Customer Segmentation Survey Guide — Dividing Customers with Cluster Analysis

Key driver analysis told you "support is what moves overall satisfaction." Importance-performance analysis decided "support and price are the priorities to fix." But those conclusions carry a hidden assumption — the assumption that customers are a single, uniform mass.

In reality, what drives satisfaction for a price-sensitive new customer is completely different from what drives it for a heavy user who wants depth of functionality. When the overall average says "support matters," that result is the blend of two distinct customer groups averaged together, and it's accurate for neither. Statisticians call this the "the average customer doesn't exist" problem. Customer segmentation surveys are how you solve it: find the "clusters of similar people" in your survey data and change strategy cluster by cluster. This guide works through the two major approaches to segmentation, how to choose your segmentation axes, the cluster analysis methods, how to decide the number of segments, and the "criteria for a usable segment" — all with the hands-on feel of practice.

1. Why Segment — the Trap of the Overall Average

Both key driver analysis and importance-performance analysis (IPA) are powerful, but they share one weakness: they average the entire respondent pool as a single population.

When you blend heterogeneous customers and average them, you get something close to Simpson's paradox. "Support is the biggest driver of satisfaction overall," yet price is the biggest driver in Segment A and functionality is the biggest in Segment B — this happens all the time. If you make decisions on the overall average alone, you end up investing in average, mediocre initiatives that land with no one.

The goal of segmentation is simple: split customers into "clusters that respond similarly" and tailor the optimal move to each cluster. Ever since the marketing classic Smith (1956) proposed "market segmentation" in contrast to "product differentiation," segmentation has remained a foundation of marketing strategy.

When you re-run KDA / IPA segment by segment, the "different drivers within each segment that the overall view hid" come into focus. Segmentation is the final piece of the analytical trilogy (identify the factors → set priorities → classify the customers).

2. A Priori vs. Post-hoc — Two Approaches

There are two fundamentally different ways to divide customers. Confuse them and you'll get the design of your analysis wrong.

A priori (descriptor-based)

The analyst splits customers mechanically by criteria decided in advance — by age band, by contract plan, by usage frequency, and so on. The criteria are clear, it's easy to operate, and anyone can reproduce it.

The weakness: there's no guarantee those criteria actually separate customer behavior. "We split 20-somethings from 30-somethings, but it turned out their purchasing behavior was nearly identical" is a common outcome. You feel like you've divided the market, but nothing about your initiatives changes.

Post-hoc (cluster-based)

This method discovers "natural clusters" in a data-driven way from response patterns — satisfaction, needs, values, and the like. It uses cluster analysis or latent class analysis. It's a "let the data speak" approach, and it surfaces segments you'd never spot through prior assumptions.

The weakness: results vary from run to run, interpretation is hard, and it's tough to reproduce operationally. It takes skill to interpret "who exactly is this segment."

Principles for choosing between them

• Start with a priori to get a read: split by demographic and behavioral data, and see whether KDA/IPA produce differences across the splits
• If no difference shows up, or you want to dig deeper, go post-hoc: use cluster analysis to find clusters based on needs and values
• In practice, combine the two: take the clusters found post-hoc and "profile" them using a priori variables (age, plan) to pin down their identity (see Section 6)

3. Choosing Your Segmentation Axes — Four Types of Variables

What do you divide on? Segmentation variables fall broadly into four types, and there's a trade-off between "ease of dividing" and "ease of connecting to action."

The practical playbook

• The combination of behavioral + needs is the most likely to produce "usable" segments
• Use demographics as a profiling axis, not a segmentation axis (describe afterward: "this segment skews toward people in their thirties")
• To measure needs and benefits, Likert-scale question design is the key. See the Likert scale design guide

4. Cluster Analysis Methods — Hierarchical, k-means, Latent Class

Cluster analysis is the core of post-hoc segmentation. The three representative methods each suit a different situation.

Hierarchical clustering

You merge samples one at a time and read the structure of the clusters from a tree diagram (a dendrogram). The advantages: you don't have to decide the number of segments in advance, and you can grasp the structure visually. Ward's method is a popular choice. The weakness is that it's computationally heavy and gets slow once samples exceed a few thousand. It suits small-to-medium samples and the exploratory stage.

k-means

You specify the number of segments k up front, assign each point to its nearest centroid, update the centroids, and repeat. It's fast even on large data and the most widely used. The weaknesses: (1) you have to decide k in advance; (2) it's sensitive to starting values, so results wobble (run it multiple times and check stability); (3) it's sensitive to variable scale (always standardize before feeding it in).

Latent class analysis (LCA)

A statistical model that assumes "each respondent belongs to one of the latent classes probabilistically." It has developed in marketing ever since Kamakura & Russell (1989). Its advantages: you can choose the number of segments by a statistical criterion (such as BIC), and it handles categorical variables naturally. The weakness is that it's specialized and requires dedicated software (Latent GOLD, R's poLCA, and the like).

Preprocessing: compress dimensions with factor analysis

When you have 20 or 30 questions, feeding them straight into cluster analysis causes correlated questions to double-weight the same underlying concept. The standard move is to first compress them with factor analysis into factors like "price orientation" and "quality orientation," then cluster on the factor scores. The survey reliability and validity guide covers the relationship between factor analysis and constructs.

5. How to Decide the Number of Segments — Statistical Indices and Interpretability

"How many do we split into?" is segmentation's biggest headache. You decide using both statistical indices and business interpretability.

Statistical rules of thumb

• Elbow method: as you increase the number of clusters, pick the "elbow" point where the decline in the within-cluster sum of squares levels off
• Silhouette coefficient: evaluates, on a scale of -1 to 1, how well each point fits its own cluster and how far it sits from the neighboring cluster. It's the index from Rousseeuw (1987), and closer to 1 is better
• For latent class analysis, BIC / AIC: choose the number of clusters that minimizes the information criterion

But the final call is "interpretability"

Even when the statistical index says "six segments is optimal," it means nothing if you can't put six into words. In practice, things usually settle into 3 to 6 segments. The reason is simple: you can't tailor distinct initiatives to seven or more (the organization doesn't have the resources).

"Don't adopt a number of segments you can't act on in the business, even if it's statistically optimal." That's the iron rule of deciding segment count. Statistics merely present candidates; the final judgment is "can we deploy a different move for each of these segments?"

6. Profiling and Personas — the Six Criteria for a Usable Segment

Once the clusters emerge, you describe who each one is (profiling). For each segment, compute the mean values of demographics, behavior, and needs, put it into words like "this segment is price-driven, skews toward people in their thirties, and is a low-frequency new-customer group," and turn it into a persona if needed.

But not every statistically distinct cluster is a "usable segment." Check whether it meets the criteria for a segment that's usable in practice, as organized by Kotler.

• Measurable: the segment's size and characteristics can be measured
• Accessible: you can reach the segment through advertising or sales
• Substantial: it's large enough to justify the investment (you can't run a dedicated initiative for a 1% segment)
• Differentiable: it responds in a clearly different way from other segments
• Actionable: you can design and execute concrete initiatives for the segment
• Stable: the segment won't vanish in the short term — it's stable over time

A segment that "split cleanly in statistical terms but has no means of reach and is too small" may be analytically correct, but it's useless to the business. At the profiling stage, sift each candidate through these six criteria.

7. The Editor's View — Five Things You Must Not Do in Segmentation

Following industry cases and the voices of practitioners on an ongoing basis, here are five accidents that keep happening in segmentation.

1. Clustering without standardizing the variables

The most frequent — and least noticed — accident. Feed "income (in thousands, ranging from the hundreds to the thousands)" and "satisfaction (1 to 5)" into k-means without standardizing, and the clusters get decided by income alone because of its larger scale, while satisfaction is all but ignored. Standardize every variable (convert to z-scores) before clustering. An analysis that skips this is almost certainly wrong.

2. Splitting on demographics alone and calling it a day

You stop at "we split by 20s / 30s / 40s." If purchasing behavior is the same across age bands, that's not segmentation — it's just a tabulation. Demographics are a profiling axis, not a segmentation axis. Keep the order: divide on behavior and needs, then describe with demographics.

3. Deciding segment count on statistical indices alone

You adopt eight segments because the silhouette coefficient peaks there, then the organization can't deploy distinct initiatives and the segments get left to rot. Cap it at "the number you can actually act on differently," and let statistics choose the optimum within that range. Three to six is the realistic landing spot.

4. Using a segment scheme forever once you've built it

You keep using segments you built two years ago, even now that the market has shifted. Segments are perishable goods. When the market, the customers, or the product change, so do the clusters. Re-run the clustering periodically (roughly once a year) and verify the segments' stability (criterion 6).

5. Slicing into segments when the sample is small

Split N=150 into six segments and you average 25 people per segment. The per-segment scores fill with error, and saying "Segment A has high satisfaction" means nothing at N=20. If you're going to build on segmentation, design the sample so you can secure at least 50 to 100 per segment, ideally 100 or more each. See how to determine the necessary sample size.

8. Customer Segmentation Surveys with the Kicue Survey Tool

A segmentation survey splits into a "measure the questions that feed the classification" phase and a "find the clusters with cluster analysis" analysis phase. Kicue covers the former; the latter is a combination with external statistical tools.

• Measuring classification variables: supports Likert-scale and single/multiple-choice question design for measuring needs, values, and behavior (question types)
• Pairing in demographic and behavioral questions: capture the attributes used for profiling (age band, plan, usage frequency) in the same form
• CSV export with respondent IDs: exports in a one-row-per-response structure with every question laid out, ready to feed directly into cluster analysis. After the analysis, you can re-join "which respondent is in which segment" back into your CRM
• GT and cross-tabulation: cross-tabulation for a priori segmentation (by age band and the like) is available on the dashboard

⚠️ What Kicue Cannot Do

• No cluster analysis, k-means, hierarchical clustering, or latent class analysis features: run the statistical analysis in R (cluster, poLCA, and so on) / Python (scikit-learn) / SPSS / Latent GOLD. Kicue itself has no statistical analysis features
• No factor analysis or variable standardization either: clustering preprocessing happens on the statistical-software side after export
• No per-segment driver analysis (KDA) either: hand the CSV to an external tool and run it segment by segment
• No automatic persona generation either: turning profiling results into personas is done by hand plus a BI tool

For related reading, the key driver analysis guide, the importance-performance analysis (IPA) guide, the survey sampling methods guide, the screening question design and operation guide, and the survey reliability and validity guide together bring the whole analytical pipeline into view — "design → classify → per-segment factor analysis → prioritize."

Summary — Six Points to Make Customer Segmentation a Usable Analysis

1. Stay alert to the trap of the overall average — re-run KDA/IPA segment by segment and the hidden differences appear
2. Divide on behavior and needs, describe with demographics — splitting on demographics alone ends in "feeling like you've divided"
3. Always standardize before clustering — don't let large-scale variables hijack the clusters
4. Cap segment count at "the number you can act on differently" — let statistical indices choose the optimum within that range (3 to 6 is the realistic answer)
5. Sift with the six criteria (measurable, accessible, substantial, differentiable, actionable, stable) — a statistical cluster ≠ a usable segment
6. If you're building on segments, secure 100 or more each — slicing a small sample into pieces fills it with error

Customer segmentation isn't about "running a sophisticated cluster analysis" for its own sake. By not missing the three points of standardization, interpretability, and actionability, it becomes the analysis that anchors your strategy — letting you escape the illusion of "the average customer" and design moves that land cluster by cluster.

If you want to design the survey that feeds your segmentation, why not try the free survey tool Kicue? With Likert/choice question design for measuring needs, behavior, and attributes, plus CSV export with respondent IDs, you can get the input-data-building part of cluster analysis started on a single account (cluster analysis, factor analysis, latent class analysis, and variable standardization are run in combination with R / Python / SPSS / Latent GOLD).