If you’re an ecommerce business or selling any sort of products to consumers, you’ve likely come across product recommendation systems. Anyone making a purchase from bigger online brands like Amazon or NetFlix has likely been recommended books to buy or videos to watch. Indeed, many marketers and sales consultants have written about the effectiveness of product recommenders increasing sales.

What do product recommenders do?

The problem solved by product recommenders is simple to state but difficult to do: take all your customers’ past behaviors and find which products they are likely to buy next. Such recommendations solve two problems:

  1. Determine the next product to buy. Most companies have too many products for customers to easily manage and browse. Giving them a ranked list, optimized for their tastes, will make it easier for them to buy something new.
  2. Lead lists for outbound campaigns. Similarly, if you are a company with a large number of customers, choosing which to contact or e-mail can be a costly decision. Product recommenders let you rank which customers are likely to buy a certain product, so you can target your sales efforts more effectively.

One of the most well-known forms of product recommendation is called “collaborative filtering”. In one form of this filtering, customer purchase histories are compared to each other. Those with similar purchases are then grouped together and products that the group likes, but which a certain individual in the group has yet to buy, will be recommended.

As an example, suppose you are someone who has previously watched all Star Wars and Indiana Jones movies, but nothing else. A collaborative filter could see that others who have similar patterns have also watched the lesser-known “American Graffiti”, and this would then be recommended to you.

Best-in-class recommendation engines

Most product recommenders, particularly for smaller companies, suffer from what is known as the “cold start” problem. If you have too few recommendations or if a customer has never made a purchase, how do you know what to recommend? More importantly, once they start buying, how do you make the recommendation as accurate as possible?

At Canopy Labs, we approach the problem through “cross-domain” recommendations. Rather than taking past purchases by themselves, we look at all aspects of a customer’s behavior to better understand what they are like and make recommendations. Here is the data we take into account:

  • Email activity. Imagine that every email communication you send your customers were also a product. Those who open e-mails and click links are implicitly showing their preferences for content and offers. We use this to group people in a similar way we group through purchases.
  • Website browsing. As with emails, people browsing a website are implicitly making a decision to follow certain links, content, or ideas. We track this data to segment customers and also look at individual decisions for product recommendations.
  • Demographics. If all else fails, we group customers by demographics, geography, and related statistics. In this way, even people who do not browse your website or read your emails can still get more personalized recommendations.
  • Of course, past purchases. Finally, we take into account the actual purchases customers have made, just like most other recommendation systems.

Depending on industry and size of the company, the way you take into account the information above differs. For example, North American fashion retailers tend to benefit from past purchases and email activity — possibly because fashion trends are somewhat similar across North America. On the other hand, local commerce sites (e.g., daily deals websites, discount sites, etc.) often require demographic information to be taken into account for their product recommenders. This is likely because many of their products are location-specific and not available in all regions.

Find out more about how to personalize your emails.