The sentence you've read a thousand times
How Amazon's “customers who bought this also bought” works, and what it deliberately ignores at request time
Beat by beat.
| t (s) | Voice-over | On screen | Caption |
|---|---|---|---|
| 0–3 | “Customers who bought this also bought.” You've read it a thousand times. | Phrase typing, blinking cursor | ALSO BOUGHT |
| 3–7 | Amazon published how it works in 2003. The surprise is what it doesn't look at. Guess. | Paper S4 header | WHAT IT IGNORES |
| 7–14 | Amazon had more than 29 million customers and several million items, too big for the algorithms of the time. [S4] | Number cards | 29M CUSTOMERS |
| 14–24 | So instead of finding customers like you, it matches each item you bought to items customers tend to buy together. [S4] | Item → item graph | ITEM TO ITEM |
| 24–34 | It does the heavy work offline: a giant similar-items table. The live part depends only on how many items you bought. [S4] | Table being built → tiny lookup | OFFLINE FIRST |
| 34–42 | That's the answer: at request time it ignores the number of customers and the size of the catalog, so it runs in real time. [S4] | Strike-through “# customers”, “# items” | IGNORES SCALE |
| 42–50 | To be fair, it also lets customers see why an item was recommended. [S4] | “Why recommended” UI | EXPLAINED |
| 50–55 | Brew principle: move the expensive thinking to before the moment of choice. Fast is a feature. | Monochrome card | BEFORE THE CHOICE |
| 55–58 | Next week: a ProductBrew investigation we can't title yet. | Black frame + ProductBrew glyph | SEALED |
On a phone, swipe the table sideways. Every factual line carries a source chip.
How it works.
Mechanism: item-to-item collaborative filtering. Rather than matching you to similar customers, it matches each of your items to items that are frequently bought together.
Scale: with 29M+ customers and several million items, customer-to-customer methods were too slow; the similar-items table is computed offline.
The surprise: online cost depends only on how many items you have bought, not on catalogue size or customer count, so it answers in real time.
What a great PM would do: find the expensive decision in your product and precompute it, so the user's moment of choice is instant.
Steelman: The method is transparent: customers can see why an item is recommended. Explainability as UX.
Move the expensive thinking before the moment of choice.
Sources
[S4] G. Linden, B. Smith & J. York, “Amazon.com Recommendations: Item-to-Item Collaborative Filtering”, IEEE Internet Computing 7(1):76–80, 2003 · https://doi.org/10.1109/MIC.2003.1167344