Identify your most valuable customers and cater to their unique needs.

Retention models help brands target their existing customers more effectively based on a desired outcome, for example the likelihood to grow customer value over time. Whether the marketer wants to cross-sell products to increase basket size or upsell customers to a new product, retention models are important tools for developing targeting strategies that yield the best results.

There are several different retention models; the ones we most frequently develop are:

• Next purchase models predict the likelihood that an existing customer will purchase again in a given timeframe (e.g. the next 3-6 months). Marketers understand how important it is to target the right customers for a return visit, and these models take it a step further by identifying when (or when not) to target. Marketers can use next purchase models for communications containing strong call-to-action messages or purchase incentives to maximize returns.
• Example: we recently implemented a next purchase model for a specialty retailer and achieved 5x the purchase rate for first-time shoppers compared to the audience average..

• Response models predict which customers are likely to respond to certain marketing campaigns. A response can be defined as a purchase, coupon redemption, product activation, or even an online form submission. To ensure these models work well across marketing channels, the model data samples should include cross-channel responses.
  Example: We developed a response model for one of our retail clients specifically for the Q4 holiday season. We used data from previous holiday seasons and generated double the average response rate, 2.3 times the average campaign revenue and 12% higher AOV

• Uplift models predict which customers are more likely to require or need a marketing stimulus (communication or incentive) to respond or purchase. While response models score all customers based on their likelihood to respond or purchase, uplift models consider an additional factor – i.e., is a stimulus needed to prompt the purchase. Uplift models deprioritize customers who are likely to purchase even without a marketing stimulus and prioritize customers who are most likely to be influenced because of a marketing stimulus.
  Example: Applying an uplift model for one of our retail clients generated 3.3 times the average program ROI.

Retention models utilize historical first-party data to predict future behavior. As a refresher, first party data refers to the data collected by the marketer about their existing audience such as email activity, purchase history, and browse behavior. Generally, we recommend using enough purchase data (e.g. sales amounts, number of transactions, transaction dates, items purchased) to cover multiple purchase cycles. If purchase data is not available, retention models can be built using other first-party data sources, including product registration, coupon redemption, form submissions, and consumer engagement data.

To make retention models more accurate and effective, we recommend supplementing first party data with additional consumer data such as age, gender, presence of children, income, interests, employee size, etc. This will help us build more comprehensive models.

One pitfall is marketers tend to target programs too narrowly. Marketers often focus marketing initiatives on high-scoring consumers, while excluding low-ranked ones. Instead, marketers should test the model and initiatives across a broader base of consumers for two reasons:

1. To ensure the number of responders or purchases can scale as the program matures.
2. To quantify the true monetary benefit of the model. Over time, retention rates and customer values can be improved across the board, not just for the high-scoring segments. To do that, marketers can test various tactics (channel mix, tiered incentives and dynamic messaging) for lower-scoring segments.
   Example: Consumers who are not at the top tier of the retention model may still respond to your brand but in other marketing channels or if you incorporate a stronger offer. In this case, you want to still include the “best of the rest” segment in your targeting efforts but exclude low-scoring segments that might be a waste of resources.

A second pitfall is marketers’ tendency to overwrite scores as they are updated over time. Maintaining historic scores helps identify pockets of opportunity to target segments when they are likely to respond. For example, if a consumer’s score increases dramatically, he or she may be ready for a next purchase through a more aggressive call-to-action or a (re)activation campaign.

We generally expect retention models to perform well for 18-24 months after initial development. The timing may vary, but model performance tends to deteriorate over time due to:

1. Changes in product merchandising. As brands launch new products and services, they may observe an early adopter period or an influx of new types of customers. A new product or service may be a frequent add-on to an existing product thereby causing an artificial increase in basket size versus historical norms.
2. Changes in customer base or landscape. A brand’s audience composition may change as the brand increasingly penetrates certain markets or expands beyond its current footprint. Alternatively, as customers age but remain loyal, their needs and shopping habits evolve.
3. Fluctuations in economic factors. These may influence consumer spending habits, particularly related to discretionary spending on non-essential products

Regardless of the scenario, it’s best to measure score changes over time so brands can make necessary adjustments and proactively plan for model re-development.