clicks over wellbeing, and how transparent platforms are about why a given item was shown. Generative AI is now being mixed in, for example LLMs that describe why an item was suggested or chat assistants that recommend products, which overlaps with semantic search.
Sources
- Google Machine Learning: Recommendation systems overview — Describes the common architecture of candidate generation, scoring and re-ranking, and content-based versus collaborative filtering approaches; checked October 24, 2026.
- Covington, Adams and Sargin: Deep Neural Networks for YouTube Recommendations — Describes YouTube's two-stage system: a candidate generation model followed by a separate ranking model; checked October 24, 2026.
