The intent scoring framework that does not break at scale

Scoring rules written on the day a CRM was implemented rarely get touched again, which is why most contact records carry a score nobody in the building would defend if asked to explain it.

What is intent scoring?

Intent scoring ranks contacts and accounts by how close they are to buying, using what they have actually done rather than who they are on paper. A pricing page revisited twice says more about readiness than a job title ever will.

A score left unattended is not a system, only a number sitting next to a contact record nobody opens. It earns its keep once something consistent decides what counts, what it is worth, and when to act on it, and that consistency is a framework rather than a one-off exercise.

What is an intent scoring framework?

An intent scoring framework is the set of rules that converts raw buyer behaviour into a priority order: which signals count, what each is worth, how fast that worth decays, and what threshold puts an account in front of a person.

Scoring models fail in a predictable order. They get set once from a vendor template, they never decay, they collapse two unrelated questions into a single number, and once volume arrives they rank the wrong accounts with total confidence.

A framework has three parts, and the score is only the middle one.

• Admission: which behaviours enter the model, and which are excluded as unactionable
• Weighting: what each behaviour is worth relative to the others, and how quickly that worth falls away
• Decision: what the total triggers, who receives it, within what time, and what happens on a week when nothing clears

Teams build the middle part and skip the other two. Points get attached to a handful of events, a number appears on the contact record, and nothing downstream changes.

That last clause is the one frameworks omit. A model that always produces a top ten produces one on quiet weeks too, and a list of the ten least cold accounts is a different object from a list of warm ones. An honest empty state does more for a rep's trust than a padded list.

What is the difference between lead scoring, lead grading, intent scoring and predictive scoring?

Four mechanisms get described with the same vocabulary, and they answer different questions from different inputs.

Two of those get merged constantly and should not be. Grading is a property of the account and moves over quarters: industry, size, the tools already in place, whether the business can legally be sold to. Scoring is a property of the moment and ought to be close to worthless 14 days later.

Intent scoring is lead scoring with two corrections: behaviour weighted by proximity to a purchase rather than by ease of capture, and every point timestamped so its value falls as it ages. The full signal taxonomy, with directional starting weights, is in the working guide to B2B buyer intent signals.

What goes into an intent score?

Three decisions produce the number: what gets admitted, what each admitted thing is worth against the others, and how fast that worth falls away. The third is the one teams skip.

Which behaviours belong in the model?

Selection usually runs on availability. Whatever the CRM already records gets points, which is a reason to know what the rest of the stack is holding and never sends over, set out in how to track buying signals across tools, and the model ends up weighted towards email opens, which are cheap to capture and, since Apple's Mail Privacy Protection began loading images automatically, largely uninformative.

The defensible axis is proximity to money. Rank each behaviour by how few steps separate it from a signed contract: a call rescheduled with a named owner sits one step away, a repeat pricing visit two or three, a category surge bought from a third-party data provider outside the building entirely.

Behaviour that cannot be acted on scores zero regardless of strength, from a contact with no reachable address to a competitor's analyst reading the pricing page. Which behaviours are observed on owned property, and which are bought in, is worked through in what first-party intent data is and when third-party helps. Filtering for actionability clears a meaningful share of the top of most queues.

How are the weights set without a data team?

Only the ratio matters. Whether a demo booking scores 25 against a pricing visit at 20, or 250 against 200, changes nothing about the order. Arguing about the absolute numbers is how scoring projects stall.

Three tiers are enough to start with: behaviour indicating a decision is being made, behaviour indicating research, and behaviour indicating nothing beyond presence. Set roughly 5 to 1 between the top tier and the bottom, then adjust once outcomes are available.

Negative weights are the missing half. A model built inside a CRM can usually only add, so a cancelled call with no rebooking raises a contact's score: the booking earned points and the cancellation earns nothing. Anything reducing the likelihood of a purchase has to reduce the number.

How fast should a signal lose value?

A score with no decay is a lifetime activity total. The contact who has been on the list three years and opens everything outranks the contact who booked a demo yesterday, and the model is working exactly as configured.

Decay is easier to reason about as a half-life than as a daily subtraction. These are starting points for tuning, not benchmarks:

• Presence, such as a single homepage visit or an email open: worth little on the day, nothing 72 hours later
• One step from a decision, such as a rescheduled call or an abandoned booking: half-life of roughly 48 hours
• Research, such as repeat pricing visits or documentation reads: half-life of 7 to 14 days
• Category research bought from a third-party data provider: half-life of 30 days or more, and never a reason to contact anyone alone
• Company triggers, such as funding or a new sales leader arriving: half-life of one to three months

Stale signals have to reach zero rather than a floor. Otherwise the model builds a sediment of old activity that outweighs anything happening now.

How is a scoring framework built from closed-won deals?

The first version comes from reading what happened in the weeks before the deals that closed, against the weeks before the deals that did not.

1. List the last 20 closed-won deals and the last 20 closed-lost. Losses are half the exercise. A behaviour appearing in every win and every loss carries no information.
2. Read the 30 days before the first meeting on each. Website records, email threads, calendar entries, call notes, written down without interpretation.
3. Score each behaviour on lift, meaning how often it appears in wins minus how often it appears in losses. That difference is the only defensible source of a weight at this stage.
4. Convert lift into a ratio and stop there. The gap between the top behaviour and the bottom is the whole output. Absolute points are decoration.
5. Write the decision rule down. Who receives the alert, in which channel, within what time, and what happens on a week when nothing clears.

The fit half of the same exercise, turning that deal list into a scorecard of account attributes, is in how to build an ICP scorecard from closed-won deals.

What threshold counts as ready?

Templates use 70 and 70 is arbitrary. The right threshold is the one that produces the number of accounts a team can genuinely work in a week.

Two quota carriers running six proper first touches a day cannot work 300 flagged accounts. Flagging that many produces a backlog, and a backlog teaches reps the flag means nothing, which is how a model gets abandoned over volume before anyone questions its accuracy.

Sort last quarter's accounts by score, find the point in the distribution that yields roughly the weekly capacity, and set the threshold there. Capacity sets the threshold, and accuracy decides whether the accounts above it are the right ones. Raise it when the top of the list stops converting.

How often should the score be recalculated?

Continuously, or at minimum every 15 minutes during working hours. A nightly job hands a rep a signal that fired up to 24 hours before, which records something rather than prompting anything.

Every hour between a behaviour and a response is an hour in which a competitor replies, an internal priority displaces the work, or a budget cycle closes.

Join the MeetIQ waitlist →

What breaks as a scoring model grows?

A framework that works on 40 accounts a week does not automatically work on 400. Two failures arrive as it grows, and neither announces itself.

Why adding a criterion changes what every old score meant

A model that sums raw points has no fixed ceiling. Add a sixth criterion worth 15 points and every account can now reach 115, so the threshold set last quarter has quietly become a lower bar, and nothing reports the change.

Normalising to a 0 to 100 scale removes the problem. Each criterion contributes a share of a fixed total, so adding one changes which accounts rank highest without redefining what qualifies as ready.

Why an unaudited model looks accurate right up until it is not

Weights derived from last year's closed-won deals describe last year's buyers. The product changes, a competitor repositions, and the behaviours that predicted a win stop predicting one, while the numbers on the contact records look just as confident as before. The check belongs in somebody's calendar, because nothing in the model reports the drift on its own.

Opened an email, plus one. Downloaded a guide, plus five. Those exact rules were written the day the CRM was implemented, nobody has touched them since, and nothing can say which of the contacts sitting at 60 this morning is worth a call.

Do teams need predictive scoring to do this?

Predictive scoring is supervised learning: a model is shown past deals with their outcomes and learns which attribute and behaviour combinations correlated with a win. The requirement is unavoidable: no model learns a pattern from a sample that does not contain one.

HubSpot's own documentation places predictive lead scoring in its Enterprise tiers, and the constraint holds wherever it is sold. A business with 18 closed-won deals and patchy CRM hygiene has nothing for a model to learn from, and the output will be confident and arbitrary.

Three rules matter more than the choice of method:

• Both outcomes are needed. Wins alone teach a model what customers look like, not what separates them from the accounts that said no.
• Attributes must be captured as they stood at qualification, never read back after the sale. A field populated during onboarding correlates perfectly with winning and predicts nothing.
• A rules model reviewed monthly beats a predictive model with nothing to train on, and it is explainable to the person expected to act on it.

Explainability is not a nicety here. A score of 74 arriving with the sentence "two people from this account read the pricing page twice this week" gets acted on. Stripped of the reasoning, the same number gets treated as weather.

Fit and engagement are two scores, and blending them is what breaks at scale

Almost every implementation ends with one number per contact, because one number sorts a list and two do not. That number is being asked two unrelated questions at once: whether this is the right kind of customer, and whether they are interested right now.

Call the alternative the two-axis scoring model. Fit is scored 0 to 100 at account level, from attributes that move over quarters. Engagement runs on its own 0 to 100 scale at contact level and decays weekly.

Neither number is ever averaged into the other: not in the interface, not in the export, not in a database column.

The bottom right cell is the whole argument. A competitor's analyst, a candidate reading up before an interview: both produce high engagement against zero fit, and a blended score sends both to a rep, because strong behaviour drags an average upwards.

A single blended number cannot separate an account that is perfect and quiet from one that is unqualified and busy. Both land in the middle of the list, and the rep who works two of the second kind stops trusting the list.

Three rules run the model.

1. Fit is a gate. It decides whether an account is worth anybody's time at all.
2. Engagement decides when, and carries no information about worth.
3. An account reaches a rep only when both scores clear their own thresholds. Neither compensates for the other.

Splitting the axes makes failure diagnosable too. A queue filling with high-engagement, low-fit accounts means the targeting is wrong and the scoring works correctly, while high-fit accounts that never engage point at demand generation rather than the model. One blended number reports both as a healthy average.

What can be checked without new tooling?

Three checks, none needing a purchase or a project.

1. Open the scoring rules and read the date they were last edited. If that date is the month the CRM was implemented, the model describes a business that no longer exists.
2. Pull the ten highest-scoring contacts from last month that never booked anything. Read what the model saw in each. The failure is usually one weight, and that weight is usually email opens.
3. Count how many accounts cleared the threshold in the past seven days against how many a person actually contacted. The gap is the backlog the model is manufacturing, and it gets fixed before any weight is touched.

The five signals carrying the most predictive weight in a small stack are broken down in the five B2B intent signals that predict purchase readiness.

Scoring is a maintenance problem rather than a modelling one at this size. Weights derived once from real deals, decay that runs on a schedule, a threshold set from capacity, and an hour a month reading the accounts the model got wrong. None of that needs a data scientist or another licence.

What it does need is somebody's name against it. The rules in most CRMs were written by a person who has since moved on, and they will carry on producing numbers, in confident order, for as long as nobody opens them.