Why Not All Recommendation Engines Are Created Equal

Machine Learning|2 Minute Read

Why Not All Recommendation Engines Are Created Equal

Many brands are investing in tools to drive personalized information to existing and potential customers, yet those efforts often fall short. According to Adobe, 60% of marketers struggle to personalize content in real time, yet 77% believe real-time personalization is crucial (source).

Most marketers are providing recommendations based on the content or products that receive the most clicks or views across their entire audience or very large segments of their total audience, rather than making recommendations based on each individual’s interests. It’s the difference between Amazon’s pretty bad recommendations, and Netflix recommendations, which are frequently cited as a the gold standard for tailored suggestions. (Don’t even get us started on the almost ubiquitous “around the web” recommendations.)

A really good recommendations engine’s algorithms deliver results against business priorities based on two inputs:

  • Individual customer attributes – including her interests, preferences, and behaviors
  • Inventory – based on available content/product inventory and key areas of focus

Traditional recommendations engines don’t have access to all of the first-party data required to use all of a customer’s attributes as inputs, and marketing databases don’t have the delivery capability to feed all of the data to a recommendations engine in real time. That’s why BlueConic offers a solution that can do both. Using machine learning and proprietary algorithms, BlueConic’s Recommendations Engine combines individual profile data to deliver recommendations that not only align with the customer’s interests, preferences, and behaviors, but also align with your content or product priorities.

When using BlueConic’s Recommendations Engine,

You can read all about how our recommendations engine works here, but one of the engineers on our data science team wrote a deep dive into the nitty gritty algorithm-level details powering the engine that explains how to continually improve the performance of recommendations, in three phases:

  1. Exploration Phase

It’s important to try at least a handful of different variants with different algorithms to quickly test and learn what is the ideal algorithm combination to start with.

  1. Exploitation Phase

During the exploitation phase, use what you learned during the exploration phase to get the maximum amount of value out of BlueConic’s Recommendations Engine. This means turning off all variants except the best performing one.

  1. Continuous Improvement Phase

Once you’ve figured out the best algorithm combination across a large audience, it’s time to start thinking about testing algorithm combinations for specific audience segments, and then repeating the process again.

If you want to start delivering recommendations that:

  • account for individual interests/preferences AND your inventory/business priorities in their output
  • offer the flexibility to appear at the right moment without specific placement restrictions
  • support different algorithm combinations for different placements
  • give you (as a marketer) full control over the algorithms to test and optimize as often as you want

…then let’s talk.

Example content algorithms Example product algorithms
Content Recommendations Product Recommendations

See what BlueConic can do for you

Whether you’re looking for operational efficiencies or improved marketing effectiveness through data activation, our customer data platform can help.