You have a major segmentation project to execute and are beginning to panic. Everything you need cannot be learned from one article, but there are a few important steps that can significantly boost your chances of success. These are some of my favorite experience-based nuggets, and my hope is that you can benefit from them too. Without further ado let’s get started.
Step 1: Questions should be actionable
Segmentation survey questionnaires are always long (sometimes exceeding 30 minutes). How much useful information can a respondent provide when the survey is so long, and often repetitive and boring? If all of that information is important in some way, the length of the questionnaire may be a reasonable compromise. But too often we see questions that few companies can be expected to act upon. I’m looking at you “Easy to do business with”. And yes, you “Company I can trust.” Who, exactly, wants to work with companies that are hard to do business with and are not trustworthy? If ease of doing business and trustworthiness are important (which is different from treating them as givens), can they not at least be measured in an actionable manner? Perhaps have people trade-off ease of doing business and what they are willing to pay, or give up, in terms of additional bells and whistles? Does trustworthiness really belong in the questionnaire, especially for a financial services or healthcare company?
In developing a questionnaire for segmentation research, the client and the researcher need to look at every question and think about what action (short/long term, direct/indirect) they can take if that question turns up as an important differentiator between segments. If no good use comes to mind, why trouble the respondent with it?
Step 2: Variables should vary
It sounds simple enough, and even redundant. But it fails to happen in segmentation studies more so than in other types of studies. We want questions to discriminate between people. That is how we can tell when people are different from others, and that in turn makes it easier to group them into segments. When everyone answers the same way we have constants, not variables. Constants are not useful in analysis. Good questions (in content and form) make consumers think and provide a discriminating answer to reveal their true perceptions. When attention is not paid to the question wording or if scales are used gratuitously, responses are likely to be invariant, resulting in segments that do little more than separate people into response-tendency based groups (that is, those who answer high, medium or low to all questions). Using trade-off based questions (where appropriate) is a good way to get variation in responses.
Step 3: Use the funnel approach
There is a tendency in segmentation studies to keep adding questions hoping that the kitchen-sink approach will produce segments. This happens because in isolation each question seems interesting. But the reality of segmentation analysis is that fewer questions provide more clarity and discrimination. Think about the extreme case of segmenting around a single question, such as likelihood to recommend the company (the question that is used to calculate Net Promoter Scores). It would be extremely easy to form three segments (say, Promoters, Passives and Detractors) by simply grouping the appropriate scale points. But, are the segments necessarily useful, in that they react differently to marketing efforts and are actionable? Not likely. To get more nuance we can add more variables, but with each additional variable the distinction between the segments becomes murkier as each one needs to be explained using multiple variables.
So there’s a trade-off between depth and clarity. Simply adding more variables and hoping that the segmentation algorithm will magically produce a great solution is not a viable option. Instead, try using the funnel approach. First identify the areas that have an impact on the consumer decision-making process (after all, that is what the marketer is trying to influence). Then select a few questions that best represent each area or bucket. These questions will form the heart of the segmentation. Combinations of these areas/questions are going to be the segments. Think about whether those segments are plausible and actionable. If not, repeat the process. The algorithm can only work with the variables assigned to it. If we can’t envision the areas and their combinations, how can we expect the algorithm to produce segments on its own?
Step 4: Use qualitative before and after
While segmentation studies are usually large scale quantitative efforts, there is an opportunity for qualitative research to do what it does best. One way is more traditional – embarking on the qualitative before the quantitative effort to highlight the major dimensions on which consumers react to a product and help shape the questionnaire. But qualitative can also play an important role on the back end after the segments have been created, named and profiled. This typically meanspersona development to help further understand the segments and to help socialize the results.
We can only understand the segments to the extent the quantitative survey allows us to. To get a good feel for the segment archetypes, it helps to do in-depth research with typical members of each group. This is particularly useful in helping client teams understand the segments as something more than just a paragraph description or set of tables. In so doing, they are more likely to incorporate the segmentation scheme into their everyday activities.
Step 5: Modeling is necessary but not sufficient
Too often segmentation discussions are overly focused on the modeling approach to be used, with claims made for one technique over another. There are dozens of variations available to segment data, but the reality is that many of them are quite good while none of them is perfect. Hence making claims for the absolute superiority of any single technique seems dogmatic.
Think about it this way. Classifying observations that are very similar to each other is easy and can be accomplished by pretty much all algorithms. Questions arise only at the boundaries, where observations are almost equally likely to be in two different segments. But there is no single correct way to classify boundary observations, and they are a small fraction of the total anyway. So, there really isn’t any need to obsess over modeling technique. It is far better to focus on the preamble (asking the right questions, using the right sample, conducting proper pre-analysis), use a robust technique iteratively and then focus on interpretation. Obsessing over whether latent class is the best segmentation technique ever can be left to those with the time and inclination to do that.
Step 6: Collinearity is not just a regression problem
Collinearity, or the high correlation between independent variables, is a well known problem in key driver analysis. But it can also have a strong impact on segmentation analysis. The implicit assumption in segmentation analysis is that the handful of constructs that are the foundation of the analysis are about equally weighted. Each construct (such as price sensitivity) may be represented by one or a few questions.
However, if some of those questions are highly correlated, then they are not really addressing unique dimensions of the construct. They are simply adding more weight to the analysis without adding any variance. In other words, the construct (price sensitivity) is over-represented in the data. This can result in the analysis getting tilted in the direction of that construct and making it appear more important than it otherwise would be. By identifying these redundancies and removing them from the analysis we make the analysis more “fair”.
Step 7: Develop a typing tool
Segmentations are expected to be of strategic importance and to be used for several years by the client organization. One effective use is to classify respondents to new research activities into the existing segmentation scheme. This can be accomplished by using a typing tool (also known as a classification algorithm) that uses a subset of the segmentation questions, rather than making every new respondent go through the entire segmentation survey.
So for example when focus groups need to be conducted, it is a simple matter to have the appropriate questions included in the screener and classify people into their respective segments in real time so the research can proceed accordingly. Depending on the number of questions used in the typing tool, it can also be used on client websites, call centers and other places to allow every interaction to place customers and prospects into the appropriate segments, and hence direct the right marketing efforts at them. Consistent usage of the typing tool ensures that the segmentation scheme remains alive and well in the future research activities of the team.
So there you have it – seven recommendations for a successful segmentation.
These are not the only ones, of course, but these have arisen from my experience with practical segmentation studies, rather than culled from textbooks. Hopefully you will find them useful too. Good luck with your study!
- Fort Washington, Pennsylvania
- Research & analytics firm with expertise in new product research, message optimization, pricing research, conjoint, segmentation, brand equity value.