How a Simple Recommendation Engine Boosts Revenue over 70%

David Jones · Aug 18, 2014 · 2 min read · Share:

United Cellars is an award-winning online wine retailer based in Australia. Their site is built on the Spree Commerce platform and supported by a well-trained telemarketing team. We proposed that a recommendation engine would improve the visitor experience and increase conversions, so we set up an A/B test to find out.

The site currently has over:

  • 16,000 product views
  • 60,000 orders
  • 3,000 reviews

With this valuable data we knew we could make quality product recommendations using Machine Learning algorithms provided by Prediction.IO - an open source prediction server.

Getting started with a recommendation engine

Because we have a long history with United we were already familiar with their business goals and customer behaviors. While everybody appreciates a good deal, price is not the main motivator for United’s best customers. They’re looking for wine they haven’t tried before that’s similar in style to vintages they’ve already purchased.

With this in mind, we set up a recommendation engine and A/B testing so we could compare the results to the baseline. We ran the test until we had statistically significant data.

We spent one week getting everything set up and running smoothly.

  • Prediction.IO server setup on Amazon Web Services
  • Push ratings, orders, and product views data to Prediction.IO server
  • Connect Prediction.IO server to Ruby Prediction.IO SDK running in the store
  • Retrieve recommended products
  • Display recommended products on the site
  • Set up A/B test

Results of the A/B Test

We ran the first test for 15 days. 50% of the traffic saw the Prediction.IO recommendations; 50% viewed the original site. The test demonstrated

  • 45% longer average session
  • 22% increase in conversion rate
  • 37% increase in average order value
  • 71% increase in revenue

The significant revenue increase is due to a combination of higher conversion rate and bigger average order value. The results were so compelling it was a no-brainer to start showing the recommendations to all visitors.

More optimization opportunities

In addition to on-site product recommendations, Prediction.IO unlocks many more possibilities. To build on their initial success, United is considering

  • A tool for the telemarketing staff to offer customers personalized suggestions over the phone
  • Email marketing campaigns that send individualized recommendations to each recipient
  • Applying the recommendation engine to customize search results

We hope this glimpse into the potential benefits of Machine Learning will inspire you to think about your site is a different way.


"How a Simple Recommendation Engine Boosts Revenue over 70%" was written by David Jones.


Join The Conversation

Share and start a conversation about this post: