Account scoring models that actually work share three traits: they're built on fresh, validated data rather than stale CRM fields; they combine fit (does this account match our ICP) with intent and timing signals (is something happening now); and they're calibrated against real won/lost outcomes instead of arbitrary point weights. A model that scores accounts on data that's 22% decayed per year will confidently rank dead companies highly — which is worse than no model at all.
Almost every revenue team eventually builds an account scoring model. Most of them quietly stop trusting it within two quarters. The model isn't wrong because the math is bad — it's wrong because of what it's built on and what it leaves out. Here's what separates a scoring model that predicts revenue from one that just looks rigorous in a deck.
Why Most Account Scoring Models Fail
The typical failure isn't the formula. It's three quieter problems:
- —Stale inputs: the model scores fields that decay roughly 22% a year, so it ranks acquired, shut-down, and moved companies as if nothing changed
- —Arbitrary weights: points get assigned by committee intuition ('employee count is worth 10, industry is worth 5') with no link to actual outcomes
- —Fit without timing: the model captures whether an account looks like a good customer but not whether now is the moment to work it
A model with all three problems produces scores that feel objective and are actually just confident guesses dressed up as data. That's more dangerous than no model, because reps trust the number.
The Three Layers of a Model That Works
A scoring model that holds up over time is built in three layers, in this order.
Layer 1: Validity
Before you score an account on anything, confirm it's a real, current, in-ICP company. Is the domain live? Was it acquired or shut down? Did it relocate or downsize out of your market? This layer is binary and it comes first, because a brilliant score on a dead company is just noise with a decimal point. Most teams skip this entirely, which is why their models drift.
Layer 2: Fit
For accounts that pass validity, score how well they match your ideal customer profile: industry, size, business model, tech stack, geography. Crucially, derive these weights from your actual closed-won data, not intuition. If your best customers cluster in a certain segment, the model should reflect that empirically — not because someone felt it was important.
Layer 3: Timing
Fit tells you whether an account should buy eventually. Timing tells you whether to work it now. Layer in recent signals: funding rounds, leadership changes, hiring patterns, expansion, or competitive displacement. An account with perfect fit and no timing signal is a nurture; an average-fit account with a fresh trigger might be your best call today.
Fit and Timing Are Not the Same Thing
The single most common scoring mistake is collapsing fit and timing into one number. They answer different questions and they decay at different speeds. Fit is relatively stable — a company's industry and size don't change weekly. Timing is volatile — a funding round can flip a cold account to hot overnight, and a leadership exit can do the reverse.
Keep them as separate axes. A simple grid — high/low fit against high/low timing — is more actionable than a single blended score, because it tells the rep what action to take, not just how excited to be.
Calibrate Against Reality
A scoring model is a hypothesis about what predicts revenue. Treat it like one. Every quarter, compare scores against outcomes: did high-scoring accounts actually close at higher rates? If your top tier converts no better than your middle tier, your weights are wrong, and no amount of additional fields will fix a model that was never validated against won/lost data.
This is the discipline that separates a real model from a vanity metric. The number means nothing until it's been checked against what actually happened.
The Foundation Most Models Are Missing
If you take one thing from this: the most sophisticated scoring logic in the world can't survive bad inputs. Validity has to come first. When your model starts from accounts you've confirmed are real, current, and in-fit — and only then layers on fit and timing — the scores finally start predicting something.
That's the difference between account tiering built on data you can trust and a spreadsheet that ranks ghosts. Get the foundation right, and the model does what it was supposed to do all along: point your team at the accounts most likely to turn into revenue.