Lead scoring that reflects reality
Most lead scoring models have never been checked against a single closed deal, which means nobody in the building knows whether the score predicts anything. That is the problem worth fixing first, before adding rules. A model you can validate in an afternoon is more valuable than an elaborate one that has been accumulating points since the platform was installed and now marks half the database as sales ready.
Why do most scoring models stop meaning anything?
Because points only ever go up. A model that adds five for an email open, ten for a page view and twenty for a download, with no decay and no ceiling, will eventually push every long-standing subscriber over the qualification threshold through sheer persistence. The score stops measuring intent and starts measuring how long someone has been on the list.
The second cause is scoring on behaviour that means nothing. Views of the careers page, the support documentation and the existing customer login screen all generate points in default configurations, and all three are dominated by people who will never buy.
The tell is easy to spot. If the count of qualified leads grows steadily every month without any corresponding change in enquiries or pipeline, the model is inflating rather than qualifying, and sales has already quietly stopped trusting the flag.
Why should fit and behaviour be scored separately?
Because they answer different questions and combining them into one number destroys both. Fit asks whether this person could ever buy: the company size, the sector, the country, the job function. Behaviour asks whether they are interested right now. A single total cannot distinguish a perfect-fit prospect who has done nothing from a poor-fit student who has read everything.
Two axes give you four groups and each needs a different action. Good fit and high activity goes to sales today. Good fit and low activity goes to nurture. Poor fit and high activity goes nowhere, and recognising this group is the main saving, because it is the one that wastes sales time in a combined model.
The practical version is a letter for fit and a number for behaviour, which most platforms support directly. If yours does not, two custom fields do the job and the reporting is clearer for it.
| Signal | Type | Reasonable weight | Common mistake |
|---|---|---|---|
| Job function matches buyer or influencer | Fit | High | Scoring free-text job titles instead of a mapped picklist |
| Company size and sector in target range | Fit | High | Guessing from an email domain with no enrichment |
| Pricing or comparison page viewed twice in a week | Behaviour | High | Weighting a single view the same as repeat visits |
| Demo or contact form submitted | Behaviour | Highest, and usually a direct route to sales | Leaving it as points rather than an immediate handover |
| Email opened | Behaviour | Near zero | Treating opens as intent when privacy proxies open mail automatically |
| Careers, support or login page viewed | Neither | Zero or negative | Inheriting the platform's default page-view scoring untouched |
How fast should a score decay?
Fast enough that a burst of research three months ago no longer looks like interest today. A common and defensible approach is to halve the behavioural score every thirty days of inactivity, or to expire individual point events after a set window rather than decaying the total. Either way the fit score should not decay, because a company's size does not change because nobody visited the website.
Decay matters more than weighting. Getting the relative points slightly wrong produces a slightly noisy ranking. Having no decay at all produces a ranking by tenure, which is worse than random because it systematically promotes the least active long-term subscribers.
Set the decay period to something near your real sales cycle. If deals take two months from first contact, a signal from six months ago is history rather than intent, and treating it as current is what generates the calls that begin with a confused prospect.
How do you test whether the model works?
Export the last hundred closed-won opportunities and the last hundred closed-lost or disqualified ones, and record what each contact's score was at the moment sales first made contact. Not today's score, the score at handover, which is why you need the score history field or a snapshot written on the stage change.
Then compare the two distributions. If the won group sits noticeably higher, the model has signal and the threshold is a matter of tuning. If the two overlap almost entirely, the score is noise dressed as a number, and every hour sales spends prioritising by it is wasted.
This is an afternoon of work and almost nobody does it. It also tells you where to set the threshold honestly: pick the point where the proportion of qualified leads that actually convert is high enough for sales to keep answering, rather than the point that produces the lead volume the marketing target requires.
What belongs in negative scoring?
Signals that reliably indicate this person will not buy, rather than signals that annoy you. Competitor domains, personal email addresses when your product is sold to organisations, countries you cannot serve, students and job applicants identifiable from the form, and existing customers who should be in a different flow entirely.
Be careful with the free-mail rule. In some markets a personal address is normal for a legitimate buyer, and a blanket penalty silently suppresses a segment nobody is monitoring. If you apply it, check what it is filtering out once a quarter rather than assuming.
Unsubscribes and hard bounces should not be handled by negative points at all. They are a state, not a score, and belong in suppression logic where they cannot be outweighed by later activity.
Who should own the threshold?
Sales and marketing jointly, reviewed on a fixed cadence, with the conversion rate at the threshold as the number that decides it. A threshold owned by marketing alone drifts down whenever lead volume targets are missed, which is the mechanism behind most of the mistrust between the two teams.
Give sales an explicit way to reject a qualified lead with a reason, and read those reasons monthly. They are the cheapest source of model corrections available, and they arrive already written by the people closest to the outcome.
Then recalibrate when something structural changes: a new product, a new market, a pricing change, a website restructure that renames the pages the model scores. A model tuned to a site map that no longer exists degrades silently, because the rules still run and the points still accrue.
Common questions
- What is lead scoring?
- A set of rules that assigns points to a contact based on who they are and what they have done, so that sales can prioritise. The rules are written by a person, not learned by the software, and they only have value if they have been checked against real outcomes. A scoring model that has never been compared with closed deals is an untested guess with a number attached.
- How do you test whether a lead scoring model works?
- Take the last hundred won deals and the last hundred lost or disqualified ones, and record each contact's score at the moment sales first made contact rather than today. Compare the distributions. If the won group sits clearly higher, the model has signal and the threshold needs tuning. If they overlap almost completely, the score is noise and prioritising by it wastes sales time.
- Should lead scores decay over time?
- Behavioural scores should, fit scores should not. Without decay, points only accumulate, so the highest scores end up belonging to whoever has been on the list longest rather than whoever is interested now. Halving the behavioural score after a month of inactivity, or expiring individual point events after a fixed window, keeps the ranking about current intent. Company size does not decay, so fit stays fixed.
- Why separate fit scoring from behaviour scoring?
- Because one number cannot distinguish a perfect-fit prospect who has done nothing from a poor-fit visitor who has read everything, and those two need opposite treatment. Scoring them on separate axes produces four groups: good fit and active goes to sales, good fit and quiet goes to nurture, poor fit and active goes nowhere. Identifying that last group is where most of the saved sales time comes from.
- Should email opens count towards a lead score?
- Barely, if at all. Privacy proxies used by common mail clients fetch images automatically, which registers as an open regardless of whether a human looked at the message. That makes open-based scoring a measure of which mail client someone uses. Clicks, repeat visits to pricing or comparison pages, and form submissions carry far more information and should carry the weight.
- What should count as negative scoring?
- Signals that reliably rule someone out: competitor domains, unservable countries, job applicants and students, and existing customers who belong in a different flow. Unsubscribes and hard bounces should not be handled with negative points, because a score can be outweighed by later activity. They are a state and belong in suppression logic where nothing can override them.
