Lead Scoring: How It Works and How to Build a Model

Every sales team has a rep who works leads in the order they arrived, like mail, and a rep who somehow always seems to call the right account first. The second one is not psychic. They either built a scoring model in their head years ago, or they are about to burn out doing manually what the rest of this guide automates.

Lead scoring ranks your leads by how likely each one is to close, using points assigned to attributes like job title, company size and behaviour, so a rep works the best-fit account first instead of whatever landed in the queue most recently. Done well, it turns a pile of undifferentiated leads into a worked list. Done badly, it is a number nobody trusts, everybody quietly overrides, and eventually nobody bothers to look at.

This guide covers what actually gets scored, the difference between a manual points system and a predictive model, and how to set a threshold that matches your own close rates rather than a number that merely feels right.

What Is Lead Scoring?

Lead scoring assigns a numeric value to each lead based on how closely it resembles the accounts that have actually bought from you before. Per Salesforce's own guide, the model comes in two flavours: manual scoring, where points are assigned to attributes by hand, and predictive scoring, where machine learning analyses historical outcomes to weight attributes automatically. Either way, the output is the same thing: a single number that tells a rep where to spend the next hour. The score is an input to a decision rather than the decision itself, and our guide to lead qualification covers the verdict it feeds, and our guide to sales process optimization covers the stage-by-stage view a score is read against.

Why a number beats a rep's instinct

A rep's gut feel about which lead to call next is shaped by whichever one replied most recently, not by which one is actually most likely to close. A score built from real signal removes that recency bias, and it gives a manager something concrete to audit when a rep's pipeline stalls: not "did you work hard," but "did you work the highest-scored accounts first." It also gives marketing and sales a shared unit to argue over, which turns a subjective disagreement about lead quality into a specific, checkable claim about a single number.

A lead needs a score before it needs a rep

Scoring sits between sourcing and outreach. A record becomes a real lead once it has a reason attached to it, and scoring is how that reason gets ranked against every other lead competing for the same rep's time.

What Actually Gets Scored?

Four categories, and a model that only uses one of them is measuring a fraction of the picture, confidently.

Demographic and firmographic fit

Demographic signals describe the person: job title, seniority, decision-making authority. Firmographic signals describe the company: industry, headcount, tech stack. Neither one moves when the lead does something, both are set at the moment the record enters the pipeline, and both should be checked first, because a perfect behavioural signal on the wrong company is still the wrong company.

Behavioural and negative signals

Behavioural signals are the ones that move: a pricing page visit, a demo request, a reply to an email, attending a webinar. These decay, a pricing visit from six months ago should not carry the same weight as one from this morning, so a model that never ages its behavioural score will keep ranking stale interest above fresh interest. How fast each type ages is the part worth getting right before you pick a weight, and our guide to which buying signals decay and how fast sets out what ends each one. Negative signals subtract points rather than add them: a personal email domain on a B2B form, a location well outside your served market, a student job title. Most teams build the positive side of a model and forget the negative side exists.

Manual Scoring vs. Predictive Scoring

Every vendor pitch deck implies predictive scoring is the grown-up version and manual scoring is what teams do before they can afford the good stuff. That is marketing, not evidence. Neither approach wins by default, and the right one depends on how much closed-deal history you actually have to learn from, not on which one sounds more impressive in a demo.

A useful sequence for most teams: launch with a manual model because it works on day one, log every score alongside the deal's eventual outcome, and revisit a predictive model once there is a real base of won and lost deals to train it on. Skipping straight to predictive scoring with little history just means the model is guessing with extra steps.

Most real models end up hybrid, whatever the vendor slide deck implies

In practice, few teams run purely one or the other for long. A common pattern keeps manual rules for firmographic and demographic fit, where the logic is simple and stable, "job title contains VP or Director" does not need machine learning, and layers a predictive component on top for behavioural weighting, where the relationship between a specific action and an eventual close is genuinely harder to eyeball. The manual half stays auditable, the predictive half improves as more outcomes accumulate.

Building a Scoring Model That Reps Trust

The model itself is rarely why scoring fails. Trust fails first, and once reps stop trusting the number, they route around it, usually with a spreadsheet of their own that nobody else can see.

Calibrate the threshold against real close rates

Score a batch of already-closed deals retroactively and look for the point where win rate jumps rather than climbs gradually. Set the qualifying threshold there. A round number chosen because it feels reasonable, 70 out of 100, is exactly as arbitrary as it sounds until it has been checked against real outcomes.

Write the model down somewhere a rep can actually see it

A scoring model that lives only in a marketing automation tool's configuration screen is invisible to the rep who has to trust the number it produces. Publishing the actual weights, in a shared document, in plain language, turns a black-box score into something a rep can argue with constructively, "this account scored high on firmographics but the behavioural signal is six months stale," rather than something they either blindly trust or quietly ignore.

Review the model on a schedule, not after a complaint

A model built once and never revisited drifts as your ICP, your product and your market shift under it. The teams that keep scoring credible review the weights quarterly, on the calendar, rather than only after a rep points out that a high-scored account went nowhere. Our guide to intent providers covers where the behavioural half of a scoring model can source fresher signal than a static list.

Common Lead Scoring Mistakes

Most scoring models do not fail on the math. They fail on a handful of design choices that are easy to make once, forget to revisit, and only notice when a rep quietly stops looking at the number at all.

Weighting every behavioural signal the same

A pricing page visit and a newsletter open are not the same strength of signal, but plenty of models score them identically because both are "engagement." The fix is ranking behaviours by how close they sit to a buying decision, a pricing visit or a demo request well above a passive open, before assigning point values, rather than assigning round numbers and hoping the ranking falls out naturally.

Never checking the model against what actually closed

A model built once from assumptions and never checked against real outcomes drifts silently, because nobody notices a bad score until a rep complains about a specific dead lead months later. The cheap fix is a quarterly pull: take every deal that closed and every one that died, plot them against the score they carried, and see whether the model actually separates the two groups. If it does not, the weights are wrong regardless of how principled they looked on the whiteboard.

Scoring the account but never the timing

A perfect firmographic and behavioural score still misses if the account is not currently in a buying window. A company that matches your ICP exactly but shows no recent signal is a good target for outbound research, not a hot lead today. Folding a freshness component into the score, weighting a signal from this week above an identical one from six months ago, is what separates a model that reflects current intent from one that reflects a static database snapshot.

How Leadalf Scores a Lead

Leadalf scores every signal it finds, on LinkedIn, Reddit, X, the Meta Ads Library and job postings, against the ICP you write down rather than a fixed keyword list, which is what our guide to AI agents for sales covers in more depth. The score is attached the moment a signal is found, not after a rep manually reviews it, so the qualifying decision happens before the record ever hits a human's queue, and only the accounts that clear the bar get enriched and pushed onward.

If your pipeline coverage depends on knowing how many qualified leads a given score threshold actually produces per month, the pipeline coverage calculator turns that into a number you can plan a quarter around.

Frequently Asked Questions