Signal-based lead scoring: a model you can actually copy
The short answer
A model that scores accounts by weighted behavioural signals with recency decay applied, rather than by cumulative points that never expire. It scores at the account level and treats firmographic fit as a gate rather than as points, so activity and suitability stay separate.
Most lead scoring models have the same defect. Points only ever go up.
A prospect downloads a whitepaper in March and is still carrying those points in November. Two years of accumulated activity produces a high score for an account that stopped caring eighteen months ago, while a company that visited your pricing page three times last week sits below it.
The model is not measuring intent. It is measuring how long someone has been in your database.
Here is a model that does not have that problem, with actual numbers, so you can copy it and adjust rather than starting from a blank page.
Score accounts, not people
Before any weighting: the unit is the account.
Person-level scoring fragments a buying committee into individuals who each look lukewarm. Three people at one company each scoring 20 looks like three mediocre leads. As an account it is a 60 and a process forming.
Roll every person's activity up to the company. Keep the individual detail for context, but score, route and report at the account level. This is the same unit discipline that makes ABM metrics work.
Step 1: weight by signal strength
Not every action means the same thing. Base points:
| Signal | Points | Why |
|---|---|---|
| Pricing page visit | 10 | Closest to purchase intent of any page |
| Comparison or alternatives page | 10 | Actively evaluating, likely against a competitor |
| Specific implementation question by email | 12 | Highest-precision signal available. Nobody asks about SSO casually. |
| Demo request | 15 | Explicit |
| Docs or integration page | 6 | Serious, but often a developer on an unfunded project |
| Job posting matching your problem | 6 | Public evidence their current approach is failing |
| Technology removed from stack | 6 | A gap exists right now |
| Repeat visit within 7 days | 5 | Return behaviour is stronger than first visit |
| Executive change in the relevant function | 4 | Real but slow. Re-evaluation usually takes two quarters. |
| Webinar attendance | 3 | Attended, not evaluated |
| Content download | 2 | Traded an email for a PDF |
| Third-party surge score | 2 | Anonymous reader, coarse topics, lagging |
| Funding round | 1 | Money, not need. A multiplier, not a signal. |
| Social engagement | 1 | Reachability, not intent |
The controversial rows are the low ones. Content downloads and surge scores sit near the bottom on purpose, and they are usually the two largest contributors in a conventional model. The reasoning is in B2B buying signals.
Step 2: apply recency decay
The step that fixes the whole model.
| Age | Multiplier |
|---|---|
| 0 to 14 days | 1.0 |
| 15 to 30 days | 0.5 |
| 31 to 60 days | 0.25 |
| 60+ days | 0 |
Signals expire. A pricing page visit from four months ago is not a weaker version of a fresh one. It usually means they looked, evaluated, and chose someone else. Carrying it forward as partial credit actively misleads you.
Two exceptions worth encoding: an executive change decays over two quarters rather than two months, because the re-evaluation window is genuinely longer. A technology removal likewise stays live for about a quarter.
Step 3: multiply by committee spread
| Distinct people active | Multiplier |
|---|---|
| 1 | 1.0 |
| 2 | 1.5 |
| 3 | 2.2 |
| 4+ | 3.0 |
Deliberately superlinear. Four people looking is not four times one person looking, it is a different event. One person is curiosity. Four is an internal evaluation that someone has convened.
Add a further 1.5x when the active people sit in two or more distinct functions. Spread across functions predicts far better than depth within one, because it is the shape a real buying committee makes.
Step 4: apply fit as a gate, not points
The most common structural error is adding firmographic fit into the same score as behaviour.
Doing that lets a perfect-fit company with zero activity outrank a slightly-off-ICP company that is actively evaluating. Fit is not intent. It never was.
Use fit as a gate:
- In ICP: score normally
- Adjacent to ICP: score, flag for human review before any automated action
- Outside ICP: log it, never route it
Fit determines *whether* to pursue. Signals determine *when*. Keep them in separate columns. Building the fit model itself is covered in ideal customer profile for B2B SaaS.
Step 5: the override
Some combinations should trigger regardless of total score.
Two or more strong signals within 30 days from an in-ICP account routes immediately. No score threshold, no queue.
This matters because arithmetic can bury an urgent account. A company with exactly two signals, a pricing page visit and a specific implementation question, might total below a busier account that accumulated many weak ones. The first company is in-market. The second is on your mailing list.
Build the override before you build the threshold. It catches the cases the model gets wrong, and every model gets some wrong.
A worked example
Account A, in-ICP:
- Pricing page visit, 5 days ago: 10 × 1.0 = 10
- Repeat visit, 3 days ago: 5 × 1.0 = 5
- Comparison page, 4 days ago: 10 × 1.0 = 10
- Subtotal: 25
- Three distinct people, two functions: × 2.2 × 1.5
- Total: 82. Plus the override fires.
Account B, in-ICP:
- Content download, 45 days ago: 2 × 0.25 = 0.5
- Webinar, 90 days ago: 3 × 0 = 0
- Surge score, 20 days ago: 2 × 0.5 = 1
- One person: × 1.0
- Total: 1.5
Under a conventional cumulative model these two accounts can land close together. Under this one they are correctly two orders of magnitude apart.
Calibrate it against reality
The model above is a starting point, not a result. Tune it:
Backtest before deploying. Run it against last year's closed-won and closed-lost. Would it have surfaced the winners earlier? If not, the weights are wrong for your market.
Track precision, not volume. The question is not how many accounts crossed the threshold. It is what share of accounts that crossed it entered a real conversation within 60 days. Below roughly 20%, tighten.
Re-weight quarterly. Signal value shifts. A channel that predicted well last year may now be saturated.
Watch for the threshold trap. If your team gets more alerts than it can work, they will stop trusting all of them within about a fortnight. It is better to route ten accounts a week that are worth calling than fifty that mostly are not.
Frequently asked questions
What is signal-based lead scoring?
A model that scores accounts by weighted behavioural signals with recency decay applied, rather than by cumulative points that never expire. It scores at the account level and treats firmographic fit as a gate rather than as points, so activity and suitability stay separate.
Why do traditional lead scoring models fail?
Because points only accumulate and never expire. An account that engaged two years ago can outrank one actively evaluating this week. The model ends up measuring database tenure rather than purchase intent.
Should you score leads or accounts?
Accounts. Person-level scoring splits a buying committee into individuals who each look lukewarm, when together they represent a live evaluation. Roll all individual activity up to the company for scoring, routing and reporting.
How long should a buying signal stay in a score?
Full weight for 14 days, half to 30, quarter to 60, and zero beyond that. Executive changes and technology removals are reasonable exceptions and can decay over roughly two quarters, since their re-evaluation windows are genuinely longer.
Should firmographic fit be part of a lead score?
No. Combining them lets a perfect-fit account with no activity outrank an active one that is slightly outside ICP. Use fit as a gate that decides whether to pursue, and signals to decide when.
---
*NomiOS is RZLT's GTM and ABM engine. Point it at a target and get back finished, branded work built on a real read of that company.*
[See how NomiOS works →](https://nomios.rzlt.io)
Questions
Frequently asked
- What is signal-based lead scoring?
- A model that scores accounts by weighted behavioural signals with recency decay applied, rather than by cumulative points that never expire. It scores at the account level and treats firmographic fit as a gate rather than as points, so activity and suitability stay separate.
- Why do traditional lead scoring models fail?
- Because points only accumulate and never expire. An account that engaged two years ago can outrank one actively evaluating this week. The model ends up measuring database tenure rather than purchase intent.
- Should you score leads or accounts?
- Accounts. Person-level scoring splits a buying committee into individuals who each look lukewarm, when together they represent a live evaluation. Roll all individual activity up to the company for scoring, routing and reporting.
- How long should a buying signal stay in a score?
- Full weight for 14 days, half to 30, quarter to 60, and zero beyond that. Executive changes and technology removals are reasonable exceptions and can decay over roughly two quarters, since their re-evaluation windows are genuinely longer.
- Should firmographic fit be part of a lead score?
- No. Combining them lets a perfect-fit account with no activity outrank an active one that is slightly outside ICP. Use fit as a gate that decides whether to pursue, and signals to decide when.
