Amazon's 'Customers Also Bought' — The Recommendation Algorithm That Changed Retail

Amazon's 2001 patent describes item-to-item collaborative filtering — the 'customers who bought this also bought' algorithm that generates personalized recommendations in real time, responsible for an estimated 35% of Amazon's revenue.

What it covers

This patent describes a method for computing item-to-item similarity from purchase history, then using that similarity table to generate real-time recommendations. For each product in the catalog, the algorithm computes a list of other products most commonly purchased by the same customers. This 'item similarity table' is precomputed offline. When a user visits a page, the system looks up the similarity scores for that item and returns the most similar items as recommendations — in milliseconds, because the heavy computation has already been done. This is fundamentally different from user-to-user collaborative filtering (which compares users to similar users), and dramatically faster for large catalogs because the item table can be precomputed and cached.

What it doesn't cover

• —Content-based filtering — using product attributes (genre, price, brand) rather than purchase history to make recommendations
• —Deep learning recommendation systems — modern neural network approaches that replaced simple collaborative filtering
• —Real-time streaming updates — the original algorithm used batch-computed similarity tables; real-time ML models came later
• —Sponsored product recommendations — paid placement in recommendation slots is a separate advertising product

The clever bit

In the late 1990s, collaborative filtering at Amazon required comparing each user to every other user — with millions of users, this was computationally prohibitive in real time. Greg Linden, a researcher at Amazon, proposed a solution: instead of comparing users to users, compare items to items. Items don't change their purchase history as rapidly as users change their preferences, so an item-to-item similarity table can be computed once daily (or weekly) and cached. Real-time recommendations then require only a single table lookup — instant, regardless of catalog size. The result: recommendations that feel personalized but can be served to millions of users simultaneously with minimal computation. Amazon's product recommendation engine, powered by this algorithm, generates an estimated 35% of total Amazon revenue.

Why it matters

Recommendation systems fundamentally changed the economics of discovery online. Before them, users had to search — they needed to know what they wanted. Recommendations create demand by surfacing products users didn't know they wanted. Amazon's 35% revenue attribution to recommendations is one of the most-cited statistics in e-commerce. The algorithm also enabled the 'long tail' economics Chris Anderson described — recommendation systems surface niche products to the subset of users who want them, making it economically viable to stock thousands of low-volume items that would be impossible to surface through search or traditional retail.

Real-world examples

• 1.Amazon's 'Customers who bought this item also bought' and 'Frequently bought together' widgets are direct implementations of item-to-item collaborative filtering
• 2.Netflix estimated that 75% of content watched comes from recommendations — their early recommendation system used collaborative filtering before transitioning to deep learning
• 3.Spotify's 'Discover Weekly' and YouTube's 'Up next' both descend from the collaborative filtering framework, though with modern deep learning layers added

Glossary