Quick answer
Machine learning spots patterns that can help you predict responses or group audiences. The useful question is whether those predictions improve a marketing decision. A model can score well on a technical measure and still send sales after the wrong people.
Match the method to the decision
A model learns patterns from examples and applies them to new records. That can be useful, but only if the pattern helps you make a better decision. Don’t assume a deployed model learns from every new interaction; training, using and retraining it are separate steps.
| Method | Marketing question | Evaluation |
|---|---|---|
| Classification | Which inquiries may qualify? | Precision and recall at the chosen threshold |
| Regression | What value might an order have? | Prediction error against a simple baseline |
| Clustering | Which records have similar characteristics? | Whether groups are stable and useful |
| Recommendation | Which item may be relevant next? | Customer outcomes in a controlled test |
Accuracy can conceal the wrong mistakes
A model can look accurate when the outcome of interest is rare. Suppose only 5 of 100 inquiries qualify. Predicting that none qualify would be correct for 95 records, but it would miss every useful inquiry. Overall accuracy would conceal the failure.
Precision asks how many flagged cases are correct; recall asks how many relevant cases were found. Google’s guide to classification metrics explains these measures. The threshold should reflect the cost of unnecessary follow-up and missed opportunities, alongside the team’s capacity.
Keep the test separate
Reserve records the model has not used for training or tuning. For an operational marketing prediction, a later time period can reveal whether patterns carry forward. Check that the same person or near-duplicate records have not leaked across the split.
Use only information available when the real decision would be made. A field filled in after a sale cannot fairly help a model predict that sale. Our guide to data in marketing AI covers the preparation work.
Prediction does not tell you what caused a result
A propensity score may identify people likely to buy anyway. It does not establish that sending them another message creates extra sales. Where practical, compare the intervention with a suitable control group and measure the outcome the business cares about.
Monitor performance after launch by campaign and relevant segment. Keep a simple fallback process available when inputs are missing or the model no longer meets the agreed standard. A model is useful when it improves the decision, not merely when it produces a score.
Two useful applications are separating customer fit from engagement and testing whether a campaign changed the outcome. See our guides to B2B lead scoring and marketing incrementality testing.
