How to Build a Lead Scoring Model From Website Signals (2026)

A lead scoring model is a set of rules that ranks prospects by multiplying two independent judgments: fit (is this the kind of company and person we sell to?) and intent (are they showing buying behavior right now?). The most common failure is collapsing the two into one number — a perfectly-fitting company that visited once in March shouldn't score like a mediocre fit reading your pricing page today. This guide walks through building a defensible model from website signals in an afternoon, when to add negative scoring and decay, and when predictive (machine-learning) scoring actually earns its complexity.

Start with the two-axis skeleton

Score fit and intent separately, act on the combination:

Everything else in lead scoring is detail on top of this grid. The grid also explains why scoring projects fail when marketing owns fit and nobody owns intent: the right-hand column is where revenue lives, and it's built from behavioral signals most teams don't capture.

Scoring fit (the stable half)

Fit criteria come from your closed-won history, not your aspirations. Pull your last 30 wins and score what they share:

• Firmographics: industry, employee range, geography. Example weights: target industry +20, employee count 50-500 +15, US-based +10 (if, like many person-level identification users, your motion is US-centric).
• Role (when known): economic buyer titles +15, influencer titles +5, students and job seekers 0.
• Technographics: runs the CRM you integrate with +10.

Fit scores change rarely — quarterly at most. That stability is what makes the intent half legible.

Scoring intent from website signals (the live half)

This is where most models are starving for data: intent lives in anonymous website behavior, and without visitor identification those signals attach to nobody. Once identification is running (VisiLead does this from $29/mo — companies globally, individual people on US traffic), the signal hierarchy looks like:

• Pricing page visit: +30. The single strongest self-serve signal; nobody reads pricing recreationally.
• Comparison or alternatives page: +25. They're shortlisting — you or a competitor.
• Return visit within 7 days: +20. Frequency beats depth; three short visits in a week outrank one long one.
• Case studies or security/compliance pages: +15. Late-stage validation reading.
• Multiple people from the same company: +25 at the account level. A committee forming is the most underrated signal in B2B; it's visible only if you score accounts, not just contacts.
• Blog-only visit: +3. Educational traffic is future pipeline, not current intent.

Our buying signals guide ranks these in more depth.

The two mechanics that separate real models from decoration

1. Decay. Intent points should expire — a simple version halves behavioral scores every 14 days. Without decay, your "hot" list is an archaeology of everyone who ever binged your blog, and reps learn to ignore it.
2. Negative scoring. Careers page –20, student email domains –15, existing-customer support visits –30, competitor domains identified by your ID layer –40 (or route them somewhere more fun). Negative rules do more for rep trust than any positive weight, because they're what keeps junk off the list.

Rules-based vs predictive lead scoring

Predictive scoring replaces hand-set weights with a model trained on your historical conversions. The honest decision criteria:

• Volume: below roughly a few hundred conversions per quarter, a trained model mostly memorizes noise. Rules win.
• Explainability: a rep will act on "VP visited pricing twice this week" and will not act on "score 87, source: model." Early-stage scoring is a trust-building exercise; black boxes build none.
• Maintenance: predictive models drift with your ICP and need retraining ownership; rules need a quarterly hour.

The pragmatic sequence: run rules for two quarters, use the results to learn which signals actually precede wins, then let a predictive layer (your marketing automation platform's, or a dedicated tool's) challenge the rules — not replace them — once volume justifies it.

Wiring it up

1. Deploy identification so behavioral signals have identities attached — this is the data layer the whole model stands on.
2. Set the routing thresholds with sales, not for them: e.g. fit ≥ 40 and intent ≥ 50 posts to Slack and creates a task; high fit with low intent enters nurture; everything else accumulates quietly. VisiLead's intent scoring with ICP filters covers the fit-plus-behavior mechanics out of the box; the same logic builds in HubSpot or Salesforce with more assembly.
3. Close the loop quarterly. Pull the last quarter's closed-won and closed-lost, check what their scores were the week before the opportunity opened, and adjust the two or three weights that were most wrong. A scoring model that never changes after launch is a monument, not a system.

Frequently Asked Questions

Q: What is a good lead scoring model to start with?

A: A two-axis rules model: fit (firmographics scored from your last 30 closed-won deals) kept separate from intent (website behavior: pricing +30, comparison pages +25, 7-day return +20, with 14-day decay and negative rules for careers pages and students). Route on the combination. It takes an afternoon to build and outperforms most complex setups because reps can see why each lead scored.

Q: What website signals matter most for lead scoring?

A: In rough order: pricing page visits, comparison/alternatives pages, return frequency within a week, multiple visitors from the same company (committee formation), and late-stage validation pages like case studies or security docs. All of these require visitor identification to attach to a scoreable company or person — anonymous sessions can't be scored.

Q: When should I switch from rules-based to predictive lead scoring?

A: When three things are true: you have at least a few hundred conversions a quarter to train on, your rules model has run long enough to give you a baseline to beat, and someone owns retraining. Until then, predictive scoring mostly adds opacity — and an unexplainable score is one sales won't act on.

Q: Should lead scoring happen at the contact or account level?

A: Both, but decisions route at the account level in B2B. Individual scores miss the committee pattern — three mid-level people from one company researching simultaneously is stronger intent than any single visitor, and it's only visible when behavioral scores roll up to the account.

Q: What is a good lead score threshold for sales handoff?

A: Work backwards from capacity, not theory: if your reps can properly work 40 leads a week, set the threshold so roughly 40 leads a week cross it, then tighten as conversion data accumulates. A threshold is a routing valve, not a truth claim. Two refinements that matter more than the exact number: separate thresholds for "alert now" (high fit + high intent) versus "add to queue," and an automatic feedback field where reps mark handoffs as good or bad — that field is your recalibration dataset.

Q: How is lead scoring different from lead grading?

A: They're the two axes of the same grid wearing different names: grading usually refers to the fit judgment (A-D letter grades on firmographic match) while scoring refers to the behavioral intent number. Some platforms use the terms interchangeably, which causes real confusion. Whatever the vocabulary, the operating principle stands — keep fit and intent as separate values and route on the combination, because a hot-scoring bad-fit lead and a cold-scoring perfect-fit lead need entirely different treatment.

Q: How many criteria should a lead scoring model have?

A: Fewer than you think: five to eight fit criteria and five to eight behavioral signals cover almost everything, and each addition past that makes the model harder to explain and recalibrate. The discipline that separates working models isn't criterion count — it's that every criterion traces to closed-won evidence ("8 of our last 10 wins visited pricing twice in a week") rather than plausible-sounding intuition. Start minimal, add only what quarterly reviews prove missing.