Propensity-to-Buy Models for Email and Ads

Why a whole-list send is expensive even when email is cheap

Email feels free, so many businesses send every campaign to every subscriber. The cost shows up slowly: rising unsubscribes, falling engagement, and inbox providers deciding your mail belongs in the promotions tab or worse. Paid ads make the cost visible immediately, because you pay for every impression shown to someone with no intention of buying.

A propensity-to-buy model is the standard answer. Instead of broad segments such as 'bought in the last year', each person gets a score for how likely they are to buy in the next 30 days, and campaigns target by score.

How a propensity model is built

At its core, a propensity model is a classification model. You pick a past moment, describe every customer as they looked then, and label whether they bought in the following window. The model learns which descriptions preceded a purchase.

• Recency, frequency and value. Still the strongest signals for most retail and subscription businesses.
• Browsing and email behaviour. Product views, basket additions, clicks on recent campaigns.
• Category history. What they buy, and what customers who bought the same things went on to buy.
• Life stage of the relationship. New customers behave very differently from five-year regulars.
• Seasonal patterns. Some customers only appear before Christmas or back to school.

Product-specific propensity models, such as likelihood to buy a mattress in the next 60 days, are more useful than a generic score for businesses with distinct product lines. They cost more to maintain, so we usually start with one generic model and two or three specific ones for the categories that matter most.

Using the scores for email

The bands above are illustrative; the real cut-offs come from testing. A catalogue retailer with a 200,000-person list might find that sending only to the top half keeps most campaign revenue while halving unsubscribes and complaints. That is the kind of trade you want to measure rather than assume.

Using the scores for paid ads

Ad platforms accept customer lists for targeting and exclusion. Propensity scores turn those lists into something more useful:

1. Upload high-propensity customers as a seed for lookalike audiences, instead of all customers
2. Exclude people very likely to buy anyway from expensive retargeting
3. Exclude recent buyers of durable products who will not buy again for years
4. Build category-specific audiences for campaigns promoting particular product lines

The exclusion step is often where the quickest savings are. It pairs naturally with better measurement; our post on attribution for small marketing budgets explains how to tell whether the spend moved anything.

The mistake that flatters propensity models

Here is the trap. The customers with the highest propensity to buy are, by definition, the ones most likely to buy whether or not you email them. Target them, and your campaign's conversion rate looks wonderful. Some of that revenue would have arrived anyway.

A propensity model tells you who will buy. It does not tell you who will buy because of your message.

The only reliable check is a holdout: keep a random slice of each score band out of the campaign and compare their purchases with those who received it. If the top band buys at the same rate with or without the email, stop spending discounts on them. When that difference is the real question, uplift modelling is the tool built for it.

When propensity scoring is not worth building

Small lists do not need it. Below a few thousand active customers, simple segments based on last purchase date and category capture most of the value and anyone can understand them. Businesses where nearly every customer buys on a fixed cycle, such as annual insurance, get more from renewal-date logic than from a model.

Many email platforms now include predictive scores out of the box. For a Shopify or Klaviyo store with standard data, try those first. A custom model makes sense when you have data the platform cannot see, such as in-store purchases, trade accounts or call centre orders, or when you need scores for multiple channels from one consistent source.

What we would build first

At SpiderHunts, a first propensity project is usually a generic 30-day purchase model, scored nightly, pushed into the email platform and ad audiences as a property, with a standing holdout group in every campaign. Accuracy is reported, but the headline number is incremental revenue per thousand sends compared with the old approach.

That is typically four to six weeks of work if customer data sits in one place and longer if online and offline identities need matching. Our machine learning services page covers how we scope it. The consent side matters too: use scores only on people who agreed to marketing, and keep sensitive inferences, such as health-related purchases, out of ad audiences entirely.

Frequently asked questions