Confession: the first lead scoring model I ever built gave points for opening emails. Three opens? Hot lead! Route to sales!
You can guess how that went. Our “hottest leads” were students, competitors, and one very enthusiastic retiree. Meanwhile the actual buyers, who read one email carefully and visited the pricing page once, sat in nurture.
The model wasn’t stupid. It was starving. It scored the only signals it could see, and it couldn’t see the ones that mattered. So let’s build you a model that can.
What is a lead scoring model?
A lead scoring model ranks leads by assigning points for attributes and behaviors, so sales works the highest-scoring leads first. It’s an old idea with a solid track record; the Wikipedia entry on lead scoring traces the fundamentals if you want the formal version.
Points come from two dimensions. FIT: is this company and person our kind of buyer? And INTEREST: are they acting like someone in a buying cycle? Every scoring failure I’ve seen traces back to ignoring one of those two. Fit without interest wastes calls on companies that match the profile but couldn’t care less. Interest without fit (my email-opens fiasco) sends enthusiasm to sales regardless of whether it can ever become revenue.
📌 TL;DR: score fit and interest separately, combine them in a grid, subtract points for disqualifiers, set the threshold with sales, and review against closed deals monthly. And before touching any weights, check that the fields your model reads are actually filled.
The two-axis grid that actually holds up
Score fit and interest separately, then combine them in a grid. It’s the difference between one blurry number and a decision you can act on:
| High interest | Low interest | |
|---|---|---|
| High fit | Route to sales NOW | Nurture with intent (right buyer, not ready) |
| Low fit | Polite self-serve path | Ignore gracefully |
Fit signals live in data: industry, company size, region, role seniority, tech stack. Your ICP (ideal customer profile, the company type you win most) expressed as points. Interest signals live in behavior: pricing-page visits, demo requests, content depth, and, if you use it, third-party intent data showing research happening where you can’t see it.
What types of lead scoring models exist?
Four types cover the field, and real models combine them:
- Demographic and firmographic scoring. Points for who the lead IS: title, seniority, industry, company size. (Firmographic just means company-level attributes.) This is the fit axis.
- Behavioral scoring. Points for what the lead DOES: page visits, event attendance, content depth. The interest axis.
- Negative scoring. Points subtracted for disqualifiers: student titles, personal email domains, regions you don’t serve.
- Predictive scoring. A machine-learning model finds the weights from your historical deals instead of you assigning them. More on when that’s earned below.
So which type should you pick? Wrong question. A working model uses the first three together from day one, and graduates to the fourth when the data supports it.
How do you build a lead scoring model step by step?
Build it in five steps: mine your closed deals, assign points, add negative points, set the threshold with sales, and review monthly. Here’s each one:
- Mine your closed-won deals. What did converting leads have in common, attributes AND behaviors? Those commonalities become your point-earners. Guessing weights is astrology. Measuring them is a model.
- Assign points, small scale. Keep it comprehensible: +15 for ICP industry, +10 for target size band, +20 for pricing-page visit, +25 for demo request. If nobody can explain a score in one breath, it’s too clever.
- Add negative points. Personal email domain, student title, region you don’t serve, competitor domain: subtract. Exclusion points protect sales time better than inclusion points fill pipelines.
- Set the threshold with sales, not for them. The threshold is the score where a lead becomes an MQL (marketing qualified lead, the handoff point to sales). Pick the score that routes a volume sales can actually work, then adjust from results, not opinions.
- Review monthly against outcomes. Compare scores at routing time with what actually closed. Weights that don’t predict get cut. This meeting takes an hour, and it’s the difference between a model and a superstition.
💡 Negative-points tip: start your model with the disqualifiers, not the qualifiers. Ten minutes of listing who should NEVER reach sales (students, competitors, wrong regions) saves more rep time than a month of tuning the positive weights.
A worked example you can copy
Here’s a small point matrix to adapt. The numbers are starting points, not gospel:
| Signal | Points | Why |
|---|---|---|
| ICP industry match | +15 | Core fit signal from closed-won analysis |
| Target size band | +10 | Deals outside it stall in procurement |
| Director level or above | +10 | Junior titles rarely control budget |
| Pricing-page visit | +20 | Strongest self-serve buying signal |
| Demo request | +25 | Explicit hand-raise |
| Personal email domain | -10 | Weak accountability, poor match rates |
| Student or intern title | -25 | Research traffic, not revenue |
| Region not served | -30 | Can’t buy no matter how interested |
Now walk one lead through it. A director at a 300-person ICP-industry company (+15 +10 +10) who visited pricing (+20) scores 55. If your MQL threshold sits at 50, she routes to a rep today. A student (+0 fit, -25) who downloaded an ebook stays in nurture no matter how many emails he opens. Same inbox, radically different treatment. That’s the model doing its job.
Why enriched data decides whether scoring works
Here’s the dependency nobody puts in the scoring tutorials: fit scoring reads fields, and fields are empty by default.
A lead arrives with a name and an email. Industry? Blank. Size? Blank. Seniority? Blank. Your beautiful fit criteria evaluate to zero, not because the lead is bad, but because the record is. The model grades what it can see, and it can see nothing.
That’s why scoring and lead enrichment are one system: enrichment fills the fields at creation, THEN scoring fires, THEN routing. Sequence it the other way and every lead scores as a mystery. The plumbing side of that lives in CRM enrichment, where triggers and refresh cycles keep the scored fields current.
And when scores start feeling random months later? Audit the fill rates on your scored fields before you touch a single weight; my data quality metrics guide shows how. Scoring drift is usually data drift wearing a disguise. Bad fields are expensive everywhere, not just in scoring; a widely cited Harvard Business Review analysis put the economy-wide cost of bad data in the trillions.
What about predictive lead scoring?
Predictive lead scoring lets a machine-learning model find the weights from your historical deals instead of you assigning them. It’s genuinely better, WHEN you have the volume. The model needs enough closed-won and closed-lost examples to learn from, and complete records to learn on.
My honest guidance: under a few hundred closed deals of history, a hand-built point model reviewed monthly beats a starved algorithm. Above that, predictive earns a trial. Judge it like any classifier, with precision and recall against real outcomes, not with how impressive the dashboard looks.
And the same caveat applies as everywhere else in this article: predictive models inherit the data quality of the fields they train on. Garbage fields in, confident garbage out.
🧠 Privacy note: scoring individuals based on their behavior counts as profiling under EU and UK privacy law, and automated decisions about people carry extra duties. The UK ICO's UK GDPR guidance covers profiling in plain language. If you score EU or UK contacts, have someone qualified review the setup.
Common mistakes that quietly break scoring models
Five patterns account for most broken models I’ve audited:
- Scoring activity instead of intent. Opens and clicks measure curiosity. Pricing visits and demo requests measure buying. Weight accordingly.
- Points inflation. Every team adds signals, nobody removes them, and three years later a lead can hit the threshold without a single strong signal. Cap the model at a dozen weights.
- One number for two questions. A single blended score hides whether a lead is wrong-fit or just early. Keep fit and interest visible separately.
- Set-and-forget thresholds. The threshold that worked at 200 leads a month drowns sales at 2,000. Revisit it whenever volume shifts.
- No score decay. A pricing visit from January shouldn’t route anyone in June. Age behavioral points out after 30 to 90 days.
Notice what’s missing from that list? Tool choice. Every one of these mistakes survives a platform migration just fine. Fix the model, not the software.
My opens-era model, honestly
Back in Hamburg, that opens-scoring model of mine ran for five months before anyone checked it against outcomes. The first review meeting was brutal in the best way. Of the 60 “hot” leads we’d routed that quarter, sales had disqualified more than half within one call. And two of our biggest closed deals that quarter had never crossed the threshold at all.
We cut email opens to nearly nothing, tripled the pricing-page weight, and added the negative points we should have started with. The next quarter, the disqualification rate on routed leads dropped visibly and sales stopped joking about “marketing’s hot list.”
One hour a month reviewing outcomes. That’s all it took. That’s all it ever takes.
How we know this (and what to double-check)
This framework comes from building and reviewing scoring models hands-on across several B2B teams, plus the public sources linked above. Honest limits: point values and thresholds are business-specific, so copy the structure here, not the numbers. Decay makes every model drift (US median employee tenure is around four years, so titles and emails keep expiring). And the privacy note above is a starting point, not legal advice.
Frequently asked questions
What is the lead scoring theory?
Lead scoring theory says buying likelihood can be estimated from observable signals: who the lead is (fit) and what they do (interest). Assign points to the signals that historically preceded deals, and the total predicts which leads deserve attention first.
How is a lead score calculated?
Points get added for positive signals (ICP-matching attributes, engaged behaviors) and subtracted for negative ones, with the total compared against a routing threshold. A lead at +15 (industry) +10 (size) +20 (pricing visit) -10 (junior title) scores 35 and routes wherever 35 falls in your bands.
What is an example of lead scoring?
A director at a 300-person ICP-industry company who visited pricing twice scores, say, 55 and routes to a rep today. A student who downloaded one ebook scores -25 and stays in nurture. Same form, same day, opposite treatment.
What is a good lead score threshold?
The one that routes a lead volume your sales team can actually work at quality. There’s no universal number. Start where roughly the top 10 to 20 percent of leads land, then move it based on what sales accepts and what closes.
What is predictive lead scoring?
Predictive lead scoring uses machine learning to derive scoring weights from your historical won and lost deals, instead of humans assigning points. It outperforms manual models once you have a few hundred closed deals of clean history to train on.
How do you improve a lead scoring model over time?
Review scores against real outcomes monthly, cut weights that don’t predict, and keep the scored fields filled through automated enrichment. Most scoring drift is really data drift: fields going stale or empty under a model that hasn’t changed.
It’s time to score what matters
Run one diagnostic before touching any weights: pull last quarter’s closed-won deals and check what they scored at routing time. If your winners weren’t your high scorers, you now know exactly what this month’s review meeting is about.
You’ve got this. Tell me in the comments the silliest thing your scoring model ever rewarded. I’ve already confessed my email-opens era, so the bar is low.
