Two questions, four answers
No black box. The quadrant asks two things of every customer: would they purchase anyway, and does contact actually move them? Cross the answers and you get four classes, and only one of them deserves a discount.
The quadrant, in plain English
Would they purchase anyway?
Intent is read from the habit itself: how recently and how often they purchase, whether the rhythm is climbing or slipping. A live habit is its own predictor; it does not need to be bought.
Does contact actually move them?
Responsiveness reuses the engagement score: people who show up are people a message can reach. An incentive cannot persuade someone it never reaches, however generous it is.
One flag sits on top: a protected relationship, a champion habit or a subscription. High intent inside a protected relationship is a sleeping dog, and the discipline is to leave it alone rather than risk disturbing a routine that was working on its own.
The four classes
Would have acted anyway. An offer here buys nothing that was not already happening; every redemption is margin given away.
Acts only with a nudge. The one class where an incentive earns its cost, so this is where the whole discount budget belongs.
A healthy, protected habit. Contact with a discount can only disturb it: leave well alone and recognise, never discount.
Beyond an offer at any sensible price. No incentive; an efficient always-on presence is all this relationship justifies.
No new tracking to invent. These are events you already capture; the model reads them as they are.
- ·Offer and redemption history, who was sent what and who actually cashed it
- ·Full-price vs incentivised purchases, so intent is read before the discount muddies it
- ·Margin per purchase where it is tracked, so the model optimises profit rather than response
- ·Holdout results from incentive campaigns, the training data for a true uplift model
Models are tuned, not installed. Each step is earned on evidence, and the previous step keeps running until the next one beats it.
- v0Heuristic, day oneTransparent rules on the data you already have. Live in Braze in weeks, and everyone can see why it fired.
- v1Calibrated to your historyThresholds and weights re-fit to your own base and vertical, backtested against what actually happened.
- v2ML where it earns itA trained model replaces the rules only when it beats them on holdout data. Explainability stays a requirement.
- v3Monitored and re-tunedDrift watched, thresholds reviewed on a cadence, and every change proven with holdouts before it ships.
Starts as a heuristic on the redemption patterns you already have. It becomes a true uplift model only when it trains on your own holdout experiments from /prove, and it is re-tuned as offers, price points and seasons change.