Choosing the Number of Segments
You'll learn how to balance statistical metrics with business needs when deciding how many customer segments to create.
The Core Challenge
After engineering features and picking a clustering algorithm, you face a crucial decision: how many segments should you create? Too few, and you miss meaningful differences between customer groups. Too many, and your marketing team can't act on them—plus you risk overfitting noise in your data.
Statistical Approaches
The elbow method plots within-cluster variance (or inertia) against number of clusters. You look for the "elbow"—where adding more clusters yields diminishing returns. Think of it like buying insurance: the first few policies reduce risk dramatically, but the 20th policy barely helps.
Silhouette scores (ranging from -1 to 1) measure how well each point fits its assigned cluster versus others. Higher average scores suggest better-defined, more separated segments. A score near 0.5 or above is typically good; below 0.2 suggests weak segmentation.
The Business Reality Check
Statistics guide you, but business interpretability decides. Ask:
- Can we describe each segment in plain language? ("Price-sensitive bargain hunters" vs. "Cluster 7")
- Can teams act differently for each segment? Three email campaigns are manageable; fifteen are not
- Do segments align with existing business logic? Fighting organizational structure creates friction
- Is the segment large enough to matter? A 2% segment may not justify custom treatment
Most organizations function well with 3-7 segments. Fewer than 3 rarely justifies the effort; more than 7 overwhelms execution capacity.
Key Takeaway: Choose the number of segments where statistical quality (elbow point, silhouette score) meets business practicality—segments your organization can understand, resource, and activate differently.