Amazon

Objective

Surface the right product at the right moment for each visitor, on every surface of the site, with no manual merchandising work.

The deployment

Amazon developed a recommendation method known as item-to-item collaborative filtering. Rather than matching a customer with other similar customers, the algorithm starts from each item and precomputes the products that frequently appear together in purchase and browsing histories. When a visitor views or buys a product, the system surfaces in real time the items most associated with it. In 2003 its authors describe an in-house engine, built because existing algorithms could not handle the load of Amazon's tens of millions of customers and products, and already used on most of the site's pages as well as in many email campaigns. The method was designed around 1997 by Greg Linden, Brent Smith, and Jeremy York, then published in IEEE Internet Computing in 2003. In 2017, the journal's editorial board named it the article in its history that had best withstood the test of time. On commercial impact, Amazon has never published a figure. The 2003 paper only claims, without quantifying it, that the click-through and conversion rates of recommendations vastly exceed those of untargeted content such as banner ads or top-seller lists. The "35% of purchases on Amazon come from recommendations" line that circulates everywhere is not an Amazon figure: it is an estimate published by McKinsey in 2013, with no methodology disclosed, that the company has never confirmed. Researchers dispute it: Anuj Kumar (University of Florida) finds it hard to believe, and the study he published with Kartik Hosanagar in Information Systems Research, run at a department store rather than at Amazon, measures a different indicator - an 11% gain in sales attributable to recommendations. Finally, no public Amazon document confirms that item-to-item filtering remains the basis of the system today: the most recent primary documentation, the 2019 Amazon Science blog, retraces its history and describes more recent models that outperform it on Prime Video movie recommendations.

Results Proof C

The mechanism, the scale, and the longevity are documented by Amazon itself (the 2003 IEEE paper authored by its engineers, the Amazon Science blog), and corroborated by the 2017 IEEE recognition. But Amazon has never published a quantified commercial result for its recommendations: the case therefore cannot claim level A (financial disclosure) or level B (quantified case study). The 35% of purchases figure, widely recycled, is a 2013 McKinsey estimate that the company has never confirmed and that researchers dispute; it is not retained as a result.

1. 1
   Collect the signals AI / platform

   Purchases, cart adds, views, and clicks stream in continuously and are attached to the relevant product.

2. 2
   Precompute the similarities Data science team

   For each item, the list of the most co-consumed products is built offline. This precompute is what makes it possible to handle the load at catalog scale.

3. 3
   Serve the recommendation AI / platform

   At display, the system cross-references the items viewed or bought by the person against these tables and ranks the candidates.

4. 4
   Loop on behavior AI / platform

   The click or purchase that follows a recommendation becomes an input signal again. The system corrects itself without a merchandiser hand-picking the products.