AI lead scoring needs your own won and lost deals

A rule score adds points by hand. Open an email, get points. Download a guide, get more. H2O.ai's lead scoring page states the problem plainly: rule-based scores reward interest, not fit, because the rules only count clicks.

So a lead that downloads everything can rank above a buyer with budget and a real project. H2O.ai's fix is a model that learns from your own won and lost leads, so the score reflects the leads that actually buy. Demandbase's guide, written by Marc Perramond on 4 January 2025, defines it the same way. It describes machine learning that uses historical data and real-time behaviour to predict which leads are most likely to buy.

That's the real difference. A rule score is a guess someone wrote down. A predictive score is the model's estimate that this lead becomes an opportunity, based on the patterns in leads that already did or didn't.

HubSpot sits between the two. Its scoring lets you build scores on fit, engagement, or both. Its AI-assisted engagement scoring studies past interactions of leads that converted, then recommends a more precise score. You keep your judgement, and the model points out where it's off.

Don't confuse adoption with this. Salesforce's State of Marketing report, cited on its blog on 7 December 2023, found around 62% of marketers use AI, with predicting prospect behaviour among the top uses. Using AI isn't the same as running a model trained on your own outcomes. A vendor score built on someone else's data can still rank the wrong leads for you.

This split matters less if your product has one obvious buyer type and almost every lead that engages is a fit. Then a rule score on fit does most of the work. Here's your first decision. Write down what your current score adds points for, and mark each rule as fit or interest. If most of them are interest, you have H2O.ai's problem, whatever the tool calls itself.

The ranking guides list where scoring data comes from. Demandbase's Phase 1, data collection, names your sales systems and marketing automation tools. None of the guides here spends much time on a source you already own: your own website, read at company level. That means which companies came back, which pages they read, and in what order.

Order is the signal. Take two companies that both match your ideal customer profile. The first reads your docs, then a comparison page, then pricing, and later comes back to pricing. The second opens a few blog posts and leaves. The first is acting like a buyer checking fit. The second is curious, and curiosity is exactly what H2O.ai says rule scores over-reward.

A rule score struggles here. It might give the same points for a blog read and a pricing read. It might even rank the blog reader higher because it clicked more. Reading the path fixes that. When a company that fits your profile keeps coming back to pricing and comparison pages, you get fit and intent in one reading, and that's the gap H2O.ai describes.

It also feeds the model. A model can only learn from the inputs you give it, and H2O.ai builds its model from behaviour, firmographic and campaign data. If page order at company level is one of those behaviour inputs, the model can learn whether "docs, then comparison, then pricing" shows up more in your won deals than in your lost ones. If you never record it, no model can learn it.

Keep this at company level only. The data comes from your own site, not a bought list, so it shows interest in your product rather than your category. A third-party list can tell you a company is researching a topic. Your own pages tell you it's researching you.

A US B2B SaaS team and an Indian one read the signal the same way: company, pages, order. What changes is how many visiting companies the data can name, and which privacy rules each team works under. No source here gives coverage figures for either market. So check what your own tool can name before you trust a ranking built on it, and get your own legal review on consent.

None of this works on a site with little traffic. If only a handful of companies reach your pricing page in a month, the order tells you about those companies, not about a pattern. Here's the decision. List your pricing, comparison and docs pages, and treat a return visit to them from a company that fits as a priority signal.

H2O.ai attaches a reason code to every score. The rep can read why a lead ranks high, so the first call starts from what matters to that buyer. Without one, a priority score is just a number at the top of a list.

HubSpot handles it differently. It puts lead scores on each lead's record "for complete transparency." The score sits where reps already work, next to the history that produced it.

A bare number brings back an old argument. H2O.ai names marketing and sales arguing over what a qualified lead is as one cost of bad scoring, next to missed targets and reps who burn out on poor leads. HubSpot pitches scoring as a way to create better alignment between marketing and sales. Both are making the same point. If sales can't check the reasoning, sales stops trusting the score.

Where the score lives matters too. H2O.ai says its scores feed the systems you already run, and that it isn't a campaign tool. A score sitting in a separate dashboard nobody opens won't change which lead gets the first call.

Picture two leads at the top of a list. The first has a reason line: fits the target industry, read pricing twice, read the comparison page. The rep opens with the comparison instead of a generic intro. The second ranks high because it opened every email. Its reason line shows that's interest, not fit, so the rep can move it down.

This matters less for a founder who works every lead personally, because they already know why each lead is on the list. Reason codes start to pay off once more than one person is choosing which lead to call. Decide now: no score goes live unless it carries a plain reason a rep can read in one line.

This is the counter-case, and it's a fair one. A model learns from outcomes. H2O.ai's whole approach rests on learning from the leads that converted and the ones that didn't. Demandbase's definition relies on large datasets and historical data.

If you've closed few deals, the model has little to learn from. That's when AI lead scoring becomes a rule score with a new label. With only a handful of wins, there's no real pattern to find. The model fits the noise anyway, and hands you a number that looks precise.

There's a second trap. A model only learns from what you record. If your history is mostly form fills and email opens, the model learns to rank form fills and email opens. That's H2O.ai's interest-over-fit problem again, now hidden inside a model where nobody can read the rules.

The rule-score side also has a real point about control. HubSpot's builder page says that as your business grows, it's important your team stays in control of how leads are scored. You can explain a rule you wrote. You can change it on a Monday morning.

So here's the trade-off. A model can find patterns you'd never think to write down, but only with enough history and the right inputs. A rule score is weaker, but you can read it from day one.

This counter-case doesn't hold for a team with a long record of won and lost leads tracked against fit and behaviour. That model has real material to learn from. Base the decision on your own records, not a vendor's pitch. Pull your closed-won and closed-lost leads and ask whether there's enough there for a pattern before you pay for a model.

You don't need a vendor to find out whether a model would help. You need your own won and lost leads and your own site data, read at company level. Here's a manual pass you can run this week.

• Pull the companies behind last quarter's won and lost leads from your sales system.
• For each one, list which pricing, comparison and docs pages the company read before it became a lead, and in what order.
• Mark the paths that show up among wins and not among losses.
• Rank this month's visiting companies against those paths, with a reason line on each.

Then compare. Check whether the companies your current rule score ranks highest are the ones following winning paths. If they aren't, your score rewards interest, and that's the gap a model or a better rule has to fix. If they are, your rules are already doing the job, and a model may add little.

This works the same for a US team and an Indian one. It runs on your own deals and your own pages, not a market average, so it doesn't depend on where the buyers are.

The cost is time. This is a manual pass, and it only tells you about paths that have already happened. It's still cheaper than training a model on history that turns out to be too thin.

It doesn't work with little traffic. If only a few companies reach your pricing or comparison pages, the paths won't separate wins from losses, and you'll end up reading meaning into chance. In that case, score on fit first, the way HubSpot's builder allows. Come back to page order once more companies reach those pages.

PageFox is the wrong choice for a site with little traffic: it tells you which companies are already visiting, it does not bring new visitors.

Decide whether your won and lost history is deep enough to train a model; if it isn't, keep a rule score built on fit and add company-level page order as a priority signal. This week, pull last quarter's won and lost leads and check which of your pages those companies read before they became leads.