Lead Scoring Rebuilt Against Closed-Won Evidence

Email opens stopped being a usable scoring signal in September 2021, when Apple Mail Privacy Protection began prefetching images for every message its users receive, and in the Polish lists we measure that is between 35 and 55 percent of opens reported. Deleting open based points is the first edit we make. On a logistics client last year that single change moved the share of qualified leads that became opportunities from 9 percent to 23 percent, because the score stopped rewarding people for using an Apple device and started ranking on things they chose to do. Whether you have enough evidence for a model at all: With fewer than roughly 200 closed won and 600 closed lost records in the last eighteen months, a fitted model memorises your history instead of learning it, and it will look excellent on the data it was built from. Below that line we build an explicit rules model, weights argued out and written down, which a sales director can challenge line by line. Above it we fit, and we hold out the two most recent quarters to test against. We tell you which of the two you are before you commit, and we have talked buyers out of the larger engagement on exactly this basis. Fit is a fact, behaviour is an inference: Fit comes from registers, not from what a person typed into a form. In Poland that means PKD activity class and legal form from the GUS REGON service, VAT status from the white list, and for road transport operators the national register of carriers, which tells you licensed vehicle counts, a far better predictor for a freight seller than any self declared employee band. For food and drink, outlet counts and whether the buyer is a chain, a wholesaler or a single site. Behaviour is the small set of actions that appear disproportionately in your closed won records, which is almost never the actions your current score rewards. What comes out first: Points for opening. Points for existing customers' own staff visiting the site. Points for careers pages, competitor domains and the agency that monitors you. Points that only ever accumulate, which is why a lead from 2022 outranks one from last week. Decay gets a half life chosen from your median days to opportunity, not from a template that says thirty days because thirty is a round number. Calibrating the threshold, which is a business decision: The threshold is not a property of the model. If your team can genuinely work 40 qualified leads a week, we set the cut where the weekly volume matches that, then report the precision you are buying at that setting and what you would gain or lose by moving it either way. Accuracy is not reported, because on this class of problem it flatters everything. The part no score fixes: If sales does not record why a lead was rejected, the model cannot be retrained and the next version will be a guess. We design the disposition codes and the handover timing with your sales director in the room, and if that conversation cannot be scheduled we would rather not start. Recycling rejected leads back into nurture is part of the specification, not an afterthought. Delivered as: A written model specification with every weight and its justification, the backtest with precision and recall at the chosen threshold and at two alternatives, the implementation inside HubSpot, Marketo Engage, Account Engagement, Marketing Cloud or ActiveCampaign, and a one page explanation a sales rep can read in two minutes and argue with. Not for: Teams without an agreed definition of a qualified lead, since scoring will simply automate the disagreement. Buyers who want a predictive label they cannot explain to the person acting on it. Companies with no closed lost data, where there is nothing to learn from and we will say so on the first call.