Recommendation Systems

Recommendation systems (also called recommender systems) are AI algorithms that predict and surface content, products, or connections a user is likely to find relevant. They are among the most commercially impactful applications of machine learning, driving an estimated 35% of Amazon's revenue, 80% of Netflix viewing, and virtually all content surfaced on TikTok, YouTube, Spotify, and social media feeds.

Three fundamental approaches power recommendation systems. Collaborative filtering identifies patterns among users: "people who liked X also liked Y." It requires no understanding of the content itself, only behavioral signals (views, purchases, ratings). Content-based filtering analyzes item attributes (genre, keywords, features) and recommends items similar to what a user has previously engaged with. Hybrid approaches combine both, and modern systems use deep learning to model complex, non-linear relationships between users and items.

Modern recommendation systems are sophisticated multi-stage pipelines. Candidate generation narrows millions of possible items to a few hundred candidates using fast, approximate methods. Ranking scores candidates using detailed models that consider user history, context (time, device, location), and item features. Re-ranking applies business logic and diversity constraints to produce the final ordered list. Each stage uses different model architectures optimized for its constraints.

The intersection with LLMs is reshaping recommendations. Traditional systems operate on behavioral signals (clicks, purchases); LLMs can understand semantic meaning, enabling recommendations based on natural language descriptions of preferences. The shift from keyword-based search to conversational discovery — central to Generative Engine Optimization — transforms how recommendations are delivered. Instead of a ranked list, an LLM can explain why it recommends something and engage in dialogue to refine suggestions.

The attention economy implications are profound. Recommendation algorithms determine what billions of people see, read, watch, and buy. They shape public discourse, cultural consumption, and commercial outcomes. The optimization target matters enormously: systems optimized for engagement can amplify sensational or divisive content, while those optimized for user satisfaction may produce healthier outcomes. This is a central concern in AI governance and content moderation.

Jon Radoff's research on AI-driven discovery highlights a critical shift: as 58% of consumers rely on AI for product recommendations, the recommendation system is no longer just a feature within platforms — it's becoming the primary interface between consumers and the world of available products, content, and information. The 6x conversion rate advantage of AI search over traditional search underscores how effective AI-mediated recommendations can be when they work well.