A definition you can quote, the formula with every component defined, an illustrative worked example, and how to tell whether a scoring model is doing anything at all.
Free Forever • No Credit Card Required
Quick answer
A lead scoring model assigns each lead a number representing how likely it is to become a customer, so that limited selling attention goes to the leads most worth contacting.
The purpose is ordering, not prophecy. A scoring model does not need to predict which leads will convert; it needs to rank them well enough that working from the top of the list beats working in the order things arrived.
Score = Σ (fit weight × attribute match) + Σ (behaviour weight × recency multiplier) − Σ (disqualifying deductions).
Attributes of the lead and their organisation: industry, company size, role, country, technology in use. These are stable, known early, and describe whether this is the kind of customer you can serve well.
Actions taken: pricing viewed, demonstration requested, email replied to, trial started. These describe whether the lead is doing anything now, and they should be capped per action type so repetition cannot dominate the total.
A factor that reduces the weight of older actions. Without it, accumulated history outranks present interest, which is the opposite of what the model is for.
Deductions for signals that indicate a non-buyer: a free email domain where your customers use business addresses, a careers page visit, a competitor domain, a country you do not serve.
The cut-off at which something happens. Choose it deliberately against current data, so that the qualifying band is a workable minority rather than most of the list.
These weights are invented to demonstrate the calculation. They are not recommended values and should not be copied.
Fit. Target industry: 15. Company size within the target band: 15. Decision-making role: 10. Country served: 5. Fit subtotal: 45.
Behaviour. Pricing page viewed this week: 20. Demonstration requested: 25. Three email replies worth 5 points each, of which one was this week and two were seven weeks ago and carry a recency multiplier of 0.6, giving 5 + 3 + 3 = 11. Behaviour subtotal: 56.
Negative. Free email domain rather than a business address: minus 10.
Total = 45 + 56 − 10 = 91. Against a threshold of 70, this lead qualifies and routes to sales immediately. Just as importantly, the representative sees the breakdown: the demonstration request and the pricing view are what drove it, so the first call has an obvious opening.
It buys back attention. A team that can contact forty leads a week and receives two hundred must choose, and the alternatives to a model are choosing by arrival order, by whichever company name is recognisable, or by whoever shouted loudest in a form. All three are worse than a mediocre score.
The second function is routing. A threshold turns a number into an action: immediate assignment, an alert, entry into a sequence, or a nurture track for leads that fit but are not yet active.
Unbounded behavioural points plus no decay means every long-standing lead drifts upward until the threshold has to keep rising. Eventually most leads qualify and the model has stopped ranking anything.
When conversion is rare, a model that predicts nobody converts will look accurate. Lift by decile is the only validation that answers the question the model exists to answer.
Email opens are a particularly weak signal, since privacy features in some mail clients can register opens the recipient never performed. Weighting them heavily produces a model that promotes people who did nothing.
A model designed by marketing without sales input tends to reward marketing activity. Representatives then ignore the score, and the organisation concludes that scoring does not work.
Product, pricing, market, and traffic mix all change. A two-year-old model that nobody has tested is directing attention with false confidence.
A working model shows a clear conversion gradient across score deciles, a qualifying band that is a minority of leads, components visible on the record, negative scoring that actually excludes obvious non-buyers, and a revalidation date in the last twelve months. Sales use it without being asked to.
A failing one shows most leads above the threshold, no measurable difference in conversion between high and low scores, opaque totals, and representatives who privately work their own list. The most reliable diagnostic is to ask a representative what the score means. If they cannot say, the model is not in use regardless of what the dashboard reports.
| Aspect | Rule-based | Predictive |
|---|---|---|
| Where weights come from | Human judgement about the business | Learned from historical conversion outcomes |
| Data required | None beyond the attributes themselves | Enough past conversions to learn from |
| Explainability | High: every point can be traced | Lower, unless component contributions are exposed |
| Typical ranking quality | Adequate, limited by the assumptions | Usually better at separating the top band |
| Main failure mode | Weights reflect belief, not evidence | Learns from a biased or stale history |
Put the score and its components on the lead record, keep the score history so warming and cooling are visible, and attach a concrete action to the threshold. A score that a representative can read in five seconds and act on immediately is worth considerably more than a more sophisticated number they have to take on trust.
Continue with lead management software, AI CRM capabilities, what a CRM is, sales automation, CRM for small business, and product features.
Every team in this space loses revenue to the same recurring gaps. Here is what they cost you and how HelloGrowthCRM closes each one.
Everyone qualifies as a hot lead and the sales team has stopped looking at the score.
That is score inflation, usually caused by unbounded behavioural points and no decay. Cap the contribution of any repeated action, apply recency decay so old activity fades, and reset the threshold against current data so that the qualifying band is a minority of leads rather than most of them.Caps, decay, and a real threshold
The model was validated on an accuracy figure and nobody can tell whether it works.
Accuracy is close to meaningless when conversion is rare, because predicting that nothing converts scores well. Validate by lift instead: rank leads by score, split into deciles, and compare conversion rates across bands. A useful model shows a clear gradient from the top decile downwards.Validation by decile lift
Sales ignore the score because they cannot see why a lead is rated highly.
Show the components, not just the total: which fit attributes contributed and which actions were recorded. A representative who can see that a lead matches the target segment and requested pricing twice this week will act on it. A number with no explanation gets treated as noise, however good the underlying model is.Visible score components
The model was built two years ago and has never been revisited.
Scoring models decay as the product, the market, and the traffic mix change. Revalidate on a schedule, at least annually, by rerunning the decile lift analysis on recent outcomes. A model that no longer separates converting from non-converting leads is worse than no model, because it directs attention with false confidence.Scheduled revalidation
AI-powered CRM with the features you need to close more deals.
Common questions about using HelloGrowthCRM in your industry.
Join small businesses that close more deals with HelloGrowthCRM.
Free Forever • No Credit Card Required
Free Forever • No Credit Card Required
Prefer email? Write to sales@hellogrowthcrm.com