How to automate lead scoring: a model nobody validates

A lead score is a prediction: this person is more likely than that one to become a customer. Almost nobody treats it as one. The weights are set in a meeting, they are rarely revisited, and the question of whether a score of eighty actually converts better than a score of forty usually goes unasked for years.

The automation is trivial — adding points to a record is not the hard part. What makes a scoring system worth having is the validation, and building that from the start is what distinguishes a model from a number the sales team has learned to ignore.

What we'll cover

• Why most scores predict nothing
• Fit and intent are different scores
• Choosing signals that mean something
• Building it with CodeWords
• Decay, and why it is essential
• Validating the model
• What the score should trigger
• Getting sales and marketing to agree on it
• Making it survive
• Limits worth knowing about
• What to build first
• Frequently asked questions

Why most scores predict nothing

The weights were guessed. Somebody assigned ten points to a pricing page visit because it felt significant.

Nobody checks against outcomes. The score exists; whether it correlates with closed business is not measured.

Everything is additive, so a poor-fit lead accumulates a high score by browsing, which is exactly backwards.

Points never expire, so activity from eight months ago counts as much as yesterday's.

The scale is uncalibrated. Nobody can say what eighty means, so each rep applies their own interpretation.

Sales stops trusting it, at which point the system continues running and influences nothing, which is the most common end state.

Fit and intent are different scores

Combining them into one number is the single most common structural error, because the two require opposite responses.

Fit is about whether they should buy — size, industry, role, geography, technology. It changes slowly and is largely known up front.

Intent is about whether they are buying now — page visits, content downloaded, demo requested, return frequency. It changes daily.

Good fit with no intent is a nurture case, and reaching out hard is premature.

High intent with poor fit is usually a student, a competitor, or somebody who will not buy. A single number cannot distinguish that from a real opportunity.

Both high is the only case that warrants immediate contact, and it is a smaller group than a combined score suggests.

Report them as a pair, since two coordinates are more informative than any weighted sum of them and are far easier for a rep to act on.

Choosing signals that mean something

Prefer what they did over who they are for intent, and the reverse for fit.

Weight by rarity. A pricing page visit is common; requesting a technical evaluation is not, and the rare action carries more information.

Count repeat visits from the same account rather than the same person, since business purchases involve several people.

Watch for multiple people from one company, which is one of the strongest available signals and is frequently not captured at all.

Negative signals are real — a personal email address for enterprise software, a competitor domain, an unsupported region — and most systems have none.

Do not score what you cannot act on. A signal that changes nothing about what anybody does is a field, not a score component.

Building it with CodeWords

CodeWords connects to more than 3,000 integrations, so the CRM, the site analytics, the enrichment sources and the sales channels can meet in one automation.

1. Open CodeWords and start a new automation.
2. Describe what should happen in plain language to Cody, the automation builder: the fit and intent signals, the decay, and what each threshold triggers.
3. Authorize the connections — the CRM, the analytics source, and your enrichment provider.
4. Describe the exceptions: an existing customer, an open opportunity, a competitor, a personal address.
5. Score last year's leads retrospectively and compare the scores against what actually closed.

You describe the outcome; Cody builds it, connects it, and deploys it. The free plan covers light use, with Pro at $39 per month and Business at $100 per month as usage grows; details are on the pricing page.

Decay, and why it is essential

Intent is perishable. Somebody researching three months ago has made a decision, and it may not have involved you.

Decay the intent score over weeks, not months, so that the ranking reflects who is active now.

Do not decay fit, which is a property of the company rather than of a moment.

Without decay, the top of the list calcifies — it fills with people who were interested last quarter and stays that way.

Recency should visibly outrank volume. One strong action this week beats fifteen page views in March, and a purely additive model gets that backwards.

Reset on a real event. A closed-lost opportunity or an explicit "not now" should clear intent rather than leaving the accumulated score behind.

Validating the model

The step that turns a scoring system into something worth having, and the one that is almost always skipped.

Score historical leads retrospectively using only the data available at the time, and check whether the score separates the ones that converted.

Compare against a baseline — random ordering, or simply company size — because a model that does not beat the obvious alternative is not earning its complexity.

Measure separation, not accuracy. The question is whether the top band converts meaningfully better than the middle, which is what a rep's time depends on.

Track it forward, by band, continuously, so drift is visible.

Ask the sales team which scored leads were wrong and in which direction, since their feedback identifies the missing signal faster than any analysis.

Revisit quarterly. Your market, product and traffic mix all change, and a model tuned a year ago is describing an audience you no longer have.

What the score should trigger

A specific action per band, not just a label. Immediate contact, a nurture sequence, or nothing at all.

Routing, since the best-fit leads should reach the people most likely to close them.

Context with the handoff. A rep receiving a lead needs what it did, not the number — the score decides priority, the activity decides the conversation.

Alerting on a threshold crossing, particularly the multi-person-from-one-company signal, which is time-sensitive.

Suppression for existing customers and open opportunities, which is the most common embarrassing failure.

Nothing at all for the low band, which is a legitimate outcome and is better than a nurture sequence nobody reads.

Getting sales and marketing to agree on it

A scoring model is also a negotiated definition of what counts as a good lead, and most of the friction it produces is about that rather than about the mathematics.

Define qualified jointly, in writing, before the weights exist. Most disputes about lead quality are disputes about this definition going unstated.

Agree the volume expectation. A model tuned to produce fifty high-band leads a month when sales can work twenty is generating rejection, not qualification.

Make rejection cheap and recorded. A rep should be able to send a lead back with a reason in one click, and those reasons are the highest-quality training data available.

Review the rejections together, monthly, which is a more productive meeting than reviewing the volume.

Show conversion by band to both sides, since a shared number stops the argument being about whose impression is right.

Change the model, not the target. When the definition shifts, update the weights rather than asking sales to work leads they have already told you are wrong.

Making it survive

Make scoring idempotent and recomputable from the underlying events rather than incrementally mutated, so a bug does not corrupt every record permanently.

Version the model and record which version produced a score, or historical comparison is meaningless.

Handle enrichment failures by scoring on what you have and flagging the gap, rather than treating missing data as a negative.

Respect data protection, since profiling individuals carries obligations in several jurisdictions including the right to an explanation.

Report the outcome: score distribution, band transitions, conversion by band, and model version.

Limits worth knowing about

Anonymous traffic cannot be scored until somebody identifies themselves, which is most of your visitors.

Enrichment data is often wrong, particularly company size and industry, and wrong fit data is worse than no fit data.

Small volumes cannot validate a model — with a few dozen conversions a quarter, band differences are noise.

Scoring does not create demand. It reorders a queue, which helps when the queue is long and does nothing when it is empty.

Automated profiling has regulatory implications in some markets, which is worth checking before the model influences decisions about individuals.

What to build first

The retrospective check. Take last year's leads, score them with a simple fit-and-intent model using only data available at the time, and see whether the top band converted better than the rest. It requires no production change, and it tells you whether a scoring system is worth building at all — which for some businesses it is not.

Two habits make the difference. Keep fit and intent as two numbers rather than one, since collapsing them produces exactly the confident wrong answers that teach sales to ignore the score. And decay intent over weeks, because without it the top of the list fills with people who decided months ago and the whole ranking stops meaning anything.

Frequently asked questions

Why do most lead scoring models not work?

The weights were guessed and never checked against outcomes, everything is additive so poor-fit leads accumulate high scores by browsing, and points never expire. Sales stops trusting the number and the system keeps running without influencing anything.

Should fit and intent be one score or two?

Two. Good fit with no intent is a nurture case; high intent with poor fit is usually a student or a competitor. A single number cannot distinguish those, and each requires an opposite response.

What signals actually carry information?

Rare actions over common ones, repeat activity at the account rather than the person level, and several people from one company engaging — which is among the strongest signals and is frequently not captured at all.

Why does intent need to decay?

Because intent is perishable. Without decay the top of the list calcifies with people who were interested last quarter, and recency stops outranking volume — which is exactly backwards.

How do I know whether the score predicts anything?

Score historical leads using only data available at the time and check whether the top band converted better than the middle. Compare against a baseline like company size; a model that does not beat the obvious alternative is not earning its complexity.

Should negative signals be included?

Yes, and most systems have none. A personal email address for enterprise software, a competitor domain, or an unsupported region should all reduce the score rather than being ignored.

What should a rep receive along with a lead?

The activity, not the number. The score decides priority; what the lead actually did decides the conversation. A handoff carrying only a score produces a generic first call.

Are there legal considerations with lead scoring?

Automated profiling of individuals carries obligations in several jurisdictions, including in some cases a right to an explanation. Worth checking before the model influences decisions about people.

How do I stop arguments about lead quality?

Define qualified jointly and in writing before setting any weights, make rejection one click with a recorded reason, and review the rejections together monthly. Most disputes about lead quality are disputes about an unstated definition.

What if the model produces more leads than sales can work?

Then it is generating rejection rather than qualification. Tune the bands to the capacity that exists, since a high band nobody has time to call is indistinguishable from no score at all.

Related reading

• Automating lead capture from a website
• Automating sales pipeline updates
• HubSpot and CodeWords