Predictive targeting vs. retargeting
Compare audience selection, data requirements, and when to combine the two.
Read: predictive targeting vs. retargetingHome / Predictive lead scoring
Predictive lead scoring
A practical workflow for turning historical outcomes into a testable sales-prioritization process.
1. Define the decision before the score
Predictive lead scoring ranks leads or accounts using historical outcomes. Its practical purpose is to help a team decide where to spend limited attention. A score is useful only if it improves that decision.
Write a one-sentence specification: “Each weekday, rank eligible trial accounts by their likelihood of becoming paying accounts within 30 days.” Decide whether you are scoring a person, an account, or an opportunity; mixing these units can count the same purchase several times.
Choose a baseline, such as the team's existing qualification rules. Keep its contact capacity fixed so a model cannot appear better simply by contacting more people.
2. Build a prediction-time dataset
| Field | Available when scored? | How to use it |
|---|---|---|
| Account age | Yes | Compute relative to the scoring date. |
| Product activity in the previous 7 days | Yes, if recorded by then | Use a consistent lookback window. |
| Completed purchase in the next 30 days | No | Use as the outcome label, never as an input. |
| Sales status updated after purchase | No | Exclude; it would reveal the answer. |
Use historical snapshots instead of today's CRM values pasted onto old records. Allow the full outcome window to finish before labeling a lead as a non-converter. Document missing fields, duplicate accounts, exclusions, and sales-process changes.
3. Evaluate on later, unseen records
Train on earlier records, tune on a separate validation period, and reserve a later period for the final comparison. Avoid leaking related or duplicate records across splits. Fit preprocessing only on the training portion. scikit-learn: Common pitfalls and data leakage .
Evaluate the top group your team can realistically contact. If capacity is 50 accounts per week, measure performance at 50, not only across the entire database. Check whether performance collapses for new accounts or records with missing activity.
If you display probabilities, check calibration separately from ranking. A model can order accounts well while giving unreliable percentages. scikit-learn: Probability calibration .
4. Design the sales handoff
- Show when a score was calculated and when its underlying data was last refreshed.
- Explain which observed inputs contributed, without presenting those inputs as proof of a person's intent.
- Keep eligibility and contact preferences separate from model ranking.
- Give sales staff a way to record incorrect data and review overrides.
- Keep a fallback to the existing process if scores become stale or unavailable.
Illustrative workflow: generate a ranked list each morning, apply contact eligibility, let a reviewer select the day's accounts, and record the action taken. Logging the decision makes later evaluation possible.
5. Test the action, then monitor
A model trained on historical sales cannot tell you by itself whether a new sales call will change the result. Use a controlled test where feasible, comparing outcomes for otherwise eligible accounts assigned to the new process and the baseline.
Before launch, name an owner, choose a review schedule, and define what would trigger a pause: a broken input, an unusual score distribution, or deterioration in observed outcomes. Reassess after major pricing or product changes.
Work through the measurement example before using a score to expand outreach.
Questions to ask a lead-scoring vendor
- What outcome and time window does the score predict?
- Which fields exist at prediction time, and how is leakage checked?
- How does performance compare with our baseline at our capacity?
- Are percentages calibrated on unseen data?
- How are missing data, drift, overrides, and unavailable scores handled?
- What evidence shows the intervention changes outcomes?
The answers should be specific enough to test on your own process. A single accuracy percentage is not an evaluation plan.
Sources & further reading
Continue reading
Measure targeting performance
Work through precision, lift, calibration, and an incremental-results example.
Read: measure targeting performance