Latest / Marketing Automation with Fexingo: HubSpot, Marketo, and Automated Customer Journeys / Why Your Lead Scoring Model Is Broken and How to Fix It
Transcript
- Lucas: If these marketing conversations have sparked something you've actually used at work, there's a good chance you've run into lead scoring. The idea is simple: assign points to leads based on actions and fit, then pass the highest-scoring ones to sales. It sounds logical. But a lot of companies are running scoring models that are quietly undermining their pipeline. Luna: Quietly undermining? I've set up a few scoring models myself. They seem to work fine on the surface. Lucas: They do, until you dig into the data. I was talking with a B2B SaaS client last quarter — they sell a project management tool for mid-market teams. Their scoring model had been running for about eighteen months. Points for email opens, webinar attendance, demo requests, job title, company size. The usual mix. Luna: Right, so activity plus firmographic. That's the textbook approach. Lucas: Exactly. But when they finally measured the correlation between score and actual closed-won deals, it was almost flat. Leads with 400 points converted at about the same rate as leads with 100 points. Something was adding noise, not signal. Luna: So what was the noise? Was it just too many points for trivial actions? Lucas: Partially. But deeper analysis revealed three specific biases baked into their model. The first is recency bias — they were giving a lot of points for recent activity, like an email open in the last 24 hours. But that activity doesn't correlate with purchase intent. Someone could open three emails in one day because they're cleaning their inbox, not because they're ready to buy. Luna: Huh. So the model rewards frantic but low-intent behavior. Lucas: Exactly. The second issue was data silo bias. Their marketing automation platform — HubSpot in this case — tracked email clicks and form fills. But their CRM had sales call logs and meeting notes. The scoring model only looked at the HubSpot data. So a lead who had three conversations with a sales rep but never opened another email would actually lose points over time. Luna: So the scoring model punished engaged leads who moved to a different channel. That's brutal. Lucas: It is. And the third was the most subtle: activity over intent. They scored for actions like downloading a whitepaper or attending a webinar. But those actions don't tell you if the person has budget or authority. A junior analyst downloading a whitepaper gets the same points as a VP of Engineering requesting a demo. The model couldn't distinguish between 'just curious' and 'actively buying'. Luna: So what did they do? Did they just tweak the point values? Lucas: They tried that first. They lowered points for email opens and raised them for demo requests. But the improvement was marginal. The real fix came when they introduced behavioral scoring with negative signals. Luna: Negative signals — like what, subtracting points? Lucas: Yes. They started deducting points for things like multiple bounced emails, unsubscribes, or visiting the careers page — which often signals a job seeker, not a buyer. They also added a decay function: if a lead went 90 days without any meaningful interaction, their score would gradually drop. That alone cleaned up a lot of stale leads clogging the pipeline. Luna: So the model started punishing inactivity and low-intent behavior. That makes sense. Lucas: But the biggest leap came when they added a machine learning layer on top of the rule-based scoring. They trained a simple logistic regression model on historical closed-won data, using about 200 features from both marketing automation and CRM. The model output a probability score between 0 and 1. Then they blended that with the traditional point score — weighting the ML prediction at 60 percent and the rule-based score at 40 percent. Luna: And how did that change conversion rates? Lucas: Within three months, their sql to close rate went from 22 percent to 27 percent. That's roughly a 23 percent improvement. And the average deal size stayed the same, so it wasn't just pushing smaller deals through. The sales team also reported higher satisfaction because leads they called were actually ready to talk. Luna: Twenty-three percent lift just from scoring changes. That's significant. Lucas: It is. And it didn't require a huge data science team. They used a third-party tool that integrated with HubSpot and Salesforce — there are several now that offer out-of-the-box predictive scoring. The key was that the model was trained on their own data, not generic benchmarks. Luna: So the lesson is: your scoring model is only as good as the data you feed it, and you need to look beyond marketing data. Lucas: Exactly. And you need to constantly validate. They now run a quarterly audit where they compare predicted scores against actual outcomes and recalibrate. That's become a standard practice for them. Luna: I imagine a lot of teams skip that step. They set it and forget it. Lucas: Most do. And that's why so many scoring models degrade over time. Buyer behavior changes, market conditions shift, new competitors emerge. A model that worked in 2024 might be useless by mid-2026. Which brings me to another point: the scoring model should be dynamic, not static. Luna: Dynamic how? Like real-time adjustments? Lucas: In a sense. One approach is to use time-weighted scoring where recent interactions count more than older ones, but with a decay curve that's calibrated to your sales cycle. For a B2B SaaS product with a typical 60- to 90-day cycle, you might give full weight to actions in the last 30 days, half weight to actions 30 to 60 days ago, and zero weight beyond 90 days. That prevents ancient interactions from artificially inflating a score. Luna: That sounds like a straightforward fix that most platforms can handle natively. Lucas: They can. HubSpot and Marketo both allow time-based scoring. But very few teams enable it because the default is usually to accumulate points forever. They'd rather have a high score than an accurate one. Luna: So it's a design choice that favors quantity over quality. Lucas: Exactly. And that leads to sales teams ignoring the scores entirely, because they learn that a high score doesn't mean a hot lead. They start relying on their own intuition, which defeats the purpose. Luna: You mentioned negative signals earlier. Can you give another example beyond the careers page? Lucas: Sure. One common negative signal is a high number of support ticket submissions. If a lead is opening tickets about product issues, they're likely an existing customer, not a new prospect. Another is frequent visits to the pricing page without requesting a demo — that can indicate price sensitivity or comparison shopping, not purchase intent. Some models even subtract points for downloading competitor comparison guides. Luna: Interesting. So negative signals help filter out the noise that traditional scoring misses. Lucas: They do. But you have to be careful not to over-penalize. For example, visiting the careers page could also mean a lead is researching your company culture before making a large commitment. So you need to combine signals, not rely on any single one. Luna: Right. Context matters. Lucas: The company I mentioned initially had a lead who visited the careers page five times in a week. The model deducted 50 points. But that lead later turned out to be the VP of Engineering, who was checking out the team before a potential partnership. They almost lost a six-figure deal because of a blunt negative rule. Luna: Ouch. So there's a balance between filtering and over-correction. Lucas: Exactly. That's why the machine learning approach is more robust — it can weigh hundreds of signals simultaneously and learn which combinations are predictive, rather than applying simplistic rules. Luna: So for a smaller team without data science resources, what's the best first step? Lucas: I'd recommend starting with an audit of your current model. Export your last six months of leads, their scores, and their final disposition — won, lost, or still open. Then calculate the conversion rate by score decile. If the top decile doesn't convert at least twice as well as the bottom decile, your model is broken. Luna: That's a simple sanity check anyone can do in a spreadsheet. Lucas: Exactly. Then fix the most obvious biases: add a decay function, integrate CRM activity data, and introduce a few negative signals. You'll likely see improvement without any machine learning. The ML layer is the next step. Luna: And for those ready to try ML, are there affordable tools that plug into Marketo or HubSpot? Lucas: Yes. Tools like MadKudu, Lattice Engines, and Infer — though Infer was acquired by Dun & Bradstreet a while back. There's also CaliberMind and Leadspace. Most start around $1,000 to $2,000 per month for mid-market. For enterprise, you're looking at $5,000 plus. But the ROI can be significant if your current model is underperforming. Luna: That's a good benchmark. So the takeaway: don't assume your scoring model works just because it's been running for a while. Lucas: Right. Audit it, fix the biases, and consider layering in predictive scoring. It's one of the highest roi changes a marketing ops team can make. Luna: And it's something a listener can start on Monday with just a spreadsheet export. Lucas: Absolutely. And honestly, the fact that we can dig into episodes like this — week after week, without ads — is only possible because a small group of listeners chips in at buy me a coffee dot com slash fexingo. It's not a big ask, just a way to keep this ad-free and focused on the mechanics. Luna: Yeah, and we really appreciate that. It means we can spend time on cases like this scoring fix instead of chasing sponsors. Lucas: Exactly. So if the show has helped you sharpen your marketing ops, that's the place. But back to scoring: one final thought. The company I mentioned is now experimenting with intent data from third-party sources — like G2 reviews and LinkedIn engagement — to feed into their model. Luna: Oh, that's interesting. So they're pulling in signals from outside their own systems. Lucas: Yes. Early results show that leads who have recently researched competitor alternatives on G2 are 30 percent more likely to convert. They're adding that as a weighted feature. It's still early, but it points to where scoring is headed: more data sources, more signals, and more intelligent weighting. Luna: So the future of lead scoring is essentially a recommendation engine for prospects. Lucas: Exactly. And it's already here for teams willing to invest in the data infrastructure. The gap between the haves and have-nots is widening. Luna: Alright, something to think about. Thanks, Lucas. Lucas: Thanks, Luna. See you next time.