You've got lead scoring turned on in HubSpot, the reps sort by it every morning, and you still can't tell whether it's pointing them at the right accounts or quietly steering them away from the ones that would've closed. Somebody set up the rules a couple of years ago; nobody remembers why a pricing page visit is worth 20 points, and the number sitting on each contact record is treated like a verdict instead of a hint.
Sarah and I have built lead-scoring models dozens of times over the past decade, and one of those builds nearly cost a client a multi-million-dollar opportunity. In this post, you'll get the three jobs a score is supposed to do (prioritization is only one), the story of the negative-scored lead that almost got ignored and what it taught us about privacy gaps, and a phased way to build a model you can trust, including a review cadence Sarah borrowed from how people build AI agents.
This post is based on Episode 77 of Revenue Rewired | Ep 77: Is Your Lead Scoring System Sabotaging Good Deals?
If you'd rather listen than read, find the full episode on Apple Podcasts, YouTube, Spotify, or Amazon. It's worth your time.
Sarah opened the episode with the simplest version of it: a number, somewhere between negative 100 and positive 100 in HubSpot's case, that tells a rep whether an inbound lead looks like a fit. That's a fine starting point, but I'd sharpen it. Lead scoring doesn't tell you who's ready to buy. It tells your sales team who deserves a closer look, and that distinction matters more the more leads you're getting.
Prioritization is the job everyone knows about. A rep has two leads in front of them, one scores higher, they call that one first. Two other jobs get far less attention. The first is pattern detection: a well-built model surfaces behavior like repeat visits to a demo or pricing page, or three different contacts from the same company showing up in the same week. The second is process discipline. To score anything, marketing and sales have to agree on what "qualified" means, and I've walked into plenty of companies where marketing has one scoring model, sales has a completely different one, and the gap between them traces straight back to two departments defining that one word differently.
Sarah's framing of the whole thing is the one I keep coming back to. Done collaboratively, the score is a bridge between the two teams. If a rep gets a lead scored from marketing activity and disagrees with the score, that disagreement should flow back to marketing as feedback, and the model improves. It shouldn't live in a spreadsheet, either, and yes, we've onboarded a $25 to $30 million company that was running its scoring that way before we met them.
In one of the scenarios I described on the episode, I was the problem. When you try to replicate human judgment with an automated system, and you keep stacking on rules, the score stops meaning what you think it means. Fit and intent are two separate things, and the trap is treating a high number as proof of both at once. Marketing sees a 90, counts the MQL, moves it to SQL, and sales makes it the top priority of the day.
Then the rep gets on the call and finds a poor-fit contact. When we dig in, the reason is often something like this: the model gives 10 points for visiting two or more pages and 20 for four or more, and the person wasn't researching; they were lost. They couldn't find what they needed, clicked around the site, and landed on the contact page because they ran out of options. Sarah pointed out that this is the same lesson the analytics world already learned when GA4 moved away from time on site toward engagement rate. Long sessions can mean interest, or they can mean confusion.
So my advice for a score is to use it for intelligence and trending, not to fully dictate how you engage an opportunity. When a lead scores a 40 and "a 40 usually means great for us," the rep still needs to open the record and see what the 40 is made of. With five to fifteen demographic and behavioral factors feeding a single number, that rarely happens on its own.
About eight to ten years ago, back when lead scoring first got popular, we built a model for a client by analyzing their ten best opportunities and their ten worst. The data was clear: prospects who showed up to a sales call having done zero research, no site visits, no pricing page, no videos, turned out to be poor leads 98 percent of the time. That was validated over years of history, so we made it a negative score. Come in with none of the baseline activity everyone else showed, and you started at something like negative 25 or negative 30.
The client had just come off a conference with more leads than they could handle, so that week they leaned on the scores to decide where to spend their time. A negative-scored lead got ignored. The only reason anyone circled back is that the person filled out the form again a week later, essentially asking whether the first one had gone through, and that lead went on to become a multi-million dollar client for them.
What we eventually figured out is that this buyer was one of the early privacy-conscious people. Everything turned off, incognito mode, opted out of every cookie consent they saw. The model read that as no interest when it was really no visibility. A low score doesn't always mean low interest, and after that project I stopped letting a score be the reason a lead gets skipped.
The commercial remodeling client Sarah and I discussed ran into a cousin of the same problem. Prospects would come in thinking a kitchen refresh in an office might run $15,000 when the real number was closer to $120,000, and the client wanted to negative-score anyone with a budget that low. We talked them out of it, because the thing that set them apart from every competitor was an advisory, educational approach to explaining why commercial work costs what it does. Auto-penalizing the very prospects that approach was built to win would've quietly cut off their best differentiator.
Sarah's answer came from how people are building AI agents right now: one action at a time, because every imperfect step compounds the errors in everything downstream. Her phase one for a HubSpot user who's never known where to start looks like this. Look back over double or triple your normal sales cycle, pull your top 10 to 20 deals and your worst, and write a rubric of five to seven scoring factors. Put it in the system, have sales act on it, and get their feedback every week for four to eight weeks. Once the sales acceptance rate on scored leads hits 70 percent or better, phase two brings in customer experience data, so you're scoring for clients who stay and grow, not just prospects who convert.
I'd add that the phased approach does something beyond accuracy. The conversations it forces between marketing and sales are worth more than the model itself, and they tend to sprawl in useful directions: we need new resources on the site, our ICP needs tightening, we don't have a signal for the thing reps care about most. Sarah's point about ambiguity applies here. If you ask everyone in the room what "strategy" means, you'll get slightly different answers, and if you let that kind of ambiguity into your definition of a great-fit client, you lose the thread that connected you to your ideal customer in the first place.
Two more things I'd build in from the start. First, treat the first six months of a new model as a learning period, meaning don't ignore the score but don't fully trust it yet, because there are always kinks to work out. Second, put an expiration date on it. Sarah noted that companies running serious AI agent pilots don't let an agent run past 30 days without a full review, and a scoring model deserves the same treatment on maybe a six-month cycle. It turns "when did we last look at this?" into a scheduled conversation, and it keeps a set-it-and-forget-it score from becoming the thing your whole team prioritizes around without anyone remembering why.
Decay is the last piece, and the one I think is under-used. We've watched leads show up with an 80 out of 100, the highest score the company has ever seen, get a rep follow-up within the hour, and then go completely silent. Maybe they signed with someone else, maybe they only wanted a quote to squeeze their incumbent. Without decay, that 80 sits there, and an executive who pops into HubSpot sees a multi-million dollar lead nobody seems to be working. With decay set up properly, the score comes down on its own as the person ghosts you, and the sales team has some protection against a number that no longer reflects reality.
Q: Is lead scoring a marketing tool or a sales tool?
A: Both, and it only really works when both teams treat it that way. Marketing usually builds the criteria, but sales is the group acting on the number every day, so their feedback on whether the score matched reality is what makes the model improve over time.
Q: How many factors should a lead scoring model have?
A: Start with five to seven. Models with fifteen or more demographic and behavioral inputs produce a number nobody can interpret at a glance, and that's when reps start trusting the score instead of reading the record.
Q: Should we use negative scoring at all?
A: Yes, carefully. Negative scores are useful for patterns your history clearly supports, but be cautious with anything that could reflect a tracking gap rather than real disinterest, like a lack of site activity from someone who blocks cookies, or a budget signal that your sales process is designed to reframe.
Q: How often should we review our lead scoring rules?
A: Weekly feedback from sales during the first four to eight weeks, then a full review on a fixed cadence, six months at the outside. Treat the first six months of any new model as a trial period where you watch the score without letting it fully drive your outreach.
Q: What does HubSpot's Breeze AI change about lead scoring?
A: Breeze enriches contact and company records with outside data, like a new funding round or an open job posting, and that enrichment can shift scores up or down. It's a useful signal layer, but it's one more input your team needs to understand rather than take at face value.
StringCan Interactive works with mid-market B2B companies to fix the places where revenue leaks between marketing and sales, and lead scoring in HubSpot is a spot we're in nearly every week. We help teams define what qualified means in a way both departments will sign off on, build the phased model, and set up the feedback loop so the score gets smarter instead of stale.
If your scoring model was built a while ago and nobody's sure it's still telling the truth, that's the kind of conversation we like having. Listen to the full episode here, and if you want to look at your own model together, reach out to the team at StringCan Interactive and we'll walk through it with you.