I've seen more broken lead scoring models than working ones. Not because the teams who built them didn't understand scoring — but because they treated it like a project with a finish line. Scoring isn't a project. It's a living system. Build it like one or don't build it at all.
Here's the honest six-step approach.
Step 1: Validate what actually predicts close rate#
Before you score anything, pull 12 to 24 months of closed-won and closed-lost deals. Look for patterns — firmographic, behavioral, demographic. What shows up consistently in closed-won that doesn't in closed-lost? You can't build a scoring model on intuition. You build it on historical signal. If you don't have that data clean and accessible, fix that first.
Step 2: Build it with sales, not for sales#
Every scoring model built without sales input gets ignored by sales. Get your top two or three reps in a room. Ask them one question: what signals tell you a lead is genuinely hot before you've spoken to them? Their answers will be more predictive than any consultant's framework. Bake those signals in.
Step 3: Assign point buckets that make sense#
Keep it simple to start. Firmographic fit gets 0–30 points: right industry, right company size, right geography. Behavioral signals get 0–50 points: pricing page views, demo requests, multi-page sessions in a single visit. Negative scoring matters just as much: personal email domain, competitor employee, unsubscribe behaviour. Without negatives, your model will eventually recommend a competitor's sales rep who keeps browsing your site.
Step 4: Set MQL and SQL thresholds against capacity#
Common thresholds: nurture at 1–49, MQL at 50–74, SQL at 75 and above. Don't pick these from an article. Pick numbers that align with actual capacity. Sales should never receive more SQLs than they can act on within 24 hours. If they can only handle 30 qualified leads per week, calibrate accordingly.
Step 5: Build in decay#
A pricing-page visit from eight months ago should not carry the same weight as one from yesterday. If your platform can't auto-decay, run score recalculations on a 90-day rolling window. Stale intent is noise. Fresh intent is signal.
Step 6: Backtest before you ship#
Take your new model and re-score the last 90 days of leads. Does the rank order reflect actual close rate? Does the top 20% of scores actually convert at a meaningfully higher rate than the middle? If not, the model isn't ready. Fix it before launching. Then review monthly for the first quarter and quarterly after that.
What breaks most scoring models#
Static models in a moving market. If you haven't recalibrated in six months, your model is scoring against a buyer profile that may no longer exist.
No negative scoring. I've said it once, but it's worth repeating: without negatives, scoring will degrade into noise over time.
Measuring MQL volume instead of MQL-to-SQL conversion. High volume with poor conversion means scoring is lying to you about what's qualified. The only metric that tells you whether scoring is working is what percentage of MQLs actually become sales opportunities.