Customer segmentation analysis: groups that change the next offer

If you run a subscription business, your account team may need to help a first repeat purchase or make replenishment easier for established buyers. The analysis must identify customers before the intervention and explain why the offers should differ. The twelve illustrative customer records below let you inspect assignment and evaluation separately.

Build the first grouping from information available now

The twelve customer records below and copyable CSV record orders completed in the previous 90 days, average order value in dollars and whether the customer renewed in the subsequent period. Use the values to reconstruct the analysis, not to forecast renewals.

All twelve customer records

Renewal is coded 1 for renewed and 0 for did not renew. Every row has example_status=fictional. Retain renewal separately from the past-order inputs used to assign the group.

Copy the complete customer CSV

customer_id,orders_last_90_days,average_order_value_usd,renewed_next_period,example_status
F01,1,20,0,fictional
F02,1,30,0,fictional
F03,1,40,1,fictional
F04,1,30,0,fictional
F05,2,25,1,fictional
F06,2,35,0,fictional
F07,3,45,1,fictional
F08,3,35,0,fictional
F09,4,30,1,fictional
F10,4,40,1,fictional
F11,5,50,1,fictional
F12,6,40,0,fictional

The business question is whether an account team should offer replenishment convenience or help with a first repeat purchase. Define the groups using past orders only:

• Starter: one order in 90 days.
• Occasional: two or three orders.
• Frequent: four or more orders.

Average order value describes the groups. Later renewal is a held-out outcome; it is never an input to these rules.

For Starter, the mean is ($20 + $30 + $40 + $30) ÷ 4 = $30. The renewal pattern is a property of this example dataset. Four customers per group cannot establish a dependable commercial response.

Order count is an assignment variable, average order value describes the resulting groups, and later renewal is an evaluation outcome. That separation is crucial. If you create “loyal” customers by including future renewal, then show that loyal customers renew more, the apparent validation is circular.

Keep an input cutoff and a reproducible customer-level record. A group should be assignable when the business needs to act. If guest or marketplace purchases are absent, mark that coverage issue; otherwise Starter may describe tracking rather than the customer’s activity.

Investigate what the Starter label means before activating it

Assignment: one completed order in the prior 90 days; four accounts. Profile: mean order value USD 30; one later renewal. Question: what happened between first delivery and the first repeat opportunity?

Offer choice: compare replenishment help with usage support once timing, unused stock or experience explains the stalled repeat. Owner: the account team, using reachable customer identifiers. Evaluation: subsequent repeat orders in a new period with intervention exposure and comparison conditions recorded. The historical 1/4 renewal count motivates the question; it does not select a proven intervention.

One previous order can mean several different situations: the product has not run out, usage was difficult, the buyer bought elsewhere, or the first occasion was not recurring. These are hypotheses to investigate, not findings from the twelve rows. A replenishment reminder serves only some of them.

Ask about the time between first delivery and the first repeat opportunity, then map the explanation to a feasible intervention. Usage support may be more useful than a reminder when unused stock remains. The analysis makes the recruiting rule explicit; the needs work explains the offer difference.

Keep the existing rule while investigating. Redefining Starter after seeing which people like a reminder would turn the group into the desired result. When you decide to change the rule, save a new version and explain why the information improves assignment.

Evaluate the grouping in a future period

The order-count rule is simple enough to assign without looking at renewal. Save the 90-day input cutoff, each customer’s order count and the group version. The subsequent renewal period must begin after that cutoff; using future information to revise the original assignment would leak the outcome into the grouping.

The renewal gradient of one, two and three renewals out of four does not prove that differentiated offers cause renewal. It describes the later outcome of these groups. To evaluate the intervention, record its exposure and an appropriate comparison among eligible customers.

A segment can be analytically stable but operationally useless if the team cannot reach its members or deliver a meaningful difference. Conversely, a modest rule can earn its place when it helps the account team identify a real task and consistently supply the right support. Decide usefulness at that level.

Use the boundary problem to choose an update rule

Changing Frequent from four to five orders moves two customers, shrinking it from four to two. Look at those boundary accounts. Does their need differ from the remaining frequent buyers, or is the label responding to an arbitrary count? Keep full distributions and the buying cycle in view.

An offer that changes every time an order crosses a cutoff may create a worse customer experience than a stable rule. Consider the purchase window and update cadence the business can operate. A name should describe what the rule measures: “four-plus completed orders” is more defensible than “high-value loyalists” when loyalty has not been defined.

Choose the least complicated rule that supports the action. Complexity is valuable when it adds a need or reachability distinction that a simpler scheme misses, rather than when it produces more colorful segment names.

When clustering is worth commissioning

Use clustering when multiple relevant variables resist a simple rule. Decide how to scale variables, handle missingness and compare alternative group counts. K-means, for example, groups numeric observations around centroids and favors certain geometries; it is not a universal fit for mixed survey and transaction data. See the official scikit-learn clustering documentation for algorithm assumptions.

Fit the grouping on one dataset, document it and inspect a separate period or sample. Do not select the version solely because it tells the most attractive story. A statistical separation score also does not prove that customers need different offers.

Ask the analyst what variables the model uses and why those variables belong to the decision. Review how the variables were scaled and why that treatment fits the decision. Missing values need a declared treatment. Group count and algorithm choice should be inspected against the business task as well as mathematical separation.

Review examples from each group, boundary cases and a later period or held-out sample. Then ask whether the groups can be assigned from data available before the action. A statistical cluster is a starting point for interpretation, not proof that its members need the same message.

The acceptance test for the account team

A usable handoff includes assignment rule or model version, real source base, defining variables, needs evidence, unknown cases, reachable identifiers, candidate action and next evaluation. For Starter, the completed scorecard above directs a first-repeat investigation; for frequent buyers, the corresponding task is replenishment fit.

Accept the grouping when another person can reproduce it, the business can identify members in time and the proposed offer difference has a reason beyond the label. Reject or simplify it when those conditions fail. This is what the analysis must deliver before activation.

Use our segmentation guide and our behavioral segmentation guide, or discuss your segmentation question with Sapien.