Two scores, not one
A single blended number hides the difference between 'perfect customer, not looking' and 'wrong company, very curious'. Those two leads need different actions, and a rep who gets three of the second kind in a row stops opening the queue.
Fit is slow-moving and mostly firmographic. Behaviour is fast-moving and decays. Keeping them separate lets you route on the combination: high fit plus high behaviour goes to a rep today; high fit plus low behaviour goes to nurture; low fit is not an MQL at any level of activity.
Derive weights from closed-won data
- 01Export the last 12–24 months of closed-won and closed-lost deals with their associated contacts.
- 02For each candidate attribute, calculate the conversion rate of contacts with that attribute versus the base rate.
- 03Keep attributes that move conversion by a meaningful margin. Discard the rest, however intuitive they feel.
- 04Set weights proportional to the lift, then round hard — precision beyond five-point increments is false confidence.
- 05Backtest: score last quarter's leads with the model and check whether the leads that converted actually scored highly.
Setting the threshold
The bands only mean something if the middle column is filled with your data. Copying someone else's thresholds is how models lose credibility in the first month.
| Score band | Historical conversion to opportunity | Action |
|---|---|---|
| 0–39 | Below base rate | Nurture, no rep touch |
| 40–69 | Around base rate | Marketing sequence, review monthly |
| 70–89 | 2–3× base rate | Rep task within 24 hours |
| 90+ | 4×+ base rate | Immediate routing and alert |
Decay and negative scoring
- Halve behavioural score after 30 days of inactivity, zero it after 90.
- Score down for competitor domains, student and personal email addresses, and job applicants.
- Score down for unsubscribes and hard bounces so dead records do not sit in the MQL queue.
- Never decay fit — a 500-person manufacturer is still a 500-person manufacturer next quarter.
Make the score explainable
Write the top contributing factors to a text property on the contact record. A rep seeing 'Score 84 — VP title (+20), 3 pricing page visits in 7 days (+35), demo request (+30)' will make the call. A rep seeing '84' will not.
This single step does more for adoption than any amount of model tuning.
Review it in public
For the first three months, review the model monthly with both sales and marketing in the room: how many MQLs, how many accepted, how many converted, and which rejections had a pattern. Change one thing at a time. After the third review, quarterly is enough.
Common mistakes
- Scoring on email opens. Open tracking has been unreliable since privacy proxies became default.
- One combined score that cannot distinguish fit from intent.
- Thresholds picked in a workshop instead of derived from conversion data.
- No decay, so a spike of interest from last spring keeps a record in the queue forever.
- Never telling sales when the model changes, so their mental model and the system's diverge.