What a Propensity Score Actually Knows About Your Clients

Most recall systems run on rules someone wrote once: six weeks after a color service, six months after a cleaning, ninety days after a neurotoxin visit. The rules are fine as averages. Nobody is an average.

A propensity score replaces the rule with an estimate: how likely is this specific person to reply and book if we reach out this week? It is a number between 0 and 1, refreshed nightly, and it exists for every contact in your booking system, not just the ones who are "due."

The signals that matter

The model is trained on your brand's own history, so the exact weights differ from one business to the next. The signals are consistent:

• Personal cadence, not category cadence. Devon comes every six weeks like clockwork; at nine weeks she's a strong signal. Priya comes when she feels like it; at nine weeks she isn't late at all.
• Service mix and unfinished business. Three unused sessions on a laser package are worth more than any promo.
• How they've responded before. Who replies to texts, how fast, and at what time of day. Someone who has replied within ten minutes twice is a different contact from someone who has never answered.
• Membership and package state. A lapsed membership with a billing question in the last thread is high-value and high-risk at the same time. The score flags both.
• Local demand and season. A Tuesday in late August in Round Rock looks different from the same Tuesday at the Domain.
• Recent contact load. Four messages in the last thirty days pushes a score down regardless of everything else. Fatigue is a feature, not an afterthought.

What it deliberately doesn't do

The model does not infer protected characteristics, and it does not train on other brands' data. Scores decay when contacts go quiet, and a contact without recorded consent is never eligible regardless of score. These aren't limitations we worked around; they are the design.

From a score to a plan

A ranked list isn't a plan. The plan comes from your calendar. Open capacity at each location sets the threshold: a soft week reaches a little further down the ranking, a full week barely reaches at all. Every ask carries an expected booking probability, so the plan states what it expects to fill and by how much it might miss.

Everyone below the line gets a reason, not just a no. "Too soon: four messages in thirty days." "Low intent: consult no-show, 200+ days." "Protect: regular, will book on her own." Your team can read the plan and disagree with it, which matters more than people think.

How it learns

Every outcome flows back: replied or didn't, booked or didn't, kept the appointment or didn't, opted out. Nightly retraining folds that in. New brands start from priors built on their industry and usually settle into stable, brand-specific scores within a few weeks. After that the model mostly gets quieter and more precise, which is exactly what you want from something deciding who not to text.