Why most lead scoring fails
Lead scoring has a 54% adoption rate among B2B organizations in 2026. That sounds healthy until you ask the follow-up question: of those who have scoring, how many sales teams actually use it to prioritize their day?
The answer is far lower. And the reason is almost always the same: the model was built by marketing or ops in isolation, without input from the people who work the leads. Reps tried it, saw a few "hot" leads that were clearly garbage, and went back to their gut.
This is a trust problem, not a technology problem. And it's solvable.
The two types of scoring
Before building anything, understand what you're actually scoring. These are different problems that require different models:
Fit scoring (who they are)
Does this lead match your ideal customer profile? This is static — it doesn't change based on behavior. It's based on firmographic and demographic attributes:
- Company size (employees, revenue)
- Industry and vertical
- Technology stack
- Funding stage
- Geography
- Job title and seniority
- Department
Fit scoring answers: "If this person raised their hand, would we want to talk to them?"
Intent scoring (what they're doing)
Is this lead showing buying behavior? This is dynamic — it changes daily based on actions:
- Website visits (pricing page, case studies, documentation)
- Content downloads (especially bottom-of-funnel content)
- Email engagement (opens, clicks, replies)
- Product usage (for PLG companies)
- External signals (job changes, funding, hiring)
Intent scoring answers: "Is this person actively evaluating a solution like ours right now?"
The model needs both. High fit + low intent = nurture. Low fit + high intent = disqualify. High fit + high intent = route immediately.
Step 1: Start with closed-won analysis
Don't start by guessing which signals matter. Start with data.
Pull your last 50-100 closed-won deals and analyze them for patterns:
- What company sizes closed? What was the median, not just the range?
- Which industries had the highest win rate?
- What titles were the economic buyers?
- How many stakeholders were involved?
- What content did they engage with before becoming an opportunity?
- What was their first touch? Their last touch before requesting a demo?
- Were there external signals (funding, hiring, job changes) in the 90 days before they entered pipeline?
Do the same analysis for closed-lost deals and for leads that never converted. The contrasts reveal your scoring criteria.
Example findings
A mid-market SaaS company selling to sales teams might find:
| Signal | Closed-Won Rate | Closed-Lost Rate | |--------|----------------|-----------------| | 50-200 employees | 34% | 18% | | Series A-B funded | 41% | 15% | | VP/Director title | 38% | 12% | | Visited pricing page 2+ times | 52% | 8% | | Champion job change | 47% | N/A |
Those numbers tell you exactly what to score and how to weight it.
Step 2: Keep the model simple
The most common scoring failure is over-engineering. A model with 40 weighted attributes that nobody can explain will never earn trust.
Start with 8-12 signals maximum. You can always add complexity later. For each signal, assign points on a simple scale:
Fit signals (0-50 points):
- Company size in ICP sweet spot: +15
- Target industry: +10
- Decision-maker title: +10
- Right funding stage: +10
- Uses complementary tech (e.g., Salesforce if you integrate with SF): +5
Intent signals (0-50 points):
- Pricing page visit: +15
- Case study / ROI content: +10
- Multiple sessions in 7 days: +10
- Form fill (demo request): +25
- Product signup (PLG): +20
- Champion job change into target account: +20
Negative signals (deductions):
- Personal email (gmail, yahoo): -10
- Company too small (<10 employees): -20
- Student or intern title: -25
- Competitor domain: -50
- Unsubscribed from emails: -15
Thresholds:
- 0-30: Cold. Nurture only.
- 31-60: Warm. Marketing qualified — add to targeted campaigns.
- 61-80: Hot. Route to SDR for follow-up within 4 hours.
- 81-100: On fire. Route to AE immediately. SLA: 15 minutes.
Step 3: Build it where reps can see it
A score that lives in a marketing automation platform and never surfaces in the CRM is useless. The score needs to be visible where reps work.
In Salesforce/HubSpot:
- Score displayed on the lead record (prominently — not buried in a field nobody checks)
- Score visible in list views and queue sorting
- Score change triggers included in activity feed
- High-score alerts delivered via Slack or email in real-time
In routing logic:
- Score determines routing priority (not just round-robin)
- High-score leads jump the queue
- Low-score leads get routed to automated nurture instead of a rep
In reporting:
- Score distribution over time (are you attracting higher-quality leads?)
- Score-to-conversion correlation (does a higher score actually predict closing?)
- Score accuracy by segment (does the model work equally well for enterprise vs. mid-market?)
Step 4: The calibration loop (this is where most people stop)
Building the model is 30% of the work. Calibrating it is the other 70%. Here's the loop:
Monthly review (30 minutes)
Pull two lists:
- High-score leads that didn't convert. Why? Were they genuinely qualified but churned in the process? Or did the model overweight a signal that doesn't actually predict buying?
- Low-score leads that did convert. What signals did the model miss? Is there a new pattern emerging that isn't being captured?
Adjust weights based on what you find. Small adjustments — 5 points here, 10 points there. Don't overhaul the model monthly.
Quarterly review (2 hours)
- Re-run the closed-won analysis with fresh data.
- Check if ICP is shifting (new industries closing, company size distribution changing).
- Evaluate new signals that weren't being tracked.
- Validate thresholds: Is "Hot" actually hot? What's the conversion rate at each tier?
Rep feedback (ongoing)
The fastest way to improve the model: ask reps. "When the system tells you a lead is hot, are you glad you called them?" If the answer is consistently no for a particular segment, the model is wrong for that segment.
Build a simple feedback mechanism — even a thumbs up/down on routed leads — and use it to train the model.
Step 5: The AI layer (optional, not required)
After you have a working rule-based model with a calibration loop, you can layer in machine learning for behavioral scoring. ML is good at:
- Detecting non-obvious patterns in website behavior
- Identifying sequences of actions that predict conversion
- Adapting to seasonal shifts without manual intervention
- Scoring product usage patterns (for PLG)
ML is bad at:
- Working with small datasets (you need hundreds of closed-won deals minimum)
- Explaining why a score is what it is (black box problem)
- Handling new segments with no historical data
Start with rules. Add ML when you have the data and the trust. A transparent rule-based model that reps trust will outperform a sophisticated ML model that reps ignore.
Common failure modes
Building in a vacuum. If you built the model without interviewing reps about what "good" looks like to them, start over.
Scoring everything equally. A whitepaper download is not the same signal as a pricing page visit. Weight accordingly.
Never decaying scores. A lead who visited your pricing page 6 months ago is not the same as one who visited yesterday. Scores should decay over time if no new activity occurs.
No negative scoring. Without deductions, your model will surface students, competitors, and job seekers alongside real buyers.
Threshold too low. If everything is "qualified," nothing is. Be willing to let the majority of leads stay in nurture. That's the model working, not failing.
No feedback loop. A static model degrades over time as your market, product, and ICP evolve. Plan for ongoing calibration from day one.
Getting started this week
If you have no scoring today, here's the minimum viable version:
- Pull 30 closed-won deals. Note the 5 most common firmographic traits.
- Pull 30 closed-lost deals. Note what was different.
- Create a formula field in your CRM with 5-7 weighted criteria.
- Set one threshold (score > X = route to rep immediately).
- Run it for 30 days. Check conversion rate above and below the threshold.
- Adjust and expand.
You'll have a working model in a week and meaningful calibration data in a month. That's all it takes to get started — the sophistication comes later.