A lead lands in a rep's queue with a score of 87. The rep looks at it, can't tell why it's 87, and goes back to the accounts they were already working.
Nobody complains about this, which is the problem. The score keeps being calculated, the MQL count keeps going into the monthly report, and sales quietly stopped reading it some time ago.
Most scoring models were built in a single workshop, usually around a whiteboard, usually two or three years back. Somebody gave a demo request 10 points and an email click 3, everyone nodded, and nobody has checked those numbers against a closed deal since.
Sales isn't being stubborn
The standard diagnosis is that sales hasn't bought in. Run an enablement session, walk them through the model, get alignment.
It rarely works, because the rep's skepticism is earned. They've had enough 87s that turned out to be a student downloading an ebook, a competitor reading the pricing page, or a customer who already has a contract. Every bad MQL teaches them the number is noise, and it takes a lot of good ones to unlearn that.
A score sales trusts is one that has been right often enough, and that can say why it thinks what it thinks. Most models fail both tests.
Where trust breaks
Fit and intent get added together. A perfect-fit account that hasn't done anything and a five-person company that has opened every email can land on the same number. Those are different leads that need different handling, and a single score hides which one you're looking at. The rep has to open the record to find out, at which point the score has saved them nothing.
Nothing decays. The webinar someone attended fourteen months ago is still worth 15 points. Scores only go up, so the database slowly fills with people parked just under the threshold on the strength of activity from a previous job. Eventually one newsletter click tips them over and they arrive in a rep's queue as "hot."
The activity data is dirtier than it looks. Opens have been unreliable since mail privacy features started pre-loading images; for many audiences a large share of recorded opens are a mail client, not a person. Corporate security scanners click every link in an email before the recipient sees it. A model that awards points for opens and clicks without filtering for either is partly scoring infrastructure.
It was never calibrated. The weights came from opinion. Someone believed a pricing page visit mattered more than a case study download, which sounds right, and nobody went back to check it against what actually closed. Sounding right and predicting are different properties.
Two numbers, not one
The biggest single improvement is splitting the score in two.
Fit answers "is this the kind of company and person we sell to?" It comes from firmographic and demographic data: industry, company size, role, seniority, region if it matters to you. It changes slowly, and it should come mostly from the CRM and enrichment, not from behaviour.
Intent answers "are they doing something right now?" It comes from activity: high-value pages, demo or pricing requests, product signups, repeat visits in a short window. It moves fast and it should decay.
Plot one against the other and you get a grid. High fit and high intent goes to sales now. High fit with low intent belongs in nurture or account-based work; these are accounts you want that haven't raised a hand yet. Low fit with high intent is the interesting corner, sometimes a self-serve or partner route, sometimes a sign that your fit definition is too narrow. Low fit and low intent stays in the database and out of everyone's queue.
Most MAPs support more than one score natively. If yours doesn't, two custom fields and some discipline will get you there.
The parts that keep a model honest
Negative scoring comes first. Titles like student, recruiter or consultant. Competitor domains. Existing customers, who belong in a different process entirely. Careers page visits, which mostly mean someone wants to work for you rather than buy from you. A model that can only add points drifts upward forever.
Then decay. Intent points should lose value over time, halving or resetting over a window that matches your sales cycle. A 30-day cycle and a nine-month cycle need very different rates. Fit generally shouldn't decay, though it should update when the underlying data changes.
Filter the bots. Drop opens from the intent score, or weight them close to zero. Ignore clicks that land within seconds of delivery, or that hit every link in an email at once. Those are scanner signatures.
And show the reasons. This is the one that earns trust fastest. Next to the score on the CRM record, display the last few things that moved it: pricing page three times this week, demo request, VP title, target industry. A text field the MAP updates is enough. The format matters less than the principle, which is that a rep should never have to take the number on faith.
Set the threshold from data
The MQL threshold is usually a round number someone picked because it was round. Here's how to replace it with one that means something.
Pull every lead or contact that became an opportunity in the last 12 months, with the score they had at the moment they were handed to sales. Then pull a comparable set that was handed over and went nowhere. If your system doesn't keep score history, fixing that is step zero, and until then you can approximate from activity dates.
Group both sets into score bands and look at conversion for each band: handoff to opportunity, and opportunity to closed-won. A working model shows a staircase, where higher bands convert better. A broken one shows a flat line, or a middle band that outperforms the top.
Put the threshold where conversion becomes worth a rep's time. That part is a business decision, not a statistical one. It depends on how many reps you have and how many leads each can actually work.
The same exercise shows which rules are pulling their weight. Look for activities that closed deals share and dead leads don't. Some of your 10-point actions will turn out to be noise, and something you gave 2 points will turn out to be the best predictor in the model. In our experience it's often a pattern, like repeat visits inside a week, rather than any single page.
A few hundred deals won't give you statistical certainty. They'll give you something far better than the whiteboard.
Keeping it trusted
A model that's right on launch day drifts. Products change, the site changes, campaigns pull in different audiences. Trust holds if the model visibly keeps up.
Agree a handoff SLA that runs both ways. Marketing commits to a definition; sales commits to working what meets it within a set window.
Require a reason when sales rejects a lead. A picklist with five or six real options, not free text: wrong fit, no real interest, existing customer, duplicate, bad timing. That field is the only structured feedback most teams will ever get, and it tells you which rule to fix next.
Recalibrate every quarter by rerunning the band analysis. With score history in place it's an afternoon.
Keep a changelog for the model itself: what changed, when, why, and what conversion looked like before and after. When someone asks why MQL volume dropped 30% in March, the answer should take a minute to find, not a meeting.
When not to score at all
If sales can personally look at every inbound lead each week, they probably should. Scoring is a prioritisation tool for when volume outruns attention. Below that point it mostly adds logic someone has to maintain, and a number that can be wrong in ways a person reading the record wouldn't be.
A plain routing rule, demo requests straight to sales and everything else to nurture, beats a sophisticated model nobody reads.
What scoring sits on
None of this works if the records underneath are split. Engagement spread across two duplicate records means neither crosses the threshold. A CRM that disagrees with the MAP about which leads exist means the model scores records sales will never see. We've covered both: why CRM and MAP numbers don't match, and why the merge is the risky part of duplicate cleanup.
Most of the time, when scoring "stops working," the scoring was fine. It was being fed the wrong records.
Designing a scoring model sales will use, and wiring the handoff, SLA and feedback loop around it, is marketing operations work. If you're not sure whether your current model predicts anything, a stack audit will tell you, and the findings are yours either way.