Two scores, never one
A single composite score hides the difference between 'perfect customer, no interest' and 'wrong company, reading everything'. Those need different plays. Split them and the routing rules write themselves.
| Fit | Intent | Play |
|---|---|---|
| High | High | Route to AE immediately, alert in Slack |
| High | Low | Outbound sequence, targeted content |
| Low | High | Nurture, self-serve, watch for a role change |
| Low | Low | Leave in the database, no rep time |
Choosing weights you can defend
Do not invent weights. Pull the last two quarters of closed-won deals, look at what those contacts had in common on both dimensions, and weight accordingly. Where you have no evidence, use a small weight and say so.
- Fit: company size band, industry, job role or seniority, region, technology fit.
- Intent: pricing page views, demo or contact form submission, repeat sessions within a week, email replies (not opens), meeting bookings.
- Negatives: personal email domain, student or competitor, unsubscribe, job title that never buys.
Decay is non-negotiable
Intent without decay produces a database of permanently hot leads from 2024. Expire intent points after 30–60 days depending on your sales cycle. Fit scores don't decay — they change when the firmographics change, which is what data enrichment is for.
Set the threshold with sales in the room
The sign-off matters more than the maths. A model sales agreed to gets worked; a better model imposed on them gets ignored.
- 01Score the last two quarters retroactively.
- 02Show sales the leads that would have been MQLs at three candidate thresholds.
- 03Ask which list they'd have worked. Take that threshold.
- 04Write the definition down and get explicit sign-off.
- 05Agree an SLA: response time for MQLs, and what happens to leads they reject.
Manual scoring vs AI scoring
HubSpot offers AI-assisted scoring alongside manual property-based scores. AI scoring is useful once you have enough closed-won volume for a pattern to exist, and misleading before that. Our rule: start manual and transparent, because sales trust in the model comes from being able to see why a lead scored 82. Layer AI scoring on later as a second signal, not a replacement — and keep the manual model running alongside it for a quarter so you can compare.
Review monthly for the first quarter
- MQL-to-SQL acceptance rate — under 60% means the threshold or the fit criteria are wrong.
- Rejected MQLs, read individually. Patterns show up fast.
- Closed-won leads that never became MQLs — the most valuable list you'll look at.
- Score distribution: if everything clusters at one value, your weights aren't discriminating.
Common mistakes
- Scoring email opens, which are unreliable signals under modern privacy protection.
- One composite score that hides the fit-versus-intent distinction.
- No decay, so the hot list never cools.
- Launching the model without sales sign-off.
- Never reviewing it again after the launch month.