ICP Scoring Decay: Why Qualified Leads Go Cold and How to Refresh Your Model

Your ICP Score Is Lying to Your Reps

Most B2B teams build a lead scoring model once, ship it, and then act surprised 18 months later when reps stop trusting it. The "85" that used to mean "call within an hour" now means "probably a student researching for a paper." MQL-to-SQL conversion rates quietly drop. Sales blames marketing. Marketing blames data. Nobody blames the model.

That's ICP scoring decay — and it's the single biggest reason qualified leads turn into dead leads inside your CRM.

The pattern shows up in every diagnostic we run. According to a widely cited B2B benchmark, roughly 79% of marketing leads never convert to sales, and only about 25% of generated leads are qualified enough for direct sales engagement. Yet fewer than half of B2B companies use any structured scoring framework at all — and of the ones that do, most haven't recalibrated their model in over a year.

This piece breaks down why scoring models decay, how to spot it before it tanks pipeline, and the framework we use to rebuild scoring systems that actually predict revenue.

What ICP Scoring Decay Actually Means

ICP scoring decay is the gradual loss of predictive accuracy in your lead scoring model. The score still produces a number. The number just stops correlating with closed-won revenue.

There are four mechanisms behind it:

1. Firmographic drift — Your ICP itself has shifted. You've moved upmarket, expanded to new verticals, or killed an underperforming segment. The model still rewards companies that fit your 2023 ICP.
2. Technographic staleness — The tech stack signals you weighted (uses Salesforce, runs Marketo, has Segment installed) are now table stakes or, worse, irrelevant to your wedge.
3. Behavioral signal inflation — "Visited pricing page" used to mean intent. Now it means a competitor doing recon or an AI scraper crawling your site.
4. Negative signal absence — Most models only score positive attributes. They don't subtract points for disqualifiers like wrong geo, student domains, freemium-only behavior, or 6-month repeat tire-kickers.

When even one of these breaks, your high scores stop mapping to real opportunities. When all four break — which is the normal state of any model older than 12 months — your scoring system is actively misallocating your reps' time.

The Symptoms: How to Know Your Model Has Decayed

Before you rebuild, confirm the diagnosis. These are the signals we look for when running a GTM Audit:

• MQL-to-SQL conversion rate has dropped 20%+ over the last two quarters without any obvious top-of-funnel change.
• Reps have started ignoring scores. They've built their own shadow qualification system in a Notion doc or Slack channel. This is the loudest signal — and the most ignored.
• High-scoring leads have a worse close rate than mid-scoring leads. This inversion is the smoking gun. It means your model is selecting for the wrong attributes.
• Sales is closing deals that scored under 40. Pull a sample of last quarter's closed-won deals and look at what they scored when they came in. If a meaningful chunk scored below your "MQL" threshold, your model is missing real buyers.
• Time-to-disqualification is rising. Reps are spending more time on leads before discovering they're not a fit. The score isn't doing its job upfront.

If two or more of these apply, your model is decayed. Stop tuning it. Rebuild it.

The Refresh Framework: Four Layers That Actually Predict Revenue

Here's the structure we use when rebuilding scoring models for B2B clients. It maps to the way modern scoring tools — Marketo, HubSpot, Clay, Apollo enrichment layers — actually combine signals.

Layer 1: Fit Score (Firmographic + Technographic)

This is the "are they the right kind of company" layer. It should be the foundation, not an afterthought.

Build it from your closed-won data, not your wishlist. Pull the last 18 months of won deals and identify the attributes they actually share:

• Employee count band (be specific — "50-200" not "SMB")
• Revenue band where verifiable
• Industry / sub-industry (use NAICS or your own taxonomy)
• Geography (down to country, sometimes region)
• Technographic stack signals — but only ones that genuinely indicate buying readiness for your category
• Funding stage / recent funding events for venture-backed ICPs

Then build the inverse: a negative fit list. These auto-disqualify or push the score down:

• Wrong geo
• Sub-threshold employee count
• Competitor domains
• Personal email domains (when irrelevant to your motion)
• Industries you've explicitly killed

Weight fit at roughly 50–60% of total possible score. Fit should dominate intent for the simple reason that intent without fit is noise.

Layer 2: Intent Score (Behavioral + Engagement)

This is where most models over-index. Visited a page, opened an email, downloaded a whitepaper — every action gets points, and the points pile up until everyone scores high.

The fix: weight intent by specificity and recency.

• High-intent actions (pricing page, demo request, comparison page, security/compliance docs) get heavy weights but decay to zero within 14–30 days.
• Medium-intent actions (multiple blog views in a single session, webinar attendance, ROI calculator use) get moderate weights and decay within 30–60 days.
• Low-intent actions (single blog view, social engagement, newsletter open) get small weights and decay within 7–14 days. Honestly, most of these shouldn't score at all.

The decay rule is non-negotiable. A lead that hit your pricing page nine months ago is not a hot lead. They are a cold lead with old data attached.

Layer 3: Third-Party Intent Signals

This is the layer that separates 2024 scoring models from 2026 scoring models. Pull in:

• Bombora / G2 / TrustRadius intent data
• Job posting signals (hiring for roles that imply your category)
• Funding announcements
• Leadership changes in your buyer personas
• Tech stack changes detected via Clay, BuiltWith, or HG Insights

These signals are powerful because they're external to your own marketing funnel. They tell you something is happening at the account that your nurture stream didn't cause.

Weight third-party intent at 15–25% of the total score, depending on how reliable your data sources are for your category.

Layer 4: Negative Signals and Decay Rules

This is the layer almost everyone skips. It's also the layer that prevents your model from decaying in the first place.

Build explicit rules that:

• Subtract points for negative behaviors (unsubscribed, marked spam, bounced, replied "not interested")
• Cap scores for repeat tire-kickers (more than 3 demo requests with no progression, more than 6 months in MQL stage without advancement)
• Auto-expire scores after a defined window — typically 60–90 days of no new activity reverts the lead to a baseline fit-only score
• Reset on disqualification so a rep marking a lead "not a fit" actually changes the scoring state

Without a decay layer, scores only go up. That's not a scoring system — it's an accumulation system.

Implementing the Refresh in HubSpot or Marketo

The architecture matters as much as the logic. We've cleaned up too many models where the math was reasonable but the implementation was broken — properties firing twice, workflows competing, score updates happening on import instead of on behavior.

A clean implementation looks like this:

• Fit score and intent score live as separate properties. Don't blend them into one number. Reps and ops both need to see them independently.
• All scoring logic runs through one workflow or one set of HubSpot scoring properties. No parallel systems. No "marketing score" and "sales score" that don't talk to each other.
• Decay logic runs on a scheduled basis — daily for behavioral decay, weekly for fit refresh from enrichment, monthly for full model recalibration checks.
• Lead routing thresholds are documented and version-controlled. When you change a threshold, you log it. Otherwise you can't tell six months later whether conversion dropped because of the model or because of the market.

If your CRM was set up before you had a real scoring strategy — or if your scoring is duct-taped across HubSpot, Marketo, and a Google Sheet — the underlying HubSpot Architecture probably needs work before any model refresh will hold up.

Tying Scoring to Routing and Outbound

Scoring is only valuable if it changes what happens next. The refresh should change three downstream systems:

1. Lead routing thresholds Define tiers based on the new model: Tier 1 (fit ≥ 70 + intent ≥ 50) goes to AE same-day. Tier 2 (fit ≥ 70, intent < 50) goes to SDR nurture. Tier 3 (fit < 70 regardless of intent) goes to marketing nurture or auto-disqualifies. Document the SLAs.

2. Outbound prioritization High-fit accounts with low engagement are your outbound list. This is where the refreshed model directly feeds your Outbound System Engineering — your reps should be sequencing the high-fit, low-engagement segment with personalized cadences, not blasting the entire database.

3. Forecasting and attribution The new score becomes a leading indicator. If MQL volume is flat but average fit score is rising, your pipeline is healthier than the raw count suggests. This kind of signal is what makes Revenue Intelligence useful instead of decorative.

The Cadence: How Often to Refresh

Scoring models aren't "set it and forget it" — but they're also not "tune it every week."

The cadence we recommend:

• Weekly: Monitor lead-to-opportunity conversion by score band. Flag drift early.
• Monthly: Spot-check 10 high-scoring leads and 10 low-scoring leads with sales. Are reps seeing what the model sees?
• Quarterly: Recalibrate weights based on the last 90 days of closed-won and closed-lost data. Adjust intent decay windows.
• Annually: Full ICP review and model rebuild. Re-derive fit attributes from scratch using the most recent 12–18 months of revenue.

Teams that don't have RevOps headcount to maintain this cadence are exactly who should be running it through a GTM Operations Retainer — because a model that isn't maintained will decay back into noise within a year.

The Bottom Line

ICP scoring decay isn't a tech problem. It's a discipline problem. Your business changes — your ICP shifts, your buyers' behavior evolves, new intent signals emerge, old ones get gamed — and your scoring model has to change with it.

The teams that win in 2026 aren't the ones with the most sophisticated scoring math. They're the ones who treat their scoring model as a living system: built from closed-won data, weighted toward fit, decayed aggressively, and recalibrated on a real cadence.

If your reps have stopped trusting the score, the score has already failed. Rebuild it before you generate another quarter of leads it can't qualify.

---

If your lead scoring model is producing numbers your sales team ignores — or if MQL-to-SQL conversion has quietly fallen off a cliff — we should talk. Book a strategy call with Revstek and we'll walk through where your model is decaying and what it takes to rebuild it.