A recommender system predicts what a given person is most likely to want next and orders content accordingly: the products Amazon suggests, the videos a streaming service queues up, the posts in a social feed, the ads you see. It is arguably the most economically consequential form of AI in daily life, quietly shaping a large share of online purchases, viewing time, and attention, long before chatbots arrived.
A landmark technical reference is the 2009 paper “Matrix Factorization Techniques for Recommender Systems” by Yehuda Koren, Robert Bell, and Chris Volinsky (IEEE Computer, DOI 10.1109/MC.2009.263), written by members of the team whose methods dominated the Netflix Prize. Matrix factorization works by discovering hidden factors, learned numeric features, that describe both users and items, so that a user’s predicted taste for an item is computed from how their factors line up with the item’s. These learned factors are an early, influential form of the embeddings now used throughout AI, and the similarity comparisons they enable are conceptually the same operation a vector database performs at scale.
The field was galvanized by the 2006 Netflix Prize, the million-dollar public contest to improve Netflix’s recommendations by 10 percent, which popularized matrix factorization and ensemble methods. Modern systems have moved beyond pure matrix factorization to deep learning, but the core idea endures: learn compact representations of users and items, then rank by predicted preference. Recommenders rely on two ingredients, collaborative filtering (learning from the behavior of similar users) and content features (attributes of the items themselves), usually blended.
Why business readers should care: recommendation is where AI most directly touches revenue, retention, and engagement, which makes it both a powerful lever and a source of well-documented harms. Optimizing purely for clicks or watch time can amplify addictive, polarizing, or biased content, and the systems can entrench feedback loops that narrow what users ever see. Any organization deploying personalization should be explicit about what metric it optimizes, because that choice quietly governs user behavior at scale.