How to Build a Lead Scoring Model from Closed-Won Data
Most lead scoring models are built in a conference room.
Someone from marketing and someone from sales argue about whether a demo request is worth 20 points or 25. A whitepaper gets 10 because it feels like a 10. The numbers go into the CRM and route leads for the next two years.
They never get checked against a single closed deal.
The problem with scoring by opinion
Those opinions might be right, but they've never been tested.
The answer to "what does a good lead look like?" is already sitting in your CRM. It's in the deals you won and the deals you lost, and it rarely gets pulled.
So the score rewards whatever the room believed in. That tends to be downloads, because they're easy to count.
If sales ignores your MQLs, check the score first.
How to build a lead scoring model from closed-won data
How I'd do it: five decisions, in order.
1. Look at what happened
Pull closed-won deals from the last 12 to 18 months, then closed-lost deals and leads that never went anywhere. Thirty to sixty deals is plenty to start, and fifteen is still better than a guess.
For each one, write down the buyer's title, company size, industry, lead source, and what they did before the deal opened.
2. Split fit from intent
Fit is who they are. Industry, size, role, tech stack, region. It doesn't change based on what they click.
Intent is what they do. Pricing visits, demo requests, event attendance, return visits.
A perfect-fit account doing nothing is a future conversation. A poor-fit contact clicking everything is often a student, a competitor or someone job hunting.
Score them separately. A lead should need both to reach sales.
3. Find what separates wins from losses
Compare the two groups. You're looking for traits that show up a lot in wins and rarely in losses.
A trait in 70% of wins and 15% of losses earns points. A trait spread evenly across both earns nothing, no matter how strongly someone feels about it.
Do the same for intent. The answer is often a surprise. A second pricing page visit can predict a deal far better than the first download everyone loves to score.
The trait everyone agrees on is the one most worth checking.
4. Set the threshold from the data
Don't pick 100 because it's tidy. Look at where your wins would have scored, and put the MQL line where it catches most of them without flooding sales.
Then run the model backward against last year. How many wins would it have flagged? How many flagged leads closed? If it misses half your wins, the weights are wrong. Fix them before a live lead ever sees the model.
5. Let scores fade
Someone who was busy in March and silent since June is not still hot. Without decay, scores only go up, and your MQL list fills with people who crossed the line months ago. Subtract points after 30 days of silence, and give leads that stall a path back to nurture.
A lead scoring model example
A simple model looks like this when the analysis is done. The points are illustrative. Yours come from your own wins and losses.
| Criteria | Type | Points |
|---|---|---|
| Title in the buying group that shows up in most wins | Fit | +15 |
| Company size where most wins land | Fit | +15 |
| Industry that converts above average | Fit | +10 |
| Requested a demo or booked a call | Intent | +40 |
| Visited pricing twice in 14 days | Intent | +25 |
| Two or more people from the same account engaged | Intent | +10 |
| No activity in 30 days | Decay | -10 |
| Student, competitor or personal email | Fit | -50 |
Cap each category. Twenty downloads shouldn't outscore a perfect-fit buyer who asked for a demo.
Build it in a spreadsheet first
One tab for the win/loss data. One for criteria and points. One that scores a batch of old leads so you can see the results before anyone else does.
When the backtest looks right, move it into the CRM.
If you're on HubSpot, the built-in lead scoring tool needs a Professional or Enterprise subscription. On Starter, keep the score in a custom number property and update it on a schedule with an automation tool. It's less elegant and works fine.
Don't automate it until the spreadsheet version has survived a month.
Keep retuning it
A score built on closed deals keeps getting better, as long as someone retunes it when new deals close. If the argument is still about what counts as an MQL, settle the definition first and score after.
Want a second set of eyes?
Book a Health Check
Two weeks, fixed price. You get a ranked list of what's distorting your pipeline data, from the CRM sync to lifecycle stages, and how to fix each one.
Book a fit callNot ready to talk? Run the free audit first.
Notes from the Field.
Every other week, one broken thing and how I’d fix it.
One email every other week · unsubscribe anytime · Privacy policy