Prevent churn and win back customers

A customer's patterns of engagement with your brand determine their likelihood to remain your customer. By identifying customers who are not following normal patterns, you can try to save the relationship. Optimizely Data Platforms's (ODP's) customer winback model shows the relationship between days since engagement and likelihood to remain your customer, so that you can find customers before they churn. ODP evaluates both new shoppers and previous purchasers.

The major inputs into the model are event days for your brand's customers. Each day is a "hit" if the customer actively engaged with the brand that day or "miss" if they did not. The model understands your brand's patterns of engagement better as you integrate more sources of data into ODP.

The model scores customers on their pattern of events to determine the most likely path they will follow between days since last engagement and probability to remain a customer. It then assigns the customer a probability of remaining your customer, placing them in the Engaged, Winback, or Churned zone. For more detailed information about the model, see For the data scientists at the end of this article.

The cone plot shows the relationship between time since last engagement and likelihood to remain your customer for your average customer. Each customer takes a unique path through this cone based on their own patterns of engagement:

The cone shows your middle 60% of customers. Twenty-percent of customers may fade more quickly than shown (have a steeper drop), 20% average fade rate, while 20% may fade less quickly than shown.

The cone also includes areas denoted as Engaged, Winback, and Churned, specific to your brand's data.

The histogram shows the distribution of your customers by their likelihood to remain a customer. Your known customer base displays, excluding anonymous users. Build new audiences or evaluate existing audiences by clicking + Segment to apply them.

The histogram helps you understand the composition of audiences. For example, if your newsletter audience skews Churned, it could impact your open rates and deliverability. If you have high-value customers slipping into Winback, you may want to approach your relationship with them uniquely.

An unscored customer does not have a value for the customer zone (Engaged, Winback, and Churned). Unscored means a customer has only had a single day of activity and has made no purchases. Each analysis indicates the number of unscored customers. This percentage varies if you apply different audiences to the Distribution of Customer Likelihood chart.

Examine the unscored percentage because these customers may not be a good fit for your brand and can likewise impact open rates and deliverability in campaigns. These customers can also consume marketing resources in other forms of targeting (like audience sync). Consider excluding customers who are known for a while but have no value for the Winback Zone attribute.

Use the outputs of Optimizely's customer winback in two ways:

The winback campaign configuration includes a specially-targeted audience and an email campaign that sends three touchpoints over two weeks. Click Install Winback Campaign Recipe to install a pre-built campaign. 

The audience shows customers who have crossed into the winback zone each day, which is the optimal time to re-engage. A new set of customers cross this threshold each day. To use the campaign, follow the actions below:

You can also use winback insights directly from the Audience Builder to create your own campaigns, analyses, and audience sync.

Winback Zone uses the values of Engaged*, Winback,* and Churned, based on the customer's probability to remain a customer. Winback Type uses the value of Is Any Value when the customer has just been added to the winback zone the last time the model was run (overnight). The customer loses that value the next day. This value helps you target only customers new to the winback zone in ongoing campaigns.

The model starts with the Pareto NBD (negative binomial distribution) model, which is a well-accepted customer lifetime value model. While the original model predicts purchase events, the ODP model predicts any kind of engagement. This model uses the non-subscription version of the Pareto NBD, essentially assuming that future interaction is not unduly influenced by a past interaction, such as in a subscription model. ODP validated that model is predictive of future interaction by making historical predictions and then looking at subsequent re-engagement. 

The winback opportunity starts at the measure of live-ness where increasing live-ness has a larger corresponding increase on future engagement; this starting point is different for each client. The winback opportunity ends and a customer is considered churned when increasing liveness has a smaller corresponding increase on future engagement. Each section (Engaged, Winback, and Churned) has an aggregate likelihood of re-engagement with all activity held constant.