A recommendation engine is a tool that uses machine learning to filter specific items from a larger dataset. There are different types of filtering methods, but in general, the recommendation engine uses existing user insights - what a user has previously interacted with, what products were purchased, which movies were watched, and so on, to match other items similar to these, and then makes an item recommendation. The Amazon recommendation system works in a similar way.
Because 'recommendations' are made based on a user's past interactions and interests, they are personalized to each user and the user is therefore very likely to find the recommended product interesting as well.
And this is not an assumption, but a proven fact. Firstly, the very act of delivering personalized content is powerful - 80 percent of consumers are more likely to purchase from a brand that delivers personalized content, according to a study by Epsilon Marketing.
Secondly, a brand that recommends products experiences higher conversions rates than a brand that doesn't - customers who click on product recommendations are 4 times more likely to add that product to cart and complete the purchase according to a study by PracticalEcommerce..
The most potent example of the benefits of recommendations is probably Amazon. Amazon is the largest e-commerce brand in the world in terms of revenue and market share. (Statista)
This clearly elucidates the power of recommendations. In this case study, we look at how Amazon is using recommendations across its store and even off it, the technology behind the recommendation engine, and how you can implement similar strategies to increase engagement and conversions on your e-commerce store.
Decoding the Amazon Recommendation Engine
Here is a sneak peek into Amazon's efficient recommendation engine and how it works:
On-page recommendations - recommendations made on Amazon.com
User-specific product listing
Clicking on 'My Recommendations' on the menu bar takes you to a page where Amazon displays recommended products based on:
- Your past purchases,
- Your past interactions (products that you have visited but not purchased, search history, and so on)
All products listed on this page are filtered by the Amazon recommendation systemengine based on your shopping behavior. These are products you have shown an interest in earlier but not purchased, products similar to those you have already purchased, or products similar to those you have shown an interest in but not purchased.
This is one way how Amazon uses recommendations to increase engagement and conversions. This option, however, requires action from the user - they have to click and visit the recommendations page. To increase visibility, Amazon takes the recommendations to the user directly, in real-time, as you will see in the next few methods listed.
Recommendations based on purchases made by similar users
On visiting the product page for an Oculus VR headset, Amazon displays the following recommendations after the product information:
Here, Amazon is looking to cross-sell products that other users paired along with their Oculus VR headset purchase. Because users similar to you who previously purchased the Oculus headset also bought a travel case for it, Amazon's recommendation system presents the travel case with the assumption that you will find it useful as well.
You might find another section of recommendations, "Customers who viewed this item also viewed" along with the above one, like so:
Here, Amazon is filtering the recommendation by users who viewed the Oculus headset (and not bought it) and other similar items that the user then went on to view.
You might also see the following segment based on which product you are viewing:
Here the products are being filtered again based on users similar to you who viewed the product you are currently viewing.
In all these cases, the recommendation engine filters items based on two factors:
- The features of the product,
- User characteristics - to find other users similar to you who also purchased the product you are viewing
Cross-selling based on categories/product relationships
In this method, Amazon uses the features of the product to make a recommendation. On viewing the product page of a Dell gaming laptop, Amazon makes the following recommendations after the product details:
The mouse recommended in this section is a gaming mouse, and 'gaming' is a core feature of the laptop that is selected, and the backpack is a Dell gaming backpack for laptops, which again are the core features of the item that is selected.
Amazon is looking to help the user find products that accompany the current product being viewed, by filtering items based on product features and categories.
This segment is similar to the "Frequently bought together" recommendation, which pops up when you select the gaming backpack:
These recommendations seem to be made based on product features and based on what other users similar to you have purchased together in one transaction. In both samples, however, product features and their category (gaming, Dell, and so on) seem to be the primary filter.
Recommendations based on browsing history
On logging in to Amazon, users are generally presented with recommendations based on products they showed interest in but did not purchase:
The two sections "Keep shopping for" and "Pick up where you left off" are recommendations based on your browsing history - products you actually viewed but didn't purchase.
Another segment you will often see is suggestions of items similar to those in your browsing history:
These recommendations are also based on your browsing history but are items similar to those you viewed.
Narrowing based on interests and offers
Amazon takes recommendations one step further by further narrowing them based on available offers:
Here, Amazon is recommending products based on your browsing and purchase history, but only those products that have an active discount or offer. This further increases the chances of capturing interest and making a conversion.
If you visit the product page of an older version of Kindle, you will see the following recommendation:
Here, Amazon is recommending a better (and more expensive) version of the product you are viewing. The purpose of this recommendation is to upsell - to sell a better version of the product which will also bring in more revenue. The filtering system looks at product features to identify the right recommendation.
Off-page recommendations - recommendations made on platforms other than Amazon.com
Amazon uses on-page data like your purchases and browsing history to make off-page recommendations as well, and one way is through personalized emails:
Because the user browsed for VR headsets on Amazon's store, they sent a follow-up email with a product recommendation.
The above recommendation was based on browsing history - products actually viewed. Amazon also sends personalized emails with recommendations of products similar to what were viewed:
Amazon runs display ads across their partner networks. The ads are not of random products but are personalised to each user based on their past purchases and browsing history.
In this example, the user is seeing an ad for a digital camera on a blog page. This is because the user browsed for cameras on Amazon and Amazon then targeted the user with a personalized display ad.
How does Amazon’s recommendation engine work?
Here's what Amazon says about its recommendation system:
"We make recommendations based on your interests.
We examine the items you've purchased, items you've told us you own, and items you've rated. We compare your activity on our site with that of other customers, and using this comparison, recommend other items that may interest you in Your Amazon.
Your recommendations change regularly, based on a number of factors, including when you purchase or rate a new item, and changes in the interests of other customers like you."
The Amazon recommendation system uses different filtering methods to recommend products.
Item-to-item collaborative filtering
One is the item-to-item collaborative filtering system (source). For the system to work, the engine needs to collect tons of user and product data and create relationships between them. There are three relationship models that are created:
- User-item: A user specific matrix is created that contains the data of all products they have purchased and interacted with.
- Item-item: The item-item matrix contains a mapping of product feature similarities. A gaming laptop and gaming mouse have the relationship of being an electronic item, a computing item, a gaming product, and so on.
- User-user: This matrix contains a mapping of the similarities in user characteristics. Two users who purchased the same product and then gave it a rating of 4, for example, are mapped together.
The item-item collaborative filtering system uses these matrices to filter out relevant products for each user. Here's a simple example illustrating how it works:
- In this example, users Tim, Amy and John are first grouped as similar in a user-user matrix based on the similarity in their past purchases, browsing history, and product rating. In this case, the similarity is in their interest in the ice cream cone (the only common product all three have purchased).
- Next, the recommendation engine looks to create relationships between items (based on user interest and not product features). Here, because Tim and Amy have both shown interest in a sundae and an ice cream cone, the engine maps the two together.
- The sundae is then recommended to John because that item (sundae) was mapped to another item (ice cream cone) he has already shown an interest in. Hence, item-item collaborative filtering.
When you click on an Oculus VR headset and Amazon recommends a travel case, here's what is happening:
- You are a part of a user-user matrix based on a similarity in past purchases - you and other users have purchased or browsed the same products, or given the same products similar ratings.
- The engine then tries to create product relationships based on the interactions of users within this group. In this case, many users who bought the VR headset also bought the travel case, and so they were mapped with each other.
- When you show interest in the VR headset, you're shown the travel case as a recommendation because it is considered an item of interest by other users in your group who also bought the VR headset.
This is how most recommendations like the "Customers who bought this also bought" work on Amazon.
Amazon's recommendation engine also uses the content-based filtering system to recommend products. It utilizes the user-item and item-item matrices to achieve this.
Once you interact with a product, it looks for other products similar to this (based on feature relationships in this case) and then presents them to you. If you view a Dell gaming laptop, you will be recommended other gaming laptops based on similarity in features - CPU cores, processor type, RAM, storage capacity, and so on.
This is how the "related to items you have viewed" and other similar recommendations work.
Bandit and causal inference in recommendations
Bandit and causal inference in filtering are newer models that are being researched and improved.
The concept of 'bandits' is derived from the analogy of a gambler at a slot machine which has multiple arms, each with unknown probabilities for a payout. The gambler systematically tries each arm and analyzes the results to find the best option for the maximum payout.
This concept is applied by the engine to identify the best filtering algorithm for each user based on which recommended products they interact with.
Causal inference in filtering is another method of improving existing algorithms. Causal inference in statistics is a method of identifying the cause of an effect. In the case of Amazon, the effect is the user clicking a product link and the engine uses causal inference to identify the factors that might have caused this action.
3 Ways Amazon uses AI for personalized product recommendations
In-store (web and mobile) recommendations
Amazon user recommendations on its store to personalize the shopping experience for every consumer. Products shown on the homepage, suggested items on the product page, and so on, are all tailored to the individual user using a recommendation engine.
Amazon's voice assistant Alexa also uses AI to collect data points and deliver personalized recommendations. For example, based on what music a user tends to listen to on Alexa, the voice assistant creates personalized playlists and suggestions.
Amazon GO is Amazon's unmanned physical store, where consumers can walk in, pick a product, and leave. These stores use cameras and AI for computer vision to track users and products being picked, scan the barcode, and then add products to the user's cart on their Amazon GO app. Users can either pay later or the money is deducted through a preselected payment method. This purchase data is attributed to the user who then gets personalized recommendations on other Amazon platforms, like Alexa and Amazon.com.
Achieving Amazon-like personalization for your e-commerce store
Create a relevant user-journey
Amazon has created a seamless user journey that nudges the consumer forward. The homepage makes it easy for the user to find the product they want, the product page has all information laid out and then recommendations are made to upsell or cross-sell, and then users are walked through an easy to use checkout process. A great UX is important for the success of an e-commerce store.
Buy a personalized recommendation engine
We have already established that implementing a recommendation engine will increase engagement and sales in an e-commerce store. Recommendation engines, however, are not easy to build, train and enhance.
Fortunately, there are solutions available that you can plug and use to get the benefits of recommendations right from day 1. These solutions can be implemented without any coding and they are also tailored to work for your specific store.
If you would like to know more about implementing a solution for your store, talk to us and get a demo of our recommendation engine.
Follow a simple checkout process
One of the top reasons for cart abandonment is a long and complicated checkout process.
Amazon uses smart tactics to simplify its checkout process. One is the 'Buy Now' button that lets you immediately purchase a product without having to go to the cart page. Two is by saving information like your address and the preferred payment method they reduce the number of steps in the checkout process.
Use a transparent return policy
A return policy gives consumers a feeling of security. It also helps users get over last-minute resistance that often crops up when buying a product online. On the flip side, not having a return policy makes consumers suspicious of the brand especially since major e-commerce stores like Amazon offer an easy way to return or exchange products.
Make users feel valued
Customer loyalty is a big driver of sales, and loyalty is nurtured by creating value. In fact, a report by Loyalty Lion stated that loyal consumers spend 67 percent more on average than new ones. Amazon offers its Prime customers a lot of discounts and benefits that make them feel valued and special. Implement such strategies that are personal to the user and make them feel valued by your brand.
Implementing a recommendation engine like Amazon's
Recommendations enhance the level of personalization in an e-commerce store, and we explored the different ways Amazon uses recommendations to promote, upsell and cross-sell products throughout their store. In fact, once you have interacted with products on Amazon every segment on the site is essentially based on recommendations.
Are you looking to implement a custom recommendation engine on your e-commerce store? Drop us a message for a demo of our product and we'll help you launch a fully-functional recommendation engine.