Your CRM has dozens of fields on every lead. Most of them don’t matter.
When a research team trained a Gradient Boosting model on 4 years of real B2B CRM data with 16,600 cleaned lead records from Microsoft Dynamics, they found that a small set of fields drove the vast majority of conversion predictions. The model hit 98.39% accuracy. And the features that mattered most weren’t the ones most scoring systems prioritize.
Here’s what the feature importance analysis revealed.
1. Lead source
The single strongest predictor of conversion was where the lead came from.
Not what the lead did after arriving. Where they originated. Referral, inbound, event, partner, paid: the source channel carried more predictive signal than almost any behavioral metric.
This makes sense if you think about it from an ops perspective. Source encodes intent quality at the point of entry. A lead from a referral partner and a lead from a cold list have fundamentally different conversion profiles, regardless of what they do next.
Ops takeaway: If you’re not tracking source granularly and consistently in your CRM, you’re missing your most predictive field.
2. Reason for state
This field captures why a lead is in its current status (why it was qualified, disqualified, or stalled). It’s a field most teams treat as a dropdown that gets filled inconsistently.
The model found it to be the second most important predictor. It acts as a compressed summary of sales judgment, the “why” behind where the lead sits in the pipeline.
Ops takeaway: Enforce consistent values in this field. Make it required. The data quality investment pays back directly in prediction accuracy.
3. Lead classification
How the lead was categorized at intake: its segment, tier, or type. This is structural metadata that gets assigned early and rarely changes.
Classification interacts with source and product to create a compound signal. A “strategic” classification from an inbound source on a high-value product line is a different animal than an “unknown” classification from a purchased list.
Ops takeaway: Standardize your classification taxonomy. Reduce free-text entry. The model can only use this field if the values are consistent.
4. Product
Which product or solution the lead expressed interest in. Different products have different conversion profiles, sales cycles, and buyer personas.
The model uses product as a segmentation layer. Leads interested in one product line may convert at 3x the rate of another, independent of all other factors.
Ops takeaway: Make sure product interest is captured at the lead level, not just the opportunity level.
5. Number of responses
Total interactions between the lead and the company. This is the behavioral signal that traditional scoring systems try to capture with point totals, but here it’s one input among many, not the entire model.
The research found that “number of responses reflects the number of interactions with the company” and contributes meaningfully to prediction, but it ranks fifth, not first.
Ops takeaway: Track interaction count, but don’t over-index on it. Engagement matters, but source and classification matter more.
6. Account type
The segmentation of the account (commercial, educational, research, and so on). This is firmographic data that describes who the buyer is, not what they did.
Account type creates natural conversion cohorts. Commercial accounts in a B2B software company convert differently than educational institutions, and the model captures this without manual rule-writing.
Ops takeaway: Segment your accounts cleanly. If account type is blank on half your records, you’re degrading your model’s ability to score accurately.
7. Interest level
The lead’s stated level of purchase intent (immediate, near-term, exploring, and so on). Self-reported intent is noisy, but it still carries signal.
It ranked seventh, meaningful but not dominant. The model uses it as a confirming signal rather than a primary driver.
Ops takeaway: Capture it, but don’t build your scoring system around it alone. Stated intent is weaker than demonstrated behavior and structural fit.
What’s notably absent
Page views, email opens, content downloads, time on site: the behavioral signals that traditional scoring models weight most heavily didn’t make the top of the feature importance list.
This doesn’t mean they’re useless. It means they’re less predictive than the structural and contextual fields that describe who the lead is, where they came from, and what they’re interested in.
Traditional scoring over-indexes on activity. ML scoring reveals that identity and context outperform behavior as conversion predictors.
The data quality implication
Every field on this list has one thing in common: it’s only as good as the data behind it. If lead source is inconsistently tagged, if classification is free-text, if account type is blank 40% of the time, the model can’t use what isn’t there.
The researchers noted that “18.8% of ‘Title’ records were outliers, a data quality issue” requiring careful handling. Real CRM data is messy. The scoring model is only as strong as the data hygiene underneath it.
The best investment most ops teams can make isn’t a new tool. It’s enforcing consistent entry in the 7 fields that actually predict conversion.
Reference: Gonzalez-Flores, L., Rubiano-Moreno, J., & Sosa-Gomez, G. (2025). The relevance of lead prioritization: a B2B lead scoring model based on machine learning. Frontiers in Artificial Intelligence, 8, 1554325.