You buy a tent on Amazon on Monday. By Wednesday your homepage is full of sleeping bags, camping stoves and headlamps, and somehow you do need a headlamp.
That is the Amazon recommendation system at work. Here is how it picks those products, in plain words, and what a smaller store can borrow from it.
What is the Amazon recommendation system?
The Amazon recommendation system is the set of algorithms that decides which products you see on almost every Amazon page: the homepage, product pages, the cart, the order confirmation and the emails that follow.
Its job is simple to describe. Out of an enormous catalog, find the handful you are most likely to want right now, and put them in front of you.
Amazon’s own engineers described the idea back in 2003, in the paper that made the system famous:
At Amazon.com, we use recommendation algorithms to personalize the online store for each customer. The store radically changes based on customer interests, showing programming titles to a software engineer and baby toys to a new mother.
Greg Linden, Brent Smith and Jeremy York, Amazon.com, “Amazon.com Recommendations: Item-to-Item Collaborative Filtering”, IEEE Internet Computing, 2003
Two decades later that sentence still holds. Your Amazon and your neighbor’s Amazon are two different stores.
From 1998 to Rufus: a short history of Amazon’s recommendations
Amazon’s recommendations are older than most of the people who now build them for a living.
- 1998. Amazon launches item-based collaborative filtering, the method behind “customers who bought this also bought” (Smith and Linden, 2017).
- 2003. The team publishes the method in IEEE Internet Computing.
- 2017. The journal picks that paper as the one from its first 20 years that best withstood the test of time. Looking back, Amazon’s engineers cite a Microsoft Research estimate that put 30% of Amazon.com page views down to recommendations.
- 2019. Deep learning arrives. Then Amazon consumer chief Jeff Wilke called the new Prime Video recommendation models a “once-in-a-decade leap”.
- 2024 to 2025. Rufus, a generative AI shopping assistant, starts answering questions and recommending products in conversation, and Help Me Decide picks between similar items using your browsing and purchase history (Amazon, 2025).
Notice the pattern? Amazon never threw the old system away. Each new layer sits on top of the one before.
Where you meet Amazon’s recommendations
Once you start looking, you see them everywhere. The most common ones:
- Frequently bought together: the bundle under the price, usually the product plus one or two add-ons.
- Customers who viewed this item also viewed: alternatives, for when you are still comparing.
- Inspired by your browsing history and Buy it again on the homepage.
- Products related to this item: a row that also carries sponsored products, so part of it is advertising.
- Emails after a visit or an order, built from the same signals.
- Rufus, which recommends products in answer to a question such as “what do I need for a first camping trip?”
How does Amazon’s recommendation engine work?
Amazon has never published its whole system, but its papers and engineering posts describe the moving parts well enough. Think of it as four steps that repeat all day long:
Step 1: Collect signals
Everything you do on Amazon is a hint. Product views, searches, items added to the cart, purchases, returns, ratings and reviews.
Amazon reads these hints at three levels:
- You and products: what you looked at and bought.
- Products and products: which items people buy together or view one after another.
- You and other shoppers: whose habits look like yours.
Add the product data on top (category, brand, size, price, description) and the system has its raw material.
Step 2: Find matching items
This is where the two classic methods come in. Collaborative filtering looks at behavior: if people who bought what you bought also bought something else, that something is a good bet for you.
Amazon’s twist, back in 1998, was to compare products instead of people. Rather than hunting for customers like you, the system keeps a ready list of related items for every product. When you look at a tent, it simply pulls up the tent’s neighbors. That is fast, and it keeps working as the catalog grows.
Content-based filtering fills the gaps. A brand-new product has no purchase history yet, but it has a category, a brand and a description, so it can be matched to similar items from day one.
In practice Amazon blends both, and adds more: deep learning models and, with Rufus, generative AI that answers questions. Want the general picture of these methods? Our guide to what a recommendation engine is walks through them.
Step 3: Rank them for you
Finding a few hundred related products is the easy part. Deciding which eight to show, and in which order, is where the money is.
The ranking takes in your history, the page you are on (a product page needs alternatives, the cart needs add-ons), the time and the device, and what is in stock and can arrive quickly. Business rules then remove things you already bought, unless it is something you buy again and again, like coffee.
Step 4: Show, then learn
The list appears in a fraction of a second. Then the system watches what you do with it. A click, an add to cart or a scroll past are all new data for the next round.
Amazon sales results thanks to recommendations
So how much do recommendations actually sell? Here comes the honest answer: Amazon does not say. Its reports do not split out sales that came from a recommendation.
What we do know is how big the store has become while recommendations sat on nearly every page. Amazon’s net sales grew from $469.8 billion in 2021 (2022 results) to $716.9 billion in 2025, up 12% on 2024 (2025 results).
And there are a few clues about the share recommendations play in it:
- Rufus. Amazon reported that its AI shopping assistant was used by more than 300 million customers in 2025 and helped deliver nearly $12 billion in incremental annualized sales (fourth-quarter 2025 results). Shoppers who used Rufus were 60% more likely to complete a purchase (third-quarter 2025 results). This is Amazon’s own figure, by its own method.
- Page views. Amazon’s engineers cite a Microsoft Research estimate that 30% of Amazon.com page views came from recommendations.
- The famous 35%. The claim that “35 percent of what consumers purchase on Amazon” comes from recommendations traces back to a 2013 McKinsey article, which gives no method. It is quoted everywhere, but Amazon never confirmed it, so treat it as a rough guess from 2013.
For businesses in general, McKinsey’s 2021 research puts the typical revenue lift from personalization at 10 to 15%. That is a consultancy estimate across sectors, but it is a sensible range to plan with.
What smaller stores can learn from Amazon
You don’t need Amazon’s budget to copy the parts that matter most:
- Start with “bought together” and “viewed together”. Item-to-item recommendations are still the workhorse, and they need nothing but your order and view history.
- Put recommendations next to a decision: the product page, the cart, the order confirmation email.
- Use product data for new items, so they do not stay invisible until someone buys them.
- Hide what cannot be bought: sold-out items, and products the customer already has.
- Measure with an A/B test, not with a feeling. A recommendation widget that does not lift sales is just decoration.
We learned these lessons first-hand. Our own engine, Recostream, gave online stores Amazon-style recommendations with one line of JavaScript, returned each recommendation in 20 to 30 milliseconds and delivered a typical 5 to 10% sales uplift, measured by the stores in Google Analytics. GetResponse acquired Recostream in December 2022 (case study).
If you want recommendations like these on your own store or platform, our AI integration pilot adds them in about four weeks, with an A/B test and a clear go or no-go at the end. For the technical side, see how to build a recommendation system, and for the streaming equivalent, how the Netflix algorithm works.