A propensity model estimates the probability that a person will do something (buy, respond, leave). For an SME it does not need to be sophisticated: with simple rules and data you already have, you can identify who is most likely to buy and direct more relevant campaigns and messages to them.
The essentials in a definition
Propensity models estimate the probability that a customer or visitor will take an action (buy, subscribe, churn), based on behaviour and characteristics. For SMEs, simple versions (rule-based scores or basic models) help prioritise ads and contacts, using event and data warehouse data. Data processing and automated decisions require GDPR compliance.
What propensity is
It is an estimate, based on past behaviour, of the probability of a future action. Someone who visited the pricing page three times, opened the last emails and has bought before is, generally, more likely to buy again than someone who merely saw an ad.
It lets you concentrate effort where there is more return.
A simple model in a spreadsheet
Choose the signals that, historically, precede purchase (visits to key pages, number of sessions, response to emails, time since last contact). Assign points to each signal. Add up and create levels: high, medium, low propensity. Validate with real data: do the high-propensity people buy more?
Refine the points with experience.
A simple propensity model
- Identify signals that precede purchase
- Assign points and create levels
- Direct messages and budgets by level
- Validate and refine with real data
Using it in campaigns
Direct different messages and budgets to each level: high propensity (incentive to close, direct remarketing), medium (proof and comparison content), low (light content or no paid remarketing). Exclude those who have already bought. Send the segments to the platforms as audiences, with consent.
Measure conversion rate and cost per level.
Knowing who is close to buying lets you speak to them differently.
Cautions
Avoid using sensitive data or data the person would not expect. Bear in mind biases: the model reflects the past. Update the points regularly. Do not over-trust: use it as a guide, not as certainty. Respect data-protection rules.
If the base is small, keep the model very simple.
Decision table
| Approach | Data | Complexity |
|---|---|---|
| Rule-based score | Few variables (visits, pages, time) | Low |
| Simple regression | Purchase and behaviour history | Medium |
| More advanced models | Lots of data and variables | High |
| Platform models | Predictive audiences in campaigns | Low, with less control |
Common mistakes to avoid
- Complicating early. Start with rules and measure the gain.
- Using data without a legal basis. Comply with rules and minimise.
- Not validating. Compare the prediction with real results.
How to measure the result
Compare conversion rate by score band and the cost-per-conversion gain against campaigns without scoring.
Limits and exceptions
Predictive models can bias decisions and fail in new contexts. With little data, they are unstable, and the extra complexity may not pay off. Avoid using sensitive categories and keep human oversight. Regularly review the score against real results and simplify if the gain is small.
Illustrative example
Illustrative example: a shop creates a simple score based on recent visits, product pages viewed and add-to-carts. High-score visitors receive more intense remarketing, and low-score ones, informational content. The high-score group’s conversion confirms the rule’s usefulness, and the shop refines it over time.
In practice: checkpoints
Start with simple scoring based on a few strong signals, such as repeat visits, product views and past purchases, and compare results with and without it. Use the output to prioritise follow-up and offers, not to exclude customers unfairly. Keep personal data to the minimum, document the logic, retrain periodically and check that the model has not drifted. Treat the scores as probabilities, and combine them with common sense and customer service insight, since small businesses rarely have enough data for perfect predictions.
Frequently asked questions
Do I need data science?
Not to start. Simple rules already work.
What data can I use?
Site behaviour, emails and purchases, with respect for data protection.
How do I validate the model?
By checking whether high propensity corresponds to more purchases.
Do I need data science?
Not to start; simple rules and validation with real data already give value.
What legal precautions?
Inform, minimise data, have a legal basis and avoid automated decisions with significant effects without oversight.
Simple propensity models (rule-based scores) help prioritise who is most likely to buy; validate them with real data and respect the GDPR.
Sources and recommended reading
Last editorial review: October 2026. Informational content; always confirm current conditions in the official sources before deciding.