/ Personalization

5 lessons you can learn from Amazon’s recommendation engine

Amazon is huge. The ecommerce giant accounted for 43% of 2016 online retail sales in the US, according to Slice Intelligence. With its latest acquisition of Whole Foods and its foray into cashless shopping with Amazon Go, Amazon looks set to assert its dominance in the physical retail space as well.

Many factors contribute to Amazon’s success, but recently, artificial intelligence (AI) is increasingly being touted as a key pillar of Amazon’s competitive advantage. And one of Amazon’s best applications of AI is in its on-site product recommendations.

Amazon strives to create a personalized shopping experience for every customer. In a page titled ‘Your Amazon.com’, users are recommended a unique selection of products based on their past shopping behavior. According to research by McKinsey, a mind-boggling 35% of Amazon’s sales come from such recommendations.

In this article, we take a closer look at how Amazon recommends products on their site, and what makes them so effective. But before that, let us see what makes their recommendation algorithm tick.

While the details of Amazon’s proprietary recommendation algorithm are secret, its broad outlines can be found in their patent application (‘Personalized recommendations of items represented within a database’). Essentially, Amazon uses what it calls ‘item-to-item collaborative filtering’ to identify pairs of related products.

This flowchart illustrates the recommendation process. In step 108, Amazon identifies pairs of products that have large overlaps in their customer base, allowing them to recommend products to customers who have only discovered ‘one half’ of the pair. This methodology is known as affinity analysis, and is also used in some form by Spotify’s song recommendation algorithm.

Next, let’s explore the places that Amazon embeds these product recommendations and how they boost conversions.

Frequently bought together

The ‘Frequently bought together’ section is found below every product listing, and suggests a combination of complementary products. The focus here is on cross-selling products to increase order size. Notice that instead of recommending 3 laptops or 3 laptop bags, Amazon recommends a mix of products, and encourages the customer to ‘add all three to cart’. Related products do not necessarily have to be in the same product category but could also be products that are used together.

Learning point 1: Product recommendations can be a great way to sell packages or a larger suite of products. If your aim is to maximize order size, then you should consider recommending a group of products that customers can buy as a bundle.

Customers who bought this also bought

This section, found below the ‘frequently bought together’ section, focuses more on product discovery. Customers can scroll through a longer list of related products to find what interests them. The title of this section leverages on social proof, suggesting to shoppers that products here are trusted by people just like themselves. This screenshot was taken in a product listing for a laptop - like the previous section, the recommended products here are not other laptops, but accessories, peripherals, and software. This strategy encourages customers to add more items into their cart, instead of replacing the items in their cart.

Learning point 2: The name of your recommendation widget is important. It functions as a clue to customers about why those products are being shown, and it should inspire trust in your recommendations. The paradox of choice means that sometimes providing too many choices of products in the same category can overwhelm customers, so you can explore recommending complementary products instead.

Compare to similar items

In the ‘Compare to similar items’ section, Amazon lists out a few products in the same product category, and facilitates a specs comparison. Such comparisons are valuable when purchasing electronics, but are not applicable in all categories. This section is not shown in listings of books or movies, for instance.

Learning point 3: Online shoppers tend to savvy buyers who do their research and know how to compare prices. Make their shopping experience smoother and don’t give them an excuse to leave your site by providing useful information that facilitates a purchase decision.

Recommended for you

Amazon also provides recommendations based on your purchasing data. These recommendations are only available to registered members, and only appear after a user has added at least one item to their cart. A new member who has not made any purchases will find his recommendations page populated with best-sellers instead.

These recommendations can be filtered by product category, helping customers zero in on what they are looking for. Amazon encourages customers to interact with these recommendations by rating the item, declaring that ‘I own it’ or that I’m ‘Not interested’. Such interactions will improve the accuracy of the recommendations by helping Amazon understand the customer better. Amazon even has a help page teaching members how to improve the recommendations offered to them.

Learning point 4: Recommendation algorithms are never going to be perfect the first time. Algorithms may recommend products that customers have no interest in, and that’s why it could be valuable to allow them to offer feedback. Amazon also informs customers why each product was recommended, so that they would not be as put off if a strange recommendation appears.

Learning point 5: Using real-time recommendations to acquire new customers

Recommendations are great when you know something about your customers, but for most growing ecommerce stores, many visitors are first-time users who have not created an account with you. To reap the benefits of personalization, it is not enough to only target members who are registered on your site.

Like the proverbial iceberg, it is the first-time users lurking beneath the surface who can make or break your business. They may not be immediately visible because they have not created an account with you, but they form a significant portion of your potential customer base.

The key to converting visitors to customers is to use real-time recommendations.

Many recommendation algorithms make recommendations by looking at what a customer has purchased previously. Real-time recommendations go one step further. They analyze what products customers are clicking on in the ecommerce store, what categories they are browsing in, what banners and ads they are attracted to. This data is obtained for all visitors - even first-time users - and used to adapt product listings on the fly.

Like a salesperson who observes what products a customer is looking at before recommending similar products, real-time recommendations make instant suggestions based on the customer’s interactions on the ecommerce store.

Another benefit of real-time recommendations is that it gathers data and refreshes automatically, without customers having to manually give feedback. Not all customers are savvy enough to rate the recommendations given to them, and there is much to be said for a system that works seamlessly for the customer.

Over to you

Product recommendations are a mainstay of some of the world’s most powerful ecommerce stores, and for good reason - they have been shown on many occasions to dramatically boost conversion rates and average order sizes.

We hope you have found this article informative, and that you’ll walk away inspired to use product recommendations to supercharge your sales!

Check out Metisa if you'd like to supercharge conversion rates with real-time recommendations for your Shopify, BigCommerce, Magento or custom built site.

Justin Yek

Justin Yek

Partner & cofounder at Altitude Labs, creator of Metisa, former investment banker, public speaker, hobbyist musician

Read More