What is lead scoring, and does it actually work?
Assigning numerical values to leads based on who they are and what they have done, so sales attention goes to those most likely to buy. It works when built on evidence and fails — commonly — when built on assumptions.
The two components, which should be separated:
Fit (explicit) scoring — attributes of the person and organisation: industry, size, role, seniority, region, technology used. This answers should we sell to them?
Engagement (implicit) scoring — behaviour: pages visited, content downloaded, emails opened, events attended, demo requested, pricing page viewed. This answers are they interested now?
Keeping them separate matters, because a highly engaged poor-fit lead and a perfect-fit unengaged lead need completely different responses — and a single combined number hides the difference. Many teams plot them on a grid instead.
Why scoring commonly fails:
Scores set by opinion. Points assigned in a workshop by whoever felt strongly reflect beliefs, not evidence. The fix is to derive weights from historical data — what actually preceded closed deals.
Rewarding volume of activity rather than meaningful activity. Ten blog visits are not a buying signal; one visit to the pricing page frequently is.
No decay. Interest is time-sensitive, and without decay old activity keeps scores inflated indefinitely.
No negative scoring for disqualifying signals — competitors, students, job applicants, unsubscribes, and free-email domains where that matters.
Never recalibrating, so the model drifts as the market and content change.
Sales not trusting it, which is self-fulfilling: if high-scoring leads disappoint once, the score is ignored thereafter.
What actually makes it work:
Build from closed-won and closed-lost data.
Agree the definitions with sales, particularly what counts as qualified and what the handover triggers.
Validate against outcomes — if high-scoring leads do not convert better, the model is wrong.
Keep it simple enough to explain, since an unexplainable score is not actioned.
Consider predictive scoring where data volume supports it, with the same validation requirement.