What are personalized recommendations?
Whether you are reading news, watching videos on Youtube or shopping on Amazon, you have probably noticed rectangular boxes, popups or side menus at the end of a page, that suggest content you should look at next. Those content recommendations come in different shapes and have different goals depending on the type of site you are on.
On media sites, news sites and blogs, typical content recommendations come in the form of banners at the end of an article or popups suggesting the next article/video you should read/watch.
The goal for media sites is to try and guess what you are interested in and get you to read more articles and/or watch more videos. The more articles you read, the more pages you load, the more ads can be displayed. This makes sense as one common business model of media sites (also called publishers) is to get paid on CPM - Cost per 1,000 ad impressions.
Therefore, the question every media site should be asking is: "What is the best article recommendation we can give to this specific user on the specific page he is currently looking at in order to maximize our chances of him reading another article?"
On online shops, the challenge is slightly different compared to media sites. Not only do you want to increase the number of product pages that each visitor sees, but you also want each visitor to discover products he will like and ultimately buy one of them.
A visitor rarely comes to an online shop with the strong intent of buying a product: typically the conversion rate of online shops is only between 1 and 3%. In addition, most visitors don't stay for long: they look at two or three product pages, and then leave. It is therefore crucial that the first page the visitor sees is as effective as possible in guiding the visitor to a product he will ultimately buy.
Brick-and-mortar stores face the same type of challenge. In order to increase their customers' average cart value, stores attempt to make data-driven decisions regarding which products to place next to each other. Based on receipts, they can understand which products are bought together and optimize product placement accordingly.
This 2011 lecture by Alan Penn of the UCL Bartlett School of Architecture and the above picture illustrate how much effort is put in the design (or shall we say "maze"?) of Ikea stores.
What a brick-and-mortar store cannot do is rearrange their whole shop in order to fit the tastes of each individual shopper, even if they knew each shopper very well. However, online shops can do that, with personalized product recommendations.
Amazon, a pioneer of personalized product recommendations, can be regarded as the online equivalent of Ikea in that respect. Personalized product recommendations are displayed throughout the site. Indeed, personalized recommendations have been a vital part of Amazon's success story. Fortune magazine provides more details:
Judging by Amazon’s success, the recommendation system works. The company reported a 29% sales increase to $12.83 billion during its second fiscal quarter, up from $9.9 billion during the same time last year. A lot of that growth arguably has to do with the way Amazon has integrated recommendations into nearly every part of the purchasing process from product discovery to checkout.
This is in agreement with our own observations that personalized recommendations have an overwhelmingly positive effect.
How useful are personalized recommendations?
Obviously, the performance numbers vary between customer profiles (number and type of products, traffic), but here are some numbers which are a general indicator across all the shops we work with.
Visitors who click on recommendations are better customers: they visit more product pages, put a higher number of products into their carts, and are more likely to come back.
Visitors who never click on recommendations often visit just one product page, whereas users who do click on recommendations typically visit between 3 and 6 product pages. These numbers confirm that the recommendations must have been perceived as interesting by the visitors.
This effect of recommender systems is well-established and has been observed by others1 on media sites, among other examples.
Visitors who click on recommendations are also more likely to add more products to cart: between 20% and 90% more, depending on the type of products the shop is selling (the lower the price of the product, the higher the increase).
How are personalized recommendations generated?
Giving good recommendations is not easy, not only because each visitor is unique and has specific interests, but also because each recommendation has to be served very quickly (within a few tens of milliseconds) every time a new page is loaded. And this typically happens millions of times per day!
An additional difficulty is that the recommender system should work well, no matter what kind of products an online shop sells. It should not matter if the shop sells movies, books, clothing, electronics, or toys.
Finally, the recommendations generated by the system should reflect actual visitor behavior on the site. It is always possible to define hand-crafted recommendation rules, but there is no a priori way of determining if these rules are effective or not. This problem can be solved by relying on actual visitor behavior and machine learning techniques which analyze and make predictions based on this data in real time. This will be the subject of a future blog post.
When trying to increase the conversion rate of your online shop, you should consider using personalized product recommendations. SaaS recommender systems use anonymous visitor behavior data to understand the behavior of each visitor and predict which products from your catalog a visitor is interested in.
The increase in conversion is only one of the direct benefits. Personalized product recommendations also help visitors navigate your site more comfortably and makes them feel more valued.
F. Garcin et al. "Offline and online evaluation of news recommender systems at swissinfo.ch." Proceedings of the 8th ACM Conference on Recommender systems. ACM, 2014. ↩