Amazon
Item-to-item collaborative filtering product recommendation engine, served across every touchpoint
Amazon's item-to-item recommendation engine, designed around 1997 and recognized by IEEE in 2017, powers the homepage, product page, cart, and email; McKinsey estimated in 2013 that 35% of purchases on Amazon came from recommendations.
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, weighted by their proximity to what the person has already looked at. This engine feeds the homepage carousels, the "Frequently bought together" blocks on product pages, the cart recommendations, and the email campaigns. The method was designed around 1997 at Amazon by Greg Linden, Brent Smith, and Jeremy York, then described in a paper published in 2003. In 2017, the editorial board of IEEE Internet Computing named it the article in its history that had best withstood the test of time. The system remains the foundation of Amazon's recommendations, enriched since by more recent models.
Results Proof C
The mechanism and its scaling are documented by Amazon's official science blog and by an IEEE paper recognized in 2017, that is nearly three decades in production. The 35% figure is a 2013 McKinsey estimate, not a financial disclosure from Amazon, so it is not presented as a level A result.
How it works
Documented architectureThe stack in detail
- llm Filtrage collaboratif item-to-item Amazon's in-house algorithm designed around 1997 (Linden, Smith, York), published in 2003 and recognized by IEEE in 2017; still the foundation of the recommendations.
- infra Pipeline batch de pre-calcul des similarites Product-to-product tables built offline, the key to catalog-scale operation and real-time serving.
- outil Service de recommandation temps reel Ranking of candidates at display, served on the homepage, product page, cart, and email.
How it runs, concretely
For ops teams-
1Collect the signals AI / platform
Purchases, cart adds, views, and clicks stream in continuously and are attached to the relevant product.
-
2Precompute 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.
-
3Serve 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.
-
4Loop 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.
Purchase and browsing co-occurrences per product. On a new catalog or a product with no history (cold start), the engine has nothing to associate and falls back on weaker backup heuristics.
How your customers perceive this type of use
Sourced studiesLe paradoxe est documente des deux cotes : 71% des consommateurs attendent des interactions personnalisees et 76% sont frustres quand elles manquent (McKinsey, 2021), mais 75% declarent ne pas acheter aupres d'organisations auxquelles ils ne confient pas leurs donnees (Cisco, 2024). La « creepy line » est localisee : messages recus quelques secondes apres une recherche et suivi de localisation sont les pratiques qui mettent le plus mal a l'aise (Periscope by McKinsey, 2019).
Acceptance conditions
- La confiance dans le traitement des donnees precede l'achat : 75% ne achetent pas sans elle (Cisco 2024)
- Un cadre legal protecteur rassure : 59% des consommateurs disent que des lois fortes sur la vie privee les rendent plus a l'aise pour partager des informations dans des applications IA (Cisco 2024)
- La personnalisation elle-meme est attendue quand elle est consentie : environ la moitie des consommateurs (US 55%, UK 52%) disent s'inscrire souvent ou parfois a des services personnalises (Periscope by McKinsey 2019)
Red lines
- Le message declenche quelques secondes apres une recherche ou un achat : deuxieme ou troisieme cause de malaise selon les pays (Periscope by McKinsey 2019)
- Le suivi de localisation percu comme de la surveillance : 40% de malaise en Allemagne et au Royaume-Uni (Periscope by McKinsey 2019)
- Le mesusage des donnees personnelles par l'IA, devenu la premiere inquietude des consommateurs, a 53% et en hausse (Qualtrics 2025)
Sources: McKinsey & Company 2021 · Periscope by McKinsey 2019 · Cisco 2024 · Qualtrics 2025
How to replicate
Inference, not sourcedData prerequisites
- Purchase and browsing history tied to the product and the customer
- Catalog with stable product identifiers
- Enough co-occurrence volume per item to yield good associations
Org prerequisites
- A data team able to maintain a precompute pipeline and a real-time service
- Display surfaces instrumented to measure clicks and post-recommendation purchases
Possible stack
- A managed recommendation engine (AWS Personalize, Google Recommendations AI) for a first version without a heavy ML team
- Open source collaborative filtering (implicit, LightFM) for a custom version
The plan, step by step
- Step 1Consolidate the co-purchase and co-view history per product, with stable identifiers.Deliverable: Co-occurrence matrix over the catalog
- Step 2Compute the item-to-item similarity tables, in open source (implicit, LightFM) or via a managed engine.Deliverable: Similarity table evaluated offline
- Step 3Serve a frequently bought together block on the product pages of one catalog segment.Deliverable: Recommendation block in production on the pilot segment
- Step 4Run an A/B test against display without recommendations, measure CTR and assisted conversion, then extend to the cart and email.Deliverable: Test result and plan to extend to the other surfaces
First step: Wire a "frequently bought together" block onto product pages from co-purchases, and measure click rate and assisted conversion against display without recommendations.
The same pattern, proven in other industries
Function of Beauty
personalized product formulation by algorithm from a quiz
Henkel
genAI content supply chain for one-to-one personalization at scale
Hims & Hers
AI-driven personalized treatment routing on intake data (MedMatch)
Sources
- S1 The history of Amazon's recommendation algorithm - Amazon Science Primary archive pending
- S2 How retailers can keep up with consumers - McKinsey & Company Secondary archive pending
- S3 Amazon.com recommendations: item-to-item collaborative filtering - IEEE Internet Computing (Linden, Smith, York) Primary archive pending
An error, newer info, a source?
This page lives on its accuracy. If a figure has moved, if the deployment has changed, or if you have a higher-quality source, tell us. Every sourced correction is verified before publication.