Lead Scoring Built From Closed-Won, Not From a Workshop

I will not build a scoring model on fewer than 150 closed-won records from the last 18 months. Below that, anything I produce is a workshop opinion with arithmetic on top, and the honest deliverable is a simple fit rule plus a plan to instrument the data so a real model becomes possible next year. When the records exist, the model is fitted against them and against the losses, and it is validated on a quarter it has never seen. Two axes, not one number: A single score forces the sales team to treat a well-fitting company that has done nothing exactly like a poor-fit browser who read six pages. I score fit and engagement separately and hand sales a grid: fit A to D, engagement 1 to 4. The action differs by cell. A1 goes to a rep today. A4 goes into a nurture with a named owner. D1 goes nowhere and should be told so rather than left in a sequence for two years. How fit is built: From closed-won and closed-lost, not from the ICP slide. Job titles are normalised against a mapping I maintain rather than matched on strings, because Head of Growth and Growth Lead are the same buyer and Growth Marketing Intern is not. Firmographics come from your enrichment source with the null rate stated up front. A field empty on 40 percent of records cannot carry weight, whatever the correlation looks like on the rows where it exists. How engagement is built: Decay is fitted from your own data rather than taken from a default. In insurance the half-life of a premium calculator completion is measured in days; in a fleet purchase it is not. Negative scoring matters as much as positive: careers pages, support documentation, existing-customer logins, competitor domains. Every score carries the three signals that contributed most, written to a field, so a rep can argue with the number. A score nobody can interrogate is ignored inside a quarter. Validation, and the answer that means do not ship: The model is fitted on older data and tested on the most recent complete quarter. I report precision at your actual sales capacity, because if the team can work 200 leads a month then the only question that matters is what share of the top 200 convert. When lift over your current rule comes in under roughly 1.4x, I tell you the model is not worth the change management, and we spend the remaining time on the data instead. That has happened. I would rather say it than ship a number. Where the score lives: Native scoring in Marketo or Account Engagement when every input already sits in the platform. When it does not, the model runs as a scheduled job against your warehouse and writes one read-only field back, with each run logged and an alert if the write fails. Marketing does not get edit access to the scored field. Not included: I do not resell predictive scoring add-ons and I do not build black-box models here. If a rep cannot be told why a lead is an A2, the model has failed at its job whatever its AUC says. I do not clean your enrichment data as part of this, though I will quantify how bad it is and what the gaps cost you in missed routing. Not for you if: Your sales team does not disqualify. Without recorded losses and disqualification reasons there are no negative labels and there is no model, only a popularity contest among your busiest visitors. Also not for you if the threshold is a political number that has to stay where it is.