Predictive lead scoring explained — what it is, what it isn't, and how to do it well

Predictive lead scoring sounds like magic. In practice, it's a regression model trained on historical conversion data. The success or failure of an implementation comes down to data hygiene, not model sophistication.

What predictive lead scoring is

Take your historical lead → conversion data. For each lead, you know:

• Attributes at intake (firmographic: company size, industry, location; behavioral: pages viewed, content downloaded; demographic: title, seniority)
• Whether they converted (became a customer, closed-won deal, etc.)

You train a model (logistic regression, gradient boosting, neural net — doesn't matter much) that learns which combinations of attributes predict conversion. At inference time, you score new leads on the same attribute set and get a probability of conversion.

Leads above a threshold get prioritized by sales. Leads below get nurtured by automation. The bet: focus human attention on the highest-probability leads.

What predictive lead scoring isn't

• It's not magic. The model can only learn from your historical data, so it inherits whatever biases are in that data.
• It's not a substitute for sales judgment. The score is one input; rep judgment is another.
• It's not the same as "ICP fit". ICP fit is about whether you want them as a customer. Scoring is about whether they're likely to become one. A lead can be perfect ICP but score low (slow to convert) or imperfect ICP but score high (fast to convert).
• It's not real-time AI. Most implementations score on intake + on major behavioral events, not continuously.

The data hygiene problem

Predictive lead scoring fails 70% of the time. The reason: garbage-in-garbage-out on training data.

Common training data problems:

1. Inconsistent conversion definition. Some leads marked "won" were actually small one-off deals; some "lost" were qualified-out for non-fit. The model learns noise.
2. Survivor bias. Leads that converted get more complete data (more interactions, more enrichment). Leads that didn't convert get abandoned. Model learns "has lots of data" → conversion.
3. Sample size. Predictive scoring needs ~1,000+ historical conversions to learn meaningful patterns. Most startups have under 100.
4. Time decay. Your business in 2024 may have a different ICP than your business in 2026. Old conversions shouldn't weight the model equally to recent ones.
5. Feature leakage. Some features (e.g. "number of sales team interactions") are only high for leads that converted because sales already reached out to them. The model learns "sales reached out" → conversion, which isn't actionable.

What actually works

The lead scoring approach that works in practice for most B2B SaaS / agency contexts:

Step 1: Manual rule-based scoring (start here)

Before any ML, write down what you think predicts conversion. Common rules:

• Company size 50-500 employees: +20
• Industry in {SaaS, Agency, Marketing}: +15
• Title contains {VP, Director, Head of}: +10
• Downloaded pricing page: +25
• Visited 3+ pages in one session: +10
• Used a competitor in last 12 months: +30
• Gmail address (not business): -10
• Title contains {Student, Intern}: -20

Sum the rules. Score above 50 → priority. Score below 0 → nurture only.

Most teams discover that manual scoring captures 80% of the signal. The remaining 20% is where ML can add value — if the data is clean.

Step 2: Validate against historical data

Take your last 500 leads. Score them manually using the rules. Compare to actual conversion. If the manual rules predict 70-80% of conversions accurately, you have signal. If they predict 20-40%, your rules are weak (or your data is dirty) and ML won't fix it.

Step 3: Add ML as augmentation (not replacement)

If manual scoring works, ML can add a marginal lift by finding interaction effects you missed. Train a simple model on your historical data, use the model output as one additional input alongside the manual rules.

Don't let the model override manual rules entirely. Hybrid scoring is more interpretable, more debuggable, and usually about as accurate.

What to score on

Three categories of inputs:

Firmographic

• Company size (employee count)
• Revenue band
• Industry
• Location
• Funding stage
• Tech stack

Demographic (person-level)

• Title / seniority
• Department
• Tenure at company
• Email domain (business vs gmail)

Behavioral

• Pages visited (count, specific pages)
• Content downloaded
• Email engagement (opens, clicks)
• Form submissions
• Time on site
• Direct-traffic vs paid vs organic
• Repeat visits

The most predictive single signal in our experience: "visited the pricing page within 30 minutes of first session". Strong intent signal, easy to capture, doesn't require enrichment.

The Conduit implementation

Conduit ships a lead scoring engine in the Enrichment module. Both rule-based and ML-augmented:

• Visual rule builder for manual scoring
• Auto-training of a regression model on your historical data (requires 100+ historical conversions)
• Hybrid score: rule_score × 0.7 + ml_score × 0.3 (configurable weighting)
• Score recomputed on intake + on behavioral events
• Threshold-based routing: high score → notify sales, mid score → nurture sequence, low score → no action

What to expect

Well-implemented lead scoring (manual + light ML) typically lifts sales conversion by 15-30%. The lift comes from focus — sales reps spend more time on high-probability leads, less time on low-probability ones.

Don't expect ML to find magic patterns humans miss. ML mostly finds interaction effects ("CEOs at SaaS companies who downloaded the pricing page convert 3× more than the base rate") that humans could find if they looked. The value is automation, not magic.

Pitfalls to avoid

1. Don't make the model a black box. If sales can't explain why a lead is high-scored, they'll ignore the score.
2. Don't re-train constantly. Quarterly is enough. Monthly re-training adds noise.
3. Don't score on too many features. 10-30 features is plenty. 200 features overfits.
4. Don't conflate score with intent. Score is a probability across the population. Intent is a specific signal from a specific lead.
5. Don't skip the manual rules. They're the interpretability layer.

Bottom line

Predictive lead scoring works when you have clean training data, sensible features, and a hybrid rule + ML approach. It fails when you treat it as a magic ML solution to a data problem.

Conduit's scoring engine is a good starting point if you want to skip the build. See the Enrichment module or start a 14-day trial.