Lead scoring assigns each lead a number that reflects how likely they are to buy. It exists for one purpose, and it is an important one: deciding who your team calls first when it cannot call everybody.
If you get twenty leads a month, you do not need lead scoring, you need to call them. It becomes necessary when volume exceeds the team's capacity, which is exactly the point where most companies start working leads in the order they arrived, which is the worst possible criterion.
The two halves of any model
Every lead scoring model that works combines two different things, and confusing them is the most common mistake.
Explicit scoring: who they are. Company and person attributes. Sector, size, country, job title, the technology they run. It answers "do I want them as a customer?" and you can know it before they do anything at all.
Implicit scoring: what they do. Behaviour. Three visits to the pricing page, the last four emails opened, a demo requested, a colleague invited into the trial. It answers "are they interested, and is this the moment?".
A perfect fit with zero activity is a prospecting target. A very active lead who does not fit is noise, and often a student, a competitor or somebody looking for a job. Only the two dimensions together tell you who to call today.
A worked points table
Treat this as a starting point to adapt, not a standard. The numbers should come from your own data as soon as you have some.
| Signal | Type | Points |
|---|---|---|
| Decision-making title (director, CxO, owner) | Explicit | +20 |
| Company inside your size band | Explicit | +15 |
| Target sector | Explicit | +15 |
| Country you sell into | Explicit | +10 |
| Requested a demo | Implicit | +30 |
| Visited pricing twice or more | Implicit | +20 |
| Opened and clicked the last two emails | Implicit | +10 |
| Downloaded bottom-of-funnel content | Implicit | +10 |
| Registered for a webinar | Implicit | +5 |
| Personal email address | Explicit | -10 |
| Unrelated role (intern, student) | Explicit | -20 |
| Company far too small or far too large | Explicit | -15 |
| Competitor | Explicit | -50 |
| No activity in 60 days | Decay | -15 |
Two items in that table get ignored almost everywhere, and they are the ones that improve a model most:
Negative scoring. Subtracting matters as much as adding. Without negative points, anyone who clicks enough eventually looks like a buyer.
Decay. Interest cools. Someone who checked pricing four months ago and never came back is not the same lead as someone who checked yesterday. Without decay, your hot list fills up with people who already bought from somebody else.
Thresholds, where the decisions actually happen
The score alone does nothing. The bands do:
- 0 to 40: nurture. Automated content, no sales time.
- 41 to 70: MQL. Marketing signs off. Enters outreach sequences.
- 71 and above: SQL. A rep calls, and calls first.
Set the thresholds by looking backwards. Take the customers you closed last quarter, work out what they would have scored, and put the cut where most of them sat. If the cut excludes real customers you actually won, the cut is wrong, not them.
Building one in five steps
1. Define a good customer from data
Not from intuition. Take your last twenty or thirty closed-won customers and find what they share: sector, size, the title of whoever signed, the channel they came through. Do the same for the ones you lost and the ones who churned at six months. Your model is already written there.
2. Pick signals you can actually measure
A signal that is not in your CRM or marketing tool does not exist. Start with the five or six you have reliably and add more later.
3. Assign points and thresholds
Start simple and round. A ten-signal model the team understands beats a forty-signal model nobody can explain.
4. Agree in writing what happens in each band
Who calls, how quickly, on which channel. A score with no action attached is decoration on a contact record. The usual agreement: SQL contacted within 24 hours, MQL into an automated sequence, everyone else nurtured.
5. Review it quarterly
Markets move, your product moves, and your best customers from last year may not be this year's.
How to know whether the model works
A scoring model is validated with a table, not a feeling. Measure conversion to customer by band:
| Band | Leads | Customers | Conversion |
|---|---|---|---|
| 71 and above | 120 | 18 | 15% |
| 41 to 70 | 340 | 20 | 5.9% |
| 0 to 40 | 900 | 9 | 1% |
The only thing that matters is that conversion climbs clearly as the band rises. If the top band converts like the middle one, your model separates nothing and your team would do just as well calling at random. It is an uncomfortable check and it is the only one that counts.
Watch two more numbers: the share of SQLs that sales rejects (past 20% your thresholds are too loose) and how many closed-won customers never reached SQL at all (if there are many, the model is blind to something).
Manual or predictive
A manual model is the points table above: rules written by people. It is transparent, explainable in a meeting and fixable in five minutes. With fewer than a few hundred closed deals it is the only sensible option, because there is not enough data for anything else.
A predictive model lets a system find the patterns in your history. It needs volume and clean data, and in exchange it spots combinations nobody would have written by hand. Its weakness is opacity: when it tells a rep to call somebody who does not look like a fit, the rep does not call.
The sane order is to start manual, measure conversion by band, and move to predictive when you have the history and the manual model has run out of room.
Mistakes that ruin a model
- Scoring behaviour only. You end up with a list of very active browsers.
- Scoring fit only. You end up with a list of ideal companies that are not interested yet.
- Never subtracting. Everything rises, nothing falls, and in six months everyone is an SQL.
- Changing weights weekly. Without stability you cannot tell whether the model improved or merely changed.
- Building it without sales. Reps know which leads were good because they spoke to them. A model written by marketing alone is a model sales does not believe, and a model nobody believes does not get used.
- Applying it to incomplete data. If 40% of your leads have no sector or size, you are not scoring, you are penalising people who left the form blank. Enrich first.