Lead scoring rules that survive contact with real traffic

Lead scoring exists to answer one question: who should a human talk to first? Any model that does not make that decision easier is decoration. That rules out most of what gets built.

The failure mode is always the same. Someone assigns points for eleven signals, never documents the thresholds, and eighteen months later nobody can explain why a particular lead scored 74. The sales team stops trusting it, and the score becomes a column nobody sorts by.

Start with fit, then add intent

Two different things get blended into one number, and they should be added separately. Fit is who the lead is: company size, role, whether they used a work email. Intent is what they did: which pages they saw, whether they returned, whether they clicked pricing.

• Fit signals are slow-moving and mostly stable. Score them once, at capture.
• Intent signals decay. A pricing view today is worth more than a pricing view three weeks ago.
• A lead with great fit and zero intent is a list entry, not an opportunity.
• A lead with strong intent and poor fit is usually a student, a competitor or a hobbyist. Route it to self-serve, not to sales.

Keep the model small enough to explain

Five to seven signals is the practical ceiling for a team without a dedicated analyst. Every additional signal adds a coefficient somebody will eventually question, and the marginal accuracy is almost never worth the argument.

// a model you can read out loud
work email         +25
company >= 20      +20
role in {founder, growth, marketing}  +15
pricing page seen  +20
returned within 7d +10
demo requested     +10

// 0-100, clamped

Decay is not optional

Intent points should fade. A lead who viewed pricing in March and never returned is not the same lead in September, but a naive additive model keeps them at the top of the queue forever. Either recalculate the score from recent events or apply an explicit decay, and be honest in the UI that the number has a timestamp.

Close the loop or stop scoring

Once a month, take the leads that closed and the leads that were marked lost, and check whether the score separated them. If the top quartile converts at the same rate as the bottom quartile, the model is not working. That check takes an hour and is the only thing that keeps a scoring model from becoming folklore.

Beacon writes a score onto every lead from fit and session behaviour, and keeps it attached to the attribution record so a change in the model can be compared against closed deals later. The point is not precision. The point is a queue a human can trust on a Monday morning.