Lead Scoring Explained: Build a Model Your Sales Team Uses

Lead scoring is the practice of putting a number on a contact or an account so that a limited amount of sales attention lands on the leads most likely to convert. That is the entire purpose. A score does not forecast revenue, it does not replace qualification, and it is not a report. It is a queue order.

The problem it solves appears at a specific size. With thirty leads a month, an SDR reads all of them and no model beats their judgement. With three hundred, the SDR works the top of the list and the bottom is never touched — and without a score, the top of the list means whatever arrived most recently. Scoring replaces recency with something closer to likelihood.

Two things follow. First, a score is only worth building when there are more leads than capacity to work them; otherwise you are ranking a queue that gets fully worked regardless. Second, the model has to be legible to the people using it. An SDR who cannot see why a lead scored 82 will quietly go back to calling the newest arrivals, and you will not discover it for a quarter.

Every model draws on two kinds of input. Explicit signals are things stated outright by the lead or a data source — attributes. Implicit signals are things the lead does — behaviour. Both matter, and they fail in different ways.

The usual mistake is weighting implicit signals by volume. Ten blog visits is not intent, it is a subscriber. One pricing page visit followed by a returning session two days later is intent, from far fewer events. Weight an action by what it costs the person and what it implies about where they are in a decision, not by how often it fires.

Handle missing explicit data honestly too. A blank company size is not a small company. Score it as unknown and let enrichment fill it in later, or you will systematically down-rank everyone who met a short form.

Collapsing fit and intent into a single total is the most common way a model goes wrong. They answer different questions, and each combination calls for a different action.

Fit asks whether this is the kind of company you can serve and sell to: size, sector, geography, budget, and whether they have the problem you actually fix. Intent asks whether they are in motion right now: what they looked at, what they asked for, how recently.

Keep them as two scores and the quadrants tell you what to do. High fit with high intent is a call today. High fit with low intent is a nurture track, not a rejection — the account is still worth having next quarter. Low fit with high intent is the dangerous quadrant: an enthusiastic lead who will not close or will churn, and who will consume an SDR’s week if a blended score floats them to the top. Low fit with low intent is a suppression, not a queue entry.

A single number cannot separate the second case from the third, because a 40 built from fit and a 40 built from intent look identical in the list. If your tooling insists on one number, at least store both components and show the split next to it.

Start manual and small. Fifteen rules you can audit will outperform a black box you cannot, and you can assemble them in an afternoon.

1. Define the target. Not a good lead — something countable, such as a held discovery call, or a closed deal within ninety days.
2. List candidate signals you can capture today for most leads. A signal present on a tenth of records is unusable however predictive it looks.
3. Look backwards. Take your recent won and lost deals and check, signal by signal, how the conversion rate differs when it is present. Anything that barely moves it gets dropped.
4. Assign whole numbers on a fixed scale — for instance 0 to 100 for fit and 0 to 100 for intent. Round weights only. Precision you cannot justify is noise.
5. Test on held-out history before it touches the live queue: would this model have ranked last quarter’s won deals near the top?
6. Write the rules down somewhere the sales team can read them without asking you.

Whatever tool you use, the scoring rules and your lead generation sources belong in the same system: a score computed on fields the sourcing step never filled in is just a default value wearing a number. Growmindr scores leads as they enter the workspace and keeps the contributing rules visible on the contact record, so an SDR can see what produced a number before deciding how much to trust it.

Most models only add. That is why, given enough time, everything scores high and the ranking stops separating anything from anything else.

Subtract for disqualifiers you have evidence for: a competitor domain, a personal or student address on a business product, a country you cannot service, a role with no part in the decision, an existing customer sitting in an open escalation. Prefer a real subtraction to a hard exclusion wherever the lead might matter later — a negative keeps the record scoreable if circumstances change, an exclusion hides it forever.

Then add decay, because intent is perishable. A pricing page visit is a strong signal on the day, a weak one three weeks later, and no signal at all a quarter on. Without decay, someone who researched you in March and vanished still outranks someone active this week. A simple rule is enough: intent points expire after a fixed window, or halve every thirty days. Fit points should not decay — the company is still the same size — but they should be refreshed whenever enrichment updates the record.

There is a quick test for whether decay is set correctly. Pull the top twenty leads by score and check when each last did anything. If several have been silent for months, your decay is too slow.

A score with no agreed action attached to it is decoration. Before launch, agree the tiers and the response each one gets, in writing, with the sales team in the room.

Set the A threshold by capacity, not by aesthetics. If two SDRs can work sixty leads a week between them, the threshold belongs wherever roughly sixty leads a week land — not at 80 because 80 is a round number. Revisit it whenever headcount or lead volume changes, because a threshold tuned for one team size breaks silently for another. Too low and tier A becomes a backlog nobody clears; too high and your best leads sit unworked in tier B while the queue looks healthy.

A lead scoring model is fitted to a moment: your pricing, your ideal customer, your channels, the market. All four move, and the model does not notice. Recalibration is routine maintenance, not an admission that the first version was wrong.

Review on a schedule — quarterly is a sane default — and check three things.

• Discrimination. Bucket the last quarter’s leads into score bands and compare conversion per band. The bands should separate cleanly. If the 80s convert like the 50s, the model has stopped ranking and is only assigning numbers.
• Drift in the mix. If the share of leads above threshold has doubled with no matching change in sales outcomes, something upstream moved — a new channel, an edited form, a broken tracking script — and the score is now measuring that instead.
• Signal availability. Rules die quietly when a field stops being collected or an integration breaks. A signal now missing on most records subtracts accuracy while still looking healthy in the rule list.

Trigger an off-schedule review whenever you change pricing, enter a new segment, or open a channel with a different visitor mix. And take the qualitative signal seriously: when sales starts saying the scores feel wrong, they are usually right, and they usually noticed a quarter before the numbers did.

Broken models tend to fail in a handful of recognisable ways.

• Scoring engagement instead of buying. Newsletter opens and blog reads correlate with interest in your content, not with budget or authority. They inflate the scores of people who will never buy.
• One model for two products. If you sell to two different buyers, you need two models. A blended one is wrong for both, and wrong in opposite directions.
• Person scores with no account context. Three mid-level people from the same company each scoring 45 is a stronger signal than any one of them alone, and a person-only model cannot see it.
• Internal and bot traffic. Your own team and automated scanners generate page views. Exclude office ranges and known crawlers, or your highest-scoring lead will turn out to be a colleague.
• A model nobody explained. If sales was not involved in setting the weights, they will not work the queue, and you will spend your time defending a number instead of improving it.

None of these are visible in the model itself. They surface as sales quietly ignoring the score, which is the single most reliable indication that something underneath is wrong.