Lead scoring: ranking behaviour, or assessing quality at source

Lead scoring helps sales prioritise, but the result can disappoint when the model rewards activity without enough context. The model scores a newsletter skimmer highly because they opened five emails, and misses a real buyer who filled in one form and went straight to a call. Sales stops trusting the numbers. Marketing re-tunes the weights again, but re-tuning does not repair the records or change when their quality was first assessed.

Lead scoring is the practice of assigning a value to each lead to estimate how likely it is to convert, using a mix of behavioural activity and firmographic fit. Its accuracy is capped by its inputs and timing. The usual limit is not the algorithm: much of the quality the model later tries to infer was knowable at capture. Assessing some of that quality at capture makes the model that reads the record later more accurate.

What lead scoring does

A scoring model looks at what a lead has done and who they are, and turns that into a number that estimates conversion likelihood. Behavioural scoring adds points for actions: an email opened, a pricing page viewed, a demo requested. Firmographic scoring adds or removes points for fit: the right industry, company size, or job title. Predictive or AI scoring does the same job but learns the weights from past conversions rather than having a person set them by hand.

Scoring puts leads in a rough order so that limited sales time goes to the most promising. That is a real benefit: a team without scoring usually works leads in arrival order, which is worse. For prioritising a large volume of leads, scoring is hard to beat.

The limits of behaviour-based scoring

Behaviour-based scoring reads activity that has already accumulated. Unlike accumulated activity, whether the contact gave valid consent to be marketed to, where relied upon, was settled at the form. The accuracy of the email and role could be checked at capture and may have decayed since. Company fit was knowable from the first submission. The model infers quality from clicks when much of it could have been assessed directly at the start.

This is also where behavioural scoring shows its well-documented weaknesses. Low-intent actions like email opens tend to be over-weighted, so a diligent newsletter reader, or an automated inbox that opens everything, can accumulate enough points to look like a buyer. Many models are blind to recency, scoring an action from a month ago like one from this morning. Scores can be gamed, deliberately or accidentally, by anything that generates activity without intent. And firmographic rules misfire, penalising a real buyer who used a generic email domain or trusting a self-reported job title that was aspirational. None of these is fatal on its own, but together they are why David Campbell of SugarCRM, writing in a 2021 contributed article for Demand Gen Report, described conventional scoring’s ability to predict which leads will convert as “pretty unpredictable”. Predictive and AI models sold since that 2021 assessment have not resolved the underlying data problem, which is why the complaint remains relevant.

Why scoring depends on data quality

The deeper limit is the state of the fields the model reads. B2B contact data does not sit still. The enrichment firm Cleanlist, in a 2026 analysis citing Dun and Bradstreet, puts the decay at roughly 22.5% a year, around 2% a month, and higher in some industries. Cleanlist sells data enrichment, so it has an interest in the alarming version of that number. Its analysis names the Dun and Bradstreet benchmark but does not describe how that figure was produced. The direction is not in doubt: people change jobs, companies restructure, and email addresses stop working. A record scored this quarter is often no longer the record that was captured.

Scott Brinker is quoted on the same problem: when a fifth of your records are outdated, your lead scoring, routing and forecasts are “all built on a foundation of lies”. The CRM publication CX Today makes a related argument, noting that scoring and routing “run on stale fields” and that automation, including AI, tends to scale bad data rather than fix it. Its blunt summary is that “AI will not fix bad data. AI will scale it.”

Changing a weight alters the influence of a demo request or seniority match on the final score. It does nothing about whether the seniority field is still accurate or the email address still works. A model tuned to perfection on Monday is tuned against fields that are slightly more wrong by Friday, and more wrong again by the end of the quarter. The limiting factor is data truth: whether the underlying records remain accurate.

Scoring one contact in a group decision

There is a second structural mismatch. Individual behavioural scoring ranks one person’s activity, but a B2B purchase is rarely one person’s decision. Gartner’s ongoing research on the B2B buying journey finds that around 99% of B2B purchases are set off by an organisational change, and that buying is non-linear, looping across a set of buying jobs rather than moving down a tidy funnel. The multi-actor, event-driven shape of B2B buying is well established.

That has a direct consequence for scoring. The contact who fills in the form may be a researcher, a champion, a blocker, or a procurement gatekeeper, and their individual click trail says little about where the account as a whole sits. What a captured record tells you about fit and buying-group role is often more durable than the activity trail of whichever individual happened to fill in the form. Scoring the person is not wrong, but treating that score as a read on the account overstates what one contact’s behaviour can carry.

The practical implication is to weight what the form tells you about the account and the person’s role alongside the behavioural score, rather than letting clicks dominate. A senior contact from a target account who submits once is often a stronger signal than a junior researcher from a poor-fit company who opens every email, yet a click-weighted model can rank them the other way round. Distinguishing declared, implied and inferred signals helps separate what the form tells you from what other signals suggest. For scoring, the declared details captured at the form are frequently the most reliable part of the record, available before any behaviour has accumulated.

Assessing quality at capture

If much of lead quality is set or knowable at capture, the constructive move is to assess it there, before CRM entry. The checks cover three areas:

• Compliance: the basis and permissions for the intended use. Where consent is required, the ICO’s guidance says it must be freely given, specific, informed and unambiguous, given by a clear affirmative action. Consent is invalid without clear records, when obtained through a pre-ticked box, or when the controller was not identified.
• Contact truth: whether the email works and the role is current.
• Value: what the account and role indicate about fit and participation in the buying group.

These checks supply evidence for scoring. They do not replace the model or remove the need to review records later.

That gives a team something to measure. Call it the input-validity rate: the share of the records a model scores whose compliance and contact fields were validated at capture or re-verified since. Put plainly, it is the number of scored records with a current validation flag divided by the total number of scored records. On a base of 40,000 scored records, if only 26,000 carry a validation check within the decay window, the input-validity rate is about 65%, leaving roughly a third without a current check. That does not prove those records are wrong; it identifies missing or outdated validation evidence. The metric has one honest limit, in that it measures input hygiene rather than the model’s predictive lift, so read it alongside the model’s own accuracy rather than as a substitute for it. As a first diagnostic, though, it answers a question most teams cannot currently answer: how much of my scoring is running on data I have actually checked.

A low rate points at a specific fix rather than a vague one. If a third of scored records carry no current validation, the useful response is not to re-weight the model but to add a check at the form and a re-verification step for ageing records, so that the share of validated inputs climbs over time. The fix is better inputs rather than a cleverer model: the score improves because the evidence it reads is more reliable. The lead quality definitions and the minimum viable lead thresholds are where those validated inputs get their meaning.

Where to focus improvement

Scoring is worth doing. What usually limits it is the quality of its inputs and the timing of their assessment. Re-tuning cannot repair a stale role, missing consent evidence or an account represented by the wrong contact; those problems need better inputs.

Most of that fixing happens before the record reaches the CRM, at the point of capture, where a validation layer such as the LeadScale Engine applies on arrival the compliance, truth and value checks selected for the programme. Validated, current inputs mean the score the model produces later is a read on something real, and the model can do the prioritising job it is good at without inheriting a foundation of stale data.

If you want a concrete next step, measure your input-validity rate. Take the records your model scored last month, count how many carried a current validation check on their key fields, and divide by the total. That number tells you how much of your scoring is running on data you have confirmed, and whether your next move is a change to the model or a change to what happens at the form.