Build an Account Scoring Engine
Outcome: a two-part score over your whole account universe, weights validated against your own closed-won data, and a threshold set from team capacity.
- Surface
- App
- Level
- Intermediate
- Uses
- Formula columns · signal enrichments
- Credits
- 0 to score; signal columns cost separately
- Prerequisite
- Lessons 04–06
Two scores, not one
Fit — can we sell to them?
Size, geography, stack, business model. Slow-moving, refreshed monthly or quarterly.
Timing — should we call them now?
Funding, hiring, headcount growth, exec changes, stack changes, traffic trend. Recomputed every run and decayed as each signal ages.
Combine with a floor
fit >= threshold AND timing > 0, ranked by timing. Never average — averaging lets a hot signal at an unsellable account outrank a strong-fit account, which defeats the point of scoring at all.
A model to start from
Fit — max 15
| Criterion | Points |
|---|---|
| Employee band in target range | 4 |
| Country in target list | 3 |
| Competitor tool detected | 4 |
| Complement tool detected | 2 |
| Business model verdict matches ICP | 2 |
| Business model verdict Unknown | −1 |
Timing — max 15, decayed
| Criterion | Points |
|---|---|
| Funding round in last 90 days | 5 |
| 2+ relevant roles posted in last 30 days | 5 |
| 1 relevant role posted in last 30 days | 3 |
| New exec in the relevant function, under 90 days | 4 |
| Headcount growth above 20% over 6 months | 3 |
| Traffic growth above 30% over 6 months | 2 |
| Any signal aged 60–90 days rather than under 30 | −2 |
Whole numbers, six or seven criteria each. A score with twenty criteria is not more accurate — it is just harder to explain, and an unexplainable score gets ignored.
Choose weights from evidence
Do not guess. Derive them.
Score your closed-won accounts
As they were at the time you first engaged, as closely as you can reconstruct.
Score a matched set of no-opportunity accounts
Same size, same market, never went anywhere.
Compare per criterion
For each criterion, what share of each group had it?
Weight by the gap
A criterion present in 70% of won accounts and 20% of the rest deserves a high weight. One present in 60% of both deserves zero — it describes your market, not your buyer.
This is the same discriminating-attribute analysis as AI-Powered GTM lesson 07. Running it once gives you both a lookalike definition and a scoring model, which is why it is worth the half hour.
Set the threshold from capacity
- Compute how many accounts your team can genuinely work per week.
- Score the whole universe — free.
- Sort by timing within a fit floor.
- The score at your capacity row is the threshold.
- Read ten accounts just above and ten just below. If you cannot tell them apart, add a discriminating criterion.
The threshold moves with capacity. The model does not.
Validate it
Two checks before anyone works the output:
- Backtest. Score accounts as of six months ago. Did the accounts that became opportunities score highly? If your top decile contains no closed-won accounts, the model is not predictive.
- Face validity. Show the top 20 to a rep who knows the market. If they disagree with most of it, either the model is wrong or the rep’s intuition encodes a criterion you have not captured. Both are worth finding out.
Do this now
List your criteria
Split into fit and timing.
Run the won-versus-nothing comparison
Per criterion, both groups.
Assign weights by the gap
Zero out anything that does not discriminate.
Build both scores as formulas
Free, so build before anything else.
Compute capacity and read the threshold
At your capacity row.
Backtest
Six months back, against known outcomes.
Show the top 20 to a rep
And listen to the disagreements.
Check your work
- Fit and timing are separate, combined with a floor
- Every weight traces to a measured gap between won and not-won
- The threshold came from a capacity number
- The model backtests against real outcomes
- A rep has reviewed the top 20
Where this breaks
A score built from criteria that do not discriminate produces a confident ranking of noise. It will look like a working prioritization — accounts sorted, thresholds applied, queues populated — while the ordering carries no information about who will buy. The won-versus-not-won comparison is the only thing that distinguishes a real model from a plausible one, and skipping it is skipping the lesson.
Further automation
Re-score on every refresh so accounts cross the threshold on their own as signals fire. Route newly-crossed accounts to a queue or to Slack, and the market map becomes a live pipeline rather than a document. Lesson 08 builds the refresh.
Next lesson
08 — Keep your TAM fresh, the scheduled refresh that adds, updates and retires accounts.
Reference for this lesson: Signals, Raised funding recently, Headcount growth, GTM Engineering — scoring and gating.