Weighting Signals by Recency and Frequency

Most teams dump every event into one table: a website visit from RB2B, a pricing page hit from Koala, a job change from Cognism, a competitor mention from a social listener. Treated equally, a six-week-old whitepaper download outranks a pricing-page visit from an hour ago, and reps chase cold accounts while warm ones cool off. The funnel model assumed a slow linear march; intent does not work that way. Buyers research in public, in bursts, and they go quiet just as fast.

Treat the signal stream like an event log you version and query, not a static lead list. In a warehouse-backed setup with HubSpot or Salesforce as the system of record, every signal carries a timestamp and a source. Weighting is simply the function you write over that log to answer one question: which accounts deserve a human today. Without weighting, your queue is noise; with it, the queue becomes a priority list reps can trust.

Recency is best modeled as exponential decay: a signal's weight halves over a fixed window so a visit today scores far above the same visit two weeks ago. Pick a half-life per signal type, because a demo request stays hot for weeks while a generic blog read decays in days. Tools like Clay let you compute decay in an enrichment column, and Warmly or Koala expose recency natively on their account timelines. The point is to make age a first-class input, not an afterthought.

Frequency captures intensity: three pricing-page visits in two days from one account beats a single drive-by. Combine the two so a frequency burst can lift an account even if individual events are not the strongest type. A simple recency-times-frequency product works as a v1, but cap frequency so one looping bot or a single obsessive researcher does not dominate the queue. Store the resulting score back on the account in Salesforce so Smartlead, Instantly, and your SDR views all read from the same number.

Weights are code, so version them. Keep the half-lives, frequency caps, and per-signal multipliers in a config you can diff and roll back, not buried in a no-code workflow nobody can audit. When you change a multiplier, log it, and watch how the top of the queue shifts over the next week. Many teams over-weight a flashy signal like a competitor mention and then wonder why reps burn cycles on accounts with no real buying motion.

Validate against outcomes, not vibes. Pull won and lost deals from HubSpot and check whether high-weighted accounts actually converted faster, then adjust. Watch for staleness: a signal source like Leadfeeder or Snitcher can change its firing behavior after an update, quietly inflating scores. Review the weighting model on a regular cadence the same way you review a service-level objective, because a scoring system that is never re-tuned drifts away from reality.