Customer Segmentation and Propensity Scoring
Segmentation that ends in a campaign rather than a slide, with propensity scores per offer, delivered into the platform your marketing team actually sends from.
Most segmentation projects end with a well-designed document describing six customer types with memorable names, which is then referenced in two presentations and never used to send anything. The difference between that and a useful segmentation is whether each segment maps to a specific action someone can take next week.
Customer segmentation is the grouping of customers by behaviour, value and needs so that different groups can be treated differently. Propensity scoring estimates, per customer, the likelihood of a specific action such as buying a product, responding to an offer or upgrading, so that targeting decisions can be made at the individual level rather than at the group level.
Segments must map to actions
We start from the campaigns, offers and journeys you can actually run, and build segments that distinguish between them. A segmentation with no corresponding action is a description; a segmentation where two segments receive identical treatment has one segment too many.
| Segmentation basis | What it supports | When it disappoints |
|---|---|---|
| Value and recency | Prioritisation and service levels | Says nothing about what to offer |
| Behavioural and journey stage | Lifecycle and onboarding campaigns | Needs reliable event data |
| Needs and category affinity | Product and content targeting | Requires breadth of purchase history |
| Predicted value | Acquisition and retention investment | Sensitive to model assumptions |
| Propensity per offer | Individual targeting decisions | Not a segmentation; complements one |
In practice most organisations need a small, stable segmentation for strategy and communication, plus propensity scores for the individual targeting decisions. Trying to make one structure do both jobs is why so many of these projects disappoint.
How we build it
Few segments, clearly separated
Four to seven segments is usually the workable range. More than that and nobody can remember them, campaigns are not built for them, and the distinctions stop being real. We test separation statistically and, more importantly, check that each segment implies a different action.
Stable enough to act on
Customers moving between segments every month make campaign planning impossible. We measure migration rates and tune for stability, because a segmentation that reshuffles constantly cannot support a quarterly plan.
Propensity per offer, not in general
A single propensity score is close to meaningless. We build a model per meaningful action, such as likelihood to buy a specific category, to upgrade, or to respond to a particular offer type, because those are the decisions the marketing team is actually making.
Predicted value with honest uncertainty
Customer lifetime value estimates drive acquisition and retention spend, and they are sensitive to assumptions about horizon, discount rate and churn. We report ranges rather than point estimates and expose the assumptions so finance can argue with them properly.
Activate into the platform
Segments and scores are written into the CRM or marketing platform as fields that can be used in audience builders, refreshed on a schedule. A segmentation that lives in an analytics environment requires an export request every time anyone wants to use it, and therefore is not used.
Measure against a control
Targeted campaigns run against a holdout so the lift from targeting is measured rather than assumed. Otherwise every campaign result gets attributed to the segmentation, which is neither honest nor useful for deciding what to do next.
Segments describe, propensity decides
Use the segmentation for strategy, planning and communication, where a small number of memorable groups is genuinely useful. Use propensity scores for the individual targeting decision, where the group average will always be worse than the individual estimate. Most disappointing projects tried to do both with the same object.
The data that makes this work
- Transactions with product detail, not just totals, so category affinity can be learned.
- Engagement events such as opens, visits, sessions and feature use, with reliable timestamps.
- Service history, because support experience predicts behaviour more strongly than most marketing data.
- Campaign history including who was contacted, with what, and who was deliberately not, which is what makes lift measurable.
- Identity resolution across channels, without which every model is learning from fragments of the same person.
Where campaign history has no control groups, we design them into the first campaigns rather than trying to reconstruct causality from the past.
How the engagement runs
Built from the campaigns you can run, and activated where your team sends from.
Action inventory and data audit
The campaigns, offers and journeys available are catalogued, and customer data assessed including identity resolution.
Segmentation
Candidate segmentations built and tested for separation, stability and, above all, distinct implied actions.
Propensity and value models
Per-offer propensity models and predicted value with uncertainty, validated on time-based splits.
Activation
Segments and scores written into the CRM or marketing platform on a refresh schedule, with control groups configured.
Measurement and handover
First campaign lift measured against control, with the refresh and monitoring runbook.
What you receive
Segments and scores in the platform, with the first measured lift already in hand.
Segmentation
Four to seven segments with definitions, sizes, migration rates and the action each implies.
Propensity models
Per-offer scores with validation, drivers per customer and refresh cadence.
Predicted value
Customer value estimates with ranges and the assumptions exposed for review.
Platform activation
Segments and scores written into your CRM or marketing platform as usable fields.
Control group design
Holdouts configured so campaign lift can be measured from the first send.
Refresh and monitoring
Scheduled rescoring, migration monitoring and the trigger for rebuilding the segmentation.
Is this the right engagement?
Worth being direct. Customer Segmentation and Propensity Scoring is the wrong spend in some situations, and those are listed rather than buried.
Good fit if
- You run campaigns and currently target by broad rules or by everyone.
- Transaction and engagement data exists with reasonable identity resolution.
- Marketing can run different treatments per segment.
- Control groups are acceptable so lift can be measured.
- Someone owns the campaign calendar and can act on the output.
Choose something else if
- The customer base is small enough to treat individually.
- Identity is so fragmented that customers cannot be resolved across channels.
- The output would be a strategy document with no campaign attached.
- The specific question is churn, which is churn prediction.
Frequently asked questions
Marked up with FAQPage schema so these answers can surface directly in search results and inside AI assistant responses.
How many customer segments should we have?
Usually four to seven. Beyond that nobody remembers them, campaigns are not built for them, and the distinctions stop implying different actions. The test is not statistical separation but whether each segment would receive genuinely different treatment.
What is the difference between segmentation and propensity scoring?
Segmentation groups customers so they can be treated differently at a strategic level. Propensity scoring estimates an individual's likelihood of a specific action. Use segments for planning and communication, and propensity for the targeting decision, because the group average is always a worse estimate than the individual one.
How often should segments be refreshed?
Scores monthly or weekly depending on how fast behaviour changes; the segmentation structure itself far less often, typically annually. A structure that shifts every month cannot support a campaign plan, which is why we tune for stability as well as separation.
How do we know the targeting is working?
With control groups configured before the first send, so lift is measured rather than assumed. Without them, every campaign result gets credited to the segmentation and you learn nothing about what to do differently next quarter.
Can you predict customer lifetime value?
Yes, with ranges rather than point estimates, because the result is sensitive to assumptions about horizon, churn and discount rate. We expose those assumptions so finance can challenge them, which is more useful than a single confident number nobody trusts.
Often paired with this
Most clients combine two or three engagements from the Machine Learning & Predictive Analytics pillar. These are the ones that most often run immediately before or after.
Churn Prediction and Retention Modelling
Churn models aimed at persuadable customers, with uplift modelling and impact measured against a holdout.
Read more →Predictive Analytics
Predictions embedded in the decisions they are meant to improve, measured on the business outcome rather than on AUC.
Read more →Recommendation Engine Development
Recommenders tuned to a business objective, with cold start, diversity and rules handled, proven by A/B test.
Read more →Is this the right engagement?
Tell us what you are trying to build. If a different service fits better, or if you do not need us at all, we will say so.