Selling through recommendations

You've probably experienced it. In many webshops they have the feature that when you look at a product or put it in your shopping cart - you will be presented with a number of other products that you may also be interested in.

The best known example is probably Amazon.com. When I last visited the site, I got these suggestions for books I might be interested in:

Of course, this happens not simply by chance, Amazon does it because it provides additional sales. The interesting thing is that there is actually a big difference between the performance of various techniques for doing these product proposals.

The simplest version simply makes statistics on products that often end up in the same shopping cart. It may be that batteries are often needed when purchasing a certain electric product and after the first series of customers have all put both product and batteries in the shopping basket, the system will in future be able to propose batteries as an additional purchase. This kind of product proposal is primarily an optimization of the buying process, you should probably remember to buy batteries anyway so it becomes a little easier when you can just press a button with "yes" to buy the batteries as well.

The more advanced versions use a machine learning technique called "collaborative filtering". It does not just look at shopping baskets but takes advantage of the fact that customers' shopping profiles are often similar. There are many different variants and it is quite technical, but the basic idea is not that complicated so let me try to explain how it works without getting into the mathematical details.

Imagine that you are a customer in a webshop and have looked at a number of products, some of them maybe several times and a few you have been so interested in that you bought them. As you repeatedly visit the website, all of this data about you is collected and they are considered different degrees of interest you have shown. The system then uses the entire list of products you have shown interest in to create a profile of you as a customer.

Now the system knows a lot about you and what you like, and by creating the same kind of profile for all other customers, the system can get an overall picture of how different customers' patterns of interest are.

The smart trick that you use for these systems is to design these customer profiles in a special way, so you can mathematically get a figure for how similar 2 chosen profiles are. Therefore, with your profile in hand, the system can look up in the database and find other customer profiles that are similar to you and simply by looking at what these similar customers have purchased that you have not yet purchased, products that you may be interested in buying may be suggested.

These kinds of product proposals can even suggest products that are from a completely different category and that you may not have even looked at yet - solely because there are many other customers who have the same taste as you who also like the products in that category. It is not necessary that the products have ever been in the same shopping cart. If there is enough data for a good comparison basis, it simply gives better suggestions.

But is it really that important if you can give some better suggestions? The answer to that question is found in the fact that the movie service Netflix back in 2006-2009 ran a competition where the goal was to improve their algorithm by just 10% and if you could do so, you would win 1 million dollars!

Netflix has figured out that the precision of product proposals is very important to the customer experience. This is because product proposals are a type of service where it is important that the customer is not disappointed by trying out the suggestions.

If you as a customer over time find that the proposals usually give you a good experience, for example in the form of a movie you hadn't seen before which was really good, you end up relying on the suggestions you get. Conversely, if the suggestions just aren't quite good enough for you to dare to trust them, then you might end up completely ignoring them. In other words, there is a kind of lower limit for precision where if the system is just slightly better than that, then you as a customer will trust it, otherwise not. This also means that improving the system from perhaps being just below the limit to just over the limit actually has a disproportionate impact on how much customers use and appreciate the system.

Once you have a system on the product pages to suggest to customers other products they might be interested in, then in some other places you can also use this knowledge of the customers to your advantage.

An obvious option is to personalize the front page of the webshop, the vast majority of webshops today just have a list of popular products or seasonal items on their front page. Instead, on the front page of the webshop you can already give each customer some tailor-made suggestions for products that could be interesting (here you should give extra weight to products that the customer has previously looked at but not yet purchased).

Perhaps even more interesting is that typically the webshop has customers registered with emails and instead of sending out standardized newsletters on a regular basis, you can strategically send out offers to customers where each email contains a tailor-made selection of product offers. Some companies even go so far as to provide special discount codes in their emails that apply only to the customer and to certain products that the customer has previously looked at.

It is fine to give tailor-made suggestions to the customers, if done right it will provide additional sales. But one should not just uncritically send out personal discount codes to people to motivate them to buy. Partly, some customers will start to speculate on prices if it's too regular and then studies have shown that people lose interest if they feel it becomes too personal - they end up suspecting the company to only recommend the product because it maximizes profit margin. People love to "discover" good deals, it should preferably not seem too constructed. One possibility is therefore to leave the personal offers on only one of several sections in the newsletter.

With all the different options for using intelligent techniques to optimize and get some extra percent conversion in one's shop, however, one must always keep in mind to ensure the basic user experience on the site.

I had a strange experience a while ago when I saw a big advertisement for "Den blå avis" on extrabladet.dk. The advertisement was obviously made with machine learning techniques and the exciting thing was that the items that were suggested were actually some I found interesting. After going to the front page and finding that the suggestions placed there were of much poorer quality, I learned over time that I just had to go to the page of the page with the ads to get the good suggestions. It is quite unfortunate when your customers resort to advertising on other people's websites to find the products they are interested in, but conversely it is also a good illustration of how important product suggestions can be for the user experience. Don't let your customers' user experience be worse than watching ads.