Purchase propensity
Who is close to buying, and what
The problem
The contact strategy is the same for almost everyone. There may be three or four segments, but within each one the message is the same and the product offered is the one scheduled for that month, not the one that customer was about to buy.
The cost of that does not appear where it is generated. An extra send costs cents, so no one notices. What accumulates is unsubscribes, falling open rates, deteriorating sending reputation, and a base that responds less and less. The bill arrives three campaigns later, and by then no one attributes it.
McKinsey places the improvement in marketing return from personalisation at between 10 and 30%, and the reduction in acquisition cost at up to 50%. Personalisation decides who receives what. Propensity solves the first half of that question.
The solution
A probability per customer, for a defined time window, recalculated periodically. It is used in two ways, and it is worth choosing which one from the start because they are not built the same way.
General propensity
A single model that estimates the probability that a customer will buy again, regardless of what. It orders the entire base and answers who is worth contacting this month. This is what works when the catalogue is broad or when the decision is about contact budget.
Product propensity
One model per product or service, estimating the probability that that customer will purchase that product. It answers what to offer each one. It makes sense when the catalogue is short — three or four lines — because there a model per line can be maintained and the gain from getting the product right is large.
In both cases, the reason travels alongside the probability, which is what defines which piece corresponds to each tier.
What the model adds on top of the segmentation you already have
It does not replace it. Segmentation says who you talk to differently and in what tone. The model says, within each group, who is close to buying and what. They are two layers and they are used together.
What the model adds is that a segment describes what someone is like and a probability estimates what they are about to do. Two people in the same segment, with the same history, can be at opposite moments, and what separates them are signals from the last few weeks. A segment is a label that lasts until someone revisits it. A probability is recalculated with what has happened since the previous campaign.
More on the blog: From lab to production: three conditions agreed before training →
How it integrates
The score enters the campaign tool as a regular field, and is segmented on just like any other. With product propensity, several fields come in — one per line — and the highest for each customer marks the offer. The recalculation fires before each send, because a purchase probability ages with every action the customer takes or does not take.
The same model, other applications
A propensity model ranks a base by the probability that something happens. The event changes, the mechanics do not.
Lead scoring.
Which inbound contacts are most likely to close, so the sales team can order their week.
Call queue.
The same criterion where the bottleneck is daily contact capacity.
Account expansion.
Which customers are close to consuming more, for the team managing the portfolio.
Print run selection.
How much of the base to include in a high unit-cost send, whether mail, sample or event.