Cohort Segmentation at Direct-To-Consumer Scale
Your customers aren’t rows in a database, they’re people. Too often businesses make the mistake of thinking of their customers as a homogeneous set of participants in day-to-day business operations. It should come as no surprise that some customers are enormously more expensive than others — and as a corollary — some customers are significantly more profitable to acquire than others.
Companies that operate in other spaces have the luxury of asking questions about whether or not a particular customer or account will be profitable for their business. And in doing so, the company implicitly leverages a fundamental understanding of who their customers are and how they behave in order to optimize profitability.
That’s not a revelation. But it’s also not that easy in the direct-to-consumer space in which many subscription based e-commerce businesses operate. Why? Because it’s complex analysis at consumer scale.
The simplifying mantra of marketing has been something akin to “cast your nets wide” and see what Facebook Ads and other traffic networks can dredge up in terms of qualified traffic. And that’s a bit of a cop-out.
So let’s just say that you’re the type to push the envelope a bit. You know that business rewards the intrepid and you’re going to make sure that your name is in the running when it’s time to determine who gets that multi-million dollar subscription payday (every month.)
Subscription-based businesses are extraordinarily sensitive to retention rates.
In earlier posts, we demonstrated that outside of making sure that you have healthy profit margins on individual subscription fulfillments, it’s customer retention rates that matter in terms of expected customer value. And it’s not some modest influence: Retained customers are synonymous with profitability.
Not to understate the issue or the importance of retention rates: a 5% increase in retention can mean a 75% increase in gross revenue.
This becomes particularly noteworthy as subscription based businesses tend to extend themselves financially early into the customer lifecycle and amortize those costs against later, more profitable periods. In fact, the “trial-based” subscription vertical is distinguished by — if nothing else — this particular type of yield curve.
Discussion about understanding your customer base in order to increase overall profitability should pay particular attention to identifying which subsets (or cohorts) of your existing customer base — if any — are demonstrating statistically significant deviations from mean retention rates.
In short, ask the following question: Can we better understand our customer base and in doing so can we optimize business spend in order to prioritize our highest-value customers?
Distinguishing the apples from the oranges
Statistics are wonderful when they confirm our “gut-level” intuitions about phenomena present in datasets, but they serve us best when they provide something unexpected. For instance, if the result of your analysis explicitly contradicts assumptions you may have had about the data, while it may initially cause some pause and/or beg confirmation, it’ll also come packaged with a sense of discovery.
Developing robust, multi-dimensional customer cohorts can provide that same sense of discovery for your business. An while the kiddos at Wednesday-night yoga may remind you that “the journey is the destination”, the destination in this case is actionable information that’ll improve your bottom-line; Well substantiated cohort segmentation will guide everyday decisions across business groups.
Namaste, though.
The first step is really just to build and consistent customer dataset as the foundation for evaluation. In early iterations, lean towards inclusiveness in your dataset lieu of a clear understanding of what fields should or should not be present for evaluation. If you have guesses and suspicions, follow your nose. Otherwise, common properties like the geographic location of the customer, their age and gender, income level, education level, their particular acquisition channel, and the length of time that they have been a customer would all be perfectly relevant in an analysis of this nature.
At the same time note that a part of the complexity of the exercise is first and foremost building a representative and consistent customer dataset.
Secondarily, the complexities and nuances of the statistical methods you’ll implement need to be provided some appropriate measure of diligence, especially while constructing the customer dataset; In particular, some clustering techniques are more complex than others (please see: NP Hard) and should be treated with due consideration.
That said, you’ll certainly want to account for all the available metrics that represent “actionable” or “controllable” properties within your marketing plan. If your marketing team can target geographic areas, don’t skip including your customer’s geographic location in the dataset.
As you get better at analysis of this nature, you’ll develop tools to help you understand what parameters are most influential on particular response variables, and which can simply be omitted without consequence.
Cluster Analysis and Cohort Centroids
*If you haven’t heard the term “cohort” before, it’s just fancy industry-speak for a group of people who share some common characteristic.
K-means clustering is a popular algorithm for clustering dimensional data into k distinct groups. At the “center” of the group will exist a “centeroid” — a prototypical customer, if you will — for that cluster. You’ll need to experiment with both the clustering algorithm (there are many) that you use against your customer data as well as parameters of that algorithm.
In a very simple implementation (as seen below) of the k-means clustering algorithm on a 2-dimensional parameter space, we can see that after a handful of iterations the cohort centroids and corresponding boundaries stabilize fairly quickly. These centroids correspond to meaningful and distinct “personas” within your customer base.
For larger, more dimensional data you may choose to explore other algorithms to cluster your data even if for no other reason you become concerned about the performance of the method as your data scales. The good news is that the statistics department at your Alma Mater has been busy and you have lots of options available to you, each of them with particular benefits and possibly (probably) their own idiosyncratic complexities.
Having clustered your data appropriately, you’ll have effectively segmented your customers into k distinct groups across parameter-space. And as such, you’ll gain the ability to calculate relevant group-specific statistical properties, not least of which are centroid retention means. Other inferences might also be immediately available: For instance, if you observe significant under-representation of distinct clusters, it may be an indication that there are opportunities for improved marketing against particular groups.
Act on what you can control and react to things you can’t control.
As we mentioned earlier in this article, there are some properties that you’ll be able to specify in your “marketing flow.” For example, if you’re placing advertisements online and you find that your “educated females from New England” cohort from the above exercise converts particularly well or has an uncommonly high retention rate, you could choose to orient your marketing content and spending toward individuals that categorically fit the persona definition. This is what we’d think of as “acting” against the results of segmentation analysis.
You may also discover that there are influential properties in your data set that are beyond your ability to target. For example, people on your website that convert before their session is one-minute old might be a defining characteristic of a particularly valuable cohort — these are motivated buyers and perhaps their centroid-specific retention rate is abnormally high. The bad news is that there aren’t a lot of options available to your marketing teams in order to select these customers first and foremost; You won’t be able to target behavior metrics like that in your marketing spend. But you certainly can “react” to observed behaviors like these as customers interact with your various brand properties.
So, in order to appropriately react to observed properties of individual customers, group classifier functions that map customers to persona centroids are utilized. You’ll find, again, that daily decisions about individuals may be well-informed by understanding the real value of the customer from a cohort specific perspective.
Pick your customers. Optimize your marketing endeavors. React to customer events.
It’s not always easy to take the leap and try something new — but you wouldn’t be in your shoes if you weren’t accustomed to making sound decisions about opportunities that manifest in the course of your company’s lifetime.
Take a closer look at your customer base. Understand who they are and maybe along the way they’ll better understand your brand along the way.
And don’t hesitate to schedule a 20-minute consultation with us.
