Going Beyond Modeling to Find New Customer Acquisition Targets

When it comes to finding “prospects who look like your customers but aren’t your customers yet,” modeling is the most consistently reliable way to find your target. But this is not just another article to advocate the power of statistical modeling. Modeling is about mimicking your target group. That is why setting a proper target definition is the most important step in modeling exercise. If the target is hung in the wrong place, even an experienced shooter won’t be effective. The art of targeting in the 'multiverse' Setting the right target can be tricky because there could be many distinctive segments in a customer base. Let’s say you have over a million customers who purchased your products in past four years. Would you just use random samples of them as the target in your model? That could be a big mistake. What if you have distinctively different types of customers in your base? For example: High-value customers who spend lots of money and purchase often—obvious target, but they could make up a really small segment.High-value customers who do not purchase often.High-frequency customers who are not big spenders.New customers.Recent or active customers.Dormant or inactive customers, regardless of past values. This type of customer composition is quite common. I often profile these basic segments just to see what is going on within the customer base. Most of the time, their behavioral and geodemographic profiles are vastly different as well. What happens if you ignore such differences and lump them all together as “average customers?” Your model could be going after “phantom” targets between the segments. By definition, phantom targets may not even exist in reality—imagine the average of the second and third examples above. Multiple models for hedging There are a few cases in which. most advanced marketers would employ multiple models for targeting: There are multiple segments in the customer base, as illustrated above. Different targets call for different models.You are operating in multiple universes, such as your own customer base and various third-party data sources and their subsegments. In such cases, even if the target is stationary, comparison universes would be vastly different. You’d need multiple models because model algorithms are nothing but equational expressions of differences between target and comparison groups. If the comparison universe changes, so do algorithms.Data scientists may employ multiple modeling methodologies for the given task. Differences among techniques may matter less than changing the target or comparison definitions, but analysts often try to combine multiple methodologies to find the most optimal solutions. In general, I suggest at least two to three different types of models in the initial test stages. It is very unlikely that all of them would fail at the same time, and marketers should test multiple ideas before any big rollout to find the winning combination. The world doesn’t stay still Now, in the initial test stage, I often recommend adding not-so-obvious segments, with or without modeling. The main reason—and the most obvious downside of modeling—is that models by design go after “more of the same.” Algorithms yield high scores for the targets who are most likely to resemble predefined targets. You may have multiple targets in the mix to minimize risks, but what about the prospects who may not fit into myopic target definitions? Further, the world doesn’t remain status quo for the convenience of marketers. Models often lose their effectiveness, or even render completely useless, in the following situations: Rapid changes in market or economic conditions.Significant alterations in the product line or supply chain.Fluctuating pricing models.Shifting product adaptation cycle (from “early adopters” to “trend followers,” for example). These factors may dramatically alter the nature of buyers, and combinations of them may accelerate atrophy of models’ effectiveness. For instance, when a high-end electronics product is introduced at a high price point, it will naturally attract early technology adopters. But as the price eventually goes down, and as the manufacturer aggressively tries to expand its market, the nature of the target universe changes as well. One may initially picture high-income households in single-family housing when pushing for high-end TVs. But what if various profile reports reveal “not-so-dominant” but “significant enough” new segments in the mid- to low-income tiers? You must put them into the test plan as well—because sticking to models that go after “more of the same” will not let you expand the prospecting universe. Go after atypical targets, too When looking for new segments, do not stop short by habitually running customer profiles against the general population; run specific product buyer profiles against your overall customer base, as minute distinctiveness of each buyer group gets revealed that way. Good test strategies must include expansion plans as well. Always test new segments—if not through modeling due to a lack of good samples, do so with rule-based targeting with dominant characteristics (e.g., millennials in condos). If that works out, responders will serve as targets for the next round modeling. After all, target marketing is an endless cycle.