Building Segments for Digital Marketing

Smart segments and personas are wingmen in the flight toward personalized customer experiences and precise media and content targeting. Customers increasingly expect a personalized experience, yet many marketers feel their segmentation skills are lacking. They struggle to define realistic goals and build a useful data model to support segmentation and personas. It can be challenging for the business analyst to know which statistical approach to use, since there are so many options.

Segmentation is used by marketers for a variety of purposes, including product development, pricing, targeting, messaging and measurement. Marketers increasingly lean on customer experience as a differentiator, according to Gartner’s 2014 survey on the topic. By 2018, Gartner predicts that organizations that excel in personalization will outsell companies that don’t by 20%. Segmentation is also a key step toward meeting consumers’ demands for more relevant experiences. Meanwhile, at least 42% of marketers believe they are not very good at segmentation for digital experiences.

Most segmentations are based on the assumption that people buy products and services because they have definable needs they are trying to satisfy. In reality, these “needs” are often difficult to track. How often have you seen “stingy” or “hedonistic” as fields in a database? Yet these needs may be important to know if you are selling, say, savings accounts, or indulgent treats. Unless the marketer is using direct survey data, he or she often must employ other attributes – such as behaviors and demographics – as proxies for needs. For example, car dealers often use visits to their websites’ “Hours of Operation” page as indicating a “need” for a test drive or service visit. Such indirect clues are known as implicit data.

So needs are usually better thought of as attributes that influence decisions. How does the marketer know which needs are more important than others in influencing decisions? It is possible to group attributes with common themes using a statistical technique called factor analysis, which can reduce complexity by detecting underlying patterns across different variables. However, the most common practical criteria is how much business value the attribute influences. Business value can be defined as customer lifetime value (CLTV), average order value (AOV), basket size, or any other relevant available metric that points to the customer’s importance to the company.

Needs are often combined with business value, leaving two basic marketers’ approaches to segmentation:
•    Needs/Value methods use value as a target to build a hierarchy of needs, starting with the most important, that can be turned into a tree or dendrogram to describe segments.
•    Clustering looks at similarities in the attributes themselves and builds segments that can be thought of as “tribes” or “groups” of people with some similar features.

What is the “right” number of segments? There is a popular perception that 5-8 segments is ideal. In fact, there are statistical methods to determine the optimal number, such as latent class and two-step cluster analysis. There is no theoretical limit: a marketer can have as many segments as are statistically different and useful. The key term here is useful. Limits on data, resources and systems will set practical limits on the number of segments you want. In fact, it is always a good idea to go into a segmentation exercise with a rough idea of a maximum number your organization can actually use.

What does a needs/value model look like? Let’s say an airline wants to segment its customers. It has decided that the scope of the exercise is its own customer data and the goal is to determine the key factors (or needs) that make a person a profitable customer. Because the marketer herself is setting the target variable (value), the goal of her analysis is to determine which combinations of attributes have the greatest impact on value. Because the output is a list of influential attributes, ranked in importance, it is often helpful to express it as a branching tree.

A record is assembled for each customer who has taken at least one flight in the past 12 months. All data about that customer is pulled into a data warehouse or other data container, and key fields are defined based on factors that have been useful in the past.

The airline starts by determining which attribute has the greatest impact on the success metric, which is business value. This attribute becomes the first node of the tree, and its range of values are the branches. The model supports both categorical (e.g., male/female) and continuous (e.g., age) values, which can be put into bands which represent branches of the tree. Statistical details can get complex fast, but in general, there are three approaches:
•    Manual – customers are sorted from most to least valuable and the marketer uses experience and simple tools, such as averages and heatmaps, to arrange a list of attributes correlated with value.
•    Decision trees – these common models are a form of regression analysis, formed using a top-down algorithm that aims to minimize the standard deviation of the attributes.
•    CHAID – a form of decision tree, this method uses a reliable target value as well as some human judgment to “sanity check” the branches and the drivers.

Whichever method is used, the marketer will end up with an ordered tree of nodes (attributes) and branches (values), which are arranged from more to less strong influencers of value.

What does a hard-working marketer do next?

For that you’ll have to read my recent research note, “How To Build Segments and Personas for Digital Marketing”. It is, ahem, the most downloaded piece of research in our group for three weeks running.