5 Practical AI Techniques for Improving Customer ROI

Modeling user actions on your website or app enables you to predict behaviors and attributes that are correlated with near term actions, like purchases or churn.

For instance, an e-commerce website with a 5% session conversion rate has 95% of all sessions abandoned. But all abandoned sessions are not alike. And it’s imperative that marketers understand the true composition of this group. Typically, some are serious, likely buyers, while the rest are casual visitors who are just browsing your site. A predictive engine, can help you build a model that separates your likely buyers – that can often comprise 25% of this group – from the rest. These high intent users are typically more than 2 times more likely to respond to email messages than the average recipient and yield 7-12 times the ROI from paid advertising such as display retargeting.

Unlike user intent (which is near term), user affinity models give you an idea of the user’s persona and lifetime value. It is done by using the concepts of affinity marketing to categorize users into affinity groups based on their demographics, preferences and behavior.  Using these attributes, one segments customers by the product categories and brands they prefer, similarity in attributes to known customers such as preferences for certain authors or their preferred price bands for specific types of products.

Once you know the right set of users to target (e.g., high value users with high purchase intent), you need to understand the right content or product selection for each user. Techniques such as collaborative filtering, which makes predictions about a customer’s preferences based on the preferences of similar customers, and unsupervised clustering, which uses data analysis algorithms to find “hidden” patterns or groupings in customer data sets, can help determine the right set of products for each user.

Since promotions directly impact the bottom line, you should not only model who will be receptive to the promotions, but also drive a change in a user’s behavior by offering the promotion. The former is known as affinity modeling or response modeling, while the latter is known as uplift modeling. In uplift modeling, you try to find users who would not have transacted with you without an offer and—from among these users—find the ones who have a high likelihood of responding to your offer.

Traditional A/B testing models offer a quick path to finding an initial set of champions among creatives and data science models. You can go beyond that by using Bayesian optimization algorithms that test a new set of challengers against current winners. Auto optimization platforms can do this every time a new variant is added to the system by running these tests automatically to yield results that optimize return while minimizing variability. This is similar to the concept of efficient frontier in portfolio theory.