See Payani AI in action. Watch the demo

Payani AI
The blog
Insights

What a Lead Score Should Actually Predict

PAYANI AINovember 4, 20269 min read

What a Lead Score Should Actually Predict
ShareLinkedInX

Your lead score is probably measuring how much people like your marketing. That is a different question from who is going to buy, and in long sales cycles the two correlate far more weakly than anyone wants to admit.

A score is only worth keeping if it predicts one thing: the probability that this person, at this company, right now, becomes a closed-won deal. Not opens. Not downloads. Not webinar attendance. If you cannot show win rate rising cleanly as the score band rises, you do not have a scoring model, you have a leaderboard for your content.

Here is how to build one that survives contact with a sales team.

Why does sales ignore our lead scores?

Because the scores do not match what reps see in pipeline, and reps learn that fast.

The pattern is well documented by people who do this work. Marketing builds the model in isolation, assigns point values on instinct, hands it over as a finished artefact, and sales quietly stops using it, because the scores do not reflect the deals they are actually working (RevBlack). The same diagnosis appears elsewhere: trust breaks when sales never helped define what a sales-ready lead is in the first place, so the threshold does not reflect what reps want to call (Markivis).

There is a blunter version. When marketing defines a lead as anyone who completed a form, reps discover that most of those contacts never convert. That is a structural problem, not an attitude problem, and the fix is to raise the qualification bar before leads reach sales (DataOps Group).

So the first move is not a better points table. It is agreeing, with the people who carry the number, what a sales-ready lead actually looks like. Everything after that is arithmetic.

Build the model backward from closed-won

Almost every scoring guide builds forward: here are the behaviours we can track, assign each a value, add them up. That is how you end up scoring what is easy to measure rather than what predicts revenue.

Build it backward instead.

  1. Pull your closed-won deals, 50 to 100 as a working minimum. If you have fewer, use everything you have and accept the model is provisional.
  2. Write down what was true about each account before the deal opened. Industry, employee count, region, tech in use, who the first contact was by role, how they arrived.
  3. Write down what they did. Which pages, in what order, how many people from the account, how compressed in time, which form, whether they booked something.
  4. Now pull a matched set of closed-lost and never-qualified records and do the same.
  5. Keep only the attributes and behaviours that separate the two groups. Everything else scores zero, however much your marketing team loves it.

This is the step that makes a score predictive, and it is the step most models skip. You are not guessing at value, you are reading it off outcomes you already own.

Two honest caveats. First, tools can automate scoring, but only if you feed them clean, accurate data (RevBlack). Take job titles: VP of Marketing, Marketing VP and VP - Marketing all living in the same database, with a rule matching on VP catching some and missing others. A score built on fields that are 60 percent populated is a guess with decimals. Second, you cannot do this at all if deal outcomes live in one system and the behaviours live in another. The backward build requires the same record to hold the visit, the form, the meeting and the result.

Fit and intent are two scores, not one

A single blended number is the most common design error, and it is the one that costs you the most.

Fit and intent answer different questions and move on different clocks.

  • Fit is who they are. Industry, size, role, region, budget band. It barely changes, and when it does it changes slowly. Fit tells you whether this account is worth your sales team's time at all.
  • Intent is what they are doing. Pricing page visits, multiple people from one domain in one week, a demo form, a comparison page, a reply. It spikes and it fades.

Blend them and you lose the lever. A 70 could be a perfect-fit account doing nothing, or a terrible-fit account reading everything. Those two leads need opposite treatments: one needs an outbound reason to re-engage, the other needs to be left alone. The blended number tells your rep neither.

Score them separately and route on the pair. High fit plus high intent goes to sales today. High fit plus low intent goes to nurture and account-level demand work. Low fit plus high intent is a content signal, not a pipeline signal, and treating it as pipeline is how you teach your reps to distrust you.

A score with no expiry date is not a signal

Intent decays. Most models do not model that, and it is the quietest reason scores drift into nonsense.

Someone who read your pricing page twice in March is not hot in July. They are a historical fact. But in a cumulative points model they still carry those points, sitting at the top of the queue alongside an account that visited yesterday with three people from the same domain. Your rep calls both, gets a blank response from one, and the model loses another percent of credibility.

Three rules fix most of this.

  • Decay intent on a schedule. Pick a half-life that matches your cycle. Shorter than you think.
  • Score negatively. Unsubscribes, bounced email, a job change out of the buying role, a competitor domain, an existing customer already in another motion. These should pull the score down, not sit as neutral.
  • Reward compression, not volume. Four actions in five days is a buying signal. Four actions in ten months is a subscriber.

Fit does not decay. Only intent does. Another reason to keep them apart.

What happens at 80 points?

If you cannot answer that question with a named action and a named owner, your score is reporting, not routing.

Thresholds are only real when something happens automatically at each one. Write it as a table your sales leader signs off:

  • Band A, high fit and high intent: route to the owning rep within one hour, with the behaviour trail attached, and a task created.
  • Band B, high fit and rising intent: add to an active nurture with a specific offer, and notify the rep when it crosses into Band A.
  • Band C, high intent and low fit: marketing keeps it. No sales touch.
  • Band D: no action, and say so out loud so nobody wonders.

Every band needs an owner's name, a service-level expectation, and a visible consequence when it is missed. Automate the routing so it is not a judgment call made 40 times a day. Visual workflows triggered off score thresholds are what turn a number into a motion.

One more thing belongs in that design conversation: what you are allowed to score on. What counts as permitted data varies by jurisdiction and by sector, and it can turn on things like consent status, how de-anonymised website activity is handled, and how profiling is treated in regulated categories. Your legal or compliance reviewer decides that for your situation, not a marketing meeting.

How do you prove the model works?

Score a cohort, freeze it, and come back one full sales cycle later to check win rate by band.

That is the whole test. Not open rates, not MQL volume, not how many leads crossed 80. If Band A does not convert at a materially higher rate than Band B, and Band B than Band C, the weights are wrong and you re-derive them from a fresh closed-won set. Run it quarterly. A scoring model is a living estimate, not a configuration you finish.

The part that is usually missing

Everything above assumes one condition: that the deal outcome, the website activity, the form submission, the meeting and the score live in the same place, on the same record. Backward-built scoring is impossible when the behaviours sit in a marketing tool, the outcomes sit in a CRM, and the two are joined by a monthly export.

That is the condition Payani AI is built for. Lead scoring, website visitor and company activity tracking, forms, meetings, deals and the reporting that fuses GA4 with CRM run as one system, so win rate by score band is a report you can actually pull, and the action at each threshold is a workflow that already knows who owns it.

Frequently asked questions

Why does our sales team ignore the lead scores we send them?

Usually because the scores do not match what reps see in their pipeline. The common pattern is that marketing builds the model alone, assigns point values on instinct, and hands it over finished, at which point sales stops looking at it within about a month. Trust also breaks when a model scores activity without accounting for fit, and when sales had no part in defining what a sales-ready lead is.

How many closed-won deals do I need to build a score backward?

Aim for your last 50 to 100 closed-won deals, compared against a matched set of closed-lost and never-qualified records. If you have fewer than that, use everything you have and treat the model as provisional, then re-derive the weights as more deals close. The point is to keep only the attributes and behaviours that actually separate won from lost.

Should fit and intent be one score or two?

Two. Fit is who they are and barely changes, intent is what they are doing and fades fast. Blended into one number, a 70 could be a perfect-fit account doing nothing or a poor-fit account reading everything, and those two leads need opposite treatments. Routing on the pair tells your rep what to do; a single number does not.

How do I know if my lead scoring model actually works?

Score a cohort, freeze it, and check win rate by score band 90 days later. Not open rates, not MQL volume. If the top band does not convert at a materially higher rate than the band below it, the weights are wrong and you re-derive them from a fresh set of closed-won deals.

What data am I allowed to use in a lead score?

That depends on consent status, how you handle de-anonymised website activity, and profiling rules that differ by jurisdiction and by sector, particularly in regulated categories. It is a question for your legal or compliance reviewer rather than something to settle in a marketing meeting. Decide what belongs in the model before you build it, not after.

ShareLinkedInX

Newsletter

Get growth intelligence in your inbox.

Sharp notes on growing with AI and what actually moves revenue. No fluff, unsubscribe anytime.