Use a predictive audience

Overview

Predictive audiences rank users by modeled outcomes—conversion likelihood, churn risk, or discount sensitivity. They complement rule segments by surfacing nuanced signals you would not hand-code.

Steps

1. Confirm your workspace has enough labeled outcomes (purchases, refunds) for the model family you pick.
2. Open Audiences → Predictive → New and choose the objective; set training and holdout windows.
3. Review feature importances to sanity-check the model (no single leaky feature should dominate).

{
  "model": "conversion_7d",
  "score_threshold": 0.72,
  "refresh": "daily",
  "fallback_segment": "rule_high_intent"
}

4. Route top deciles to ad platforms and middle deciles to email experiments.

Troubleshooting

• Cold start: predictive tiers need a few thousand examples—start with rules until volume matures.
• Drift alerts: retrain when major UX or pricing changes land.
• Privacy: ensure scores are not exported to vendors that cannot process inferred data categories in your regions.