The quarterly business review usually goes the same way. Marketing presents the Tier A list: four hundred accounts, every one of them the right industry, the right headcount band, the right tech stack. Sales confirms that reps worked those accounts hard. Then someone pulls the closed-won report for the quarter and asks the uncomfortable question: how many of these deals came from Tier A? The answer is often fewer than half. The rest came from accounts scored B or C, accounts that were never on the list, or accounts that were on the list two years ago and quietly fell off.
Nobody in that room did anything wrong. The score did exactly what it was built to do. It ranked accounts by how closely they resemble the company's idea of a good customer. It was never built to tell anyone which of those accounts was in a buying cycle this quarter, and that is the only question a rep with forty working hours a week needs answered.
Those three numbers explain why static fit scoring fails. 6sense's 2024 Buyer Experience Report found that 81% of B2B buyers already have a preferred vendor when they first contact a seller, that buyers make that contact when they are roughly 70% of the way through their journey, and that 85% have already set their purchase requirements. By the time an account behaves in a way your CRM recognizes, most of the decision has been made. Forrester's State of Business Buying 2024 report puts the average buying group at 13 people, with 89% of purchases involving two or more departments, and found that 86% of B2B purchases stall at some point in the process. A score that watches one contact at a time cannot see a committee forming. And Salesforce's State of Sales report (5,500 sales professionals, surveyed March to April 2024) found that only 35% of sales professionals completely trust their organization's data, while reps say 70% of their time goes to non-selling tasks. When reps do not trust the score, they build their own lists on selling time.
This is not a problem you fix by adding more firmographic fields or buying another intent feed. It is a model problem. The score needs inputs that change when an account's buying behavior changes, and weights that come from evidence rather than a whiteboard session.
Diagnosis: why fit scores stop predicting
When we look at how account scores are built inside $3M to $30M ARR companies, the same four weaknesses appear almost every time. Together they produce a score that feels rigorous and predicts very little.
Fit describes eligibility, not timing
A firmographic fit score answers a real question: could this company plausibly buy from us? Industry, size, geography and tech stack are good filters for that. The ICP definition framework shows how to encode those filters in your CRM. But fit is nearly static. An account that scores 85 today will score 85 next quarter and the quarter after, whether it is evaluating vendors or has just signed a three-year contract with your competitor. When fit is the whole score, the ranking never moves, and reps learn to ignore a list that tells them the same thing every Monday.
The point values were invented in a meeting
Ask how most scoring models were weighted and the honest answer is a workshop. Ten points for the target industry, fifteen for more than two hundred employees, five for a pricing page visit, twenty for a demo request. Nobody tested whether a pricing page visit actually precedes revenue more often than a webinar attendance, or whether headcount matters at all once an account is inside the ICP. The model looks quantitative, but every number in it is an opinion.
The score watches people, but accounts buy
Many teams still score at the contact level and roll it up by adding. One enthusiastic analyst downloading six ebooks can push an account to the top of the queue, while an account where a VP, a director of operations and two engineers each visited once looks cold. With buying groups of a dozen or more people, the breadth of engagement across roles is usually a stronger signal than the depth of engagement from one person, and additive contact scoring gets it backwards. The identity resolution layer must connect those people to the same account before their breadth of engagement can be measured.
Nothing decays and nothing is checked
A webinar attended nine months ago still carries its full points. Intent surges from last spring still sit on the record. Without a clock, scores inflate over time until most accounts look warm. Worse, almost no team compares the score against outcomes, so a broken model can run for years while reps quietly work from their own spreadsheets.
The framework: a backtested four-signal account score
The lead scoring guide covers individual qualification; here the prediction unit is the account and its buying group. The replacement is a blended model with four components, each answering a different question. Fit asks whether the account could buy. Intent asks whether the account is researching the problem you solve, usually through third-party research signals such as topic surges, review site activity or hiring for relevant roles. Engagement asks whether the account is engaging with you, measured across first-party touchpoints like website sessions, product signups, event attendance and replies, and crucially counted as the number of distinct people and roles involved rather than raw activity. Recency asks how fresh all of that is, applied as a decay so that a signal loses weight as it ages.
The structure matters as much as the components. Fit works best as a gate plus a modest weight: accounts outside your ICP are excluded or capped, so a surge of activity from a student or a competitor does not reach a rep. Intent and engagement carry most of the weight because they are the parts that change. Recency is a multiplier on intent and engagement, not a separate bucket.
The weights are where this model departs from the workshop approach. They come from a backtest on your own history. You reconstruct what each account looked like at a past point in time, then check which signals were present in the accounts that went on to produce closed-won revenue and which were present in the ones that did not. A signal that appears three times as often in accounts that closed earns more weight than one that appears equally in winners and losers, regardless of how important it sounds.
Illustrative example. The numbers below are made up and rounded to show the method; they are not benchmarks or client results. Suppose you take a snapshot of 1,000 in-ICP accounts as they stood twelve months ago, and 100 of them produced a closed-won deal within the following two quarters: a base rate of 10%. You then check the close rate for accounts that had each signal at snapshot time. Accounts in the top fit band closed at 14%, a lift of 1.4 over the base rate. Accounts with a third-party intent surge on your category closed at 22%, a lift of 2.2. Accounts with three or more distinct engaged contacts across at least two functions closed at 30%, a lift of 3.0. And the same signals observed within the prior 14 days showed roughly twice the lift of signals more than 60 days old.
A simple way to turn that into weights is to score each component by how much it improves on the base rate. In this example fit contributes 0.4 of lift, intent 1.2 and engagement 2.0, so the normalized weights come out near 11% fit, 33% intent and 56% engagement, with a decay multiplier that halves a signal's value after roughly a month. Your own numbers will likely look very different. The weights belong to your market, your motion and your buyers, not to a template.
Before anyone uses the model, you validate it on accounts it has not seen. Hold back a portion of the history, score it with the new weights, and check whether the top tier actually captured a disproportionate share of the revenue. A model that puts 20% of accounts in its top tier should capture well over 20% of later closed-won deals in that tier, or it is not earning its place in the rep's day.
Implementation: six steps to a validated score
This is a sequence your RevOps team can run with the data it already has. Each step ends in something you can check.
Define the outcome you are predicting
Pick one outcome and one window: closed-won revenue within two quarters is a good default, qualified pipeline within one quarter works for longer cycles. Write it down. A score that tries to predict meetings, pipeline and revenue at once predicts none of them well. Check: everyone in the scoring conversation can state the outcome and the window in one sentence.
Rebuild point-in-time snapshots
For a past date, reconstruct each account's fit attributes, intent signals and engagement as they were then, not as they are now. This is the hardest step, because CRMs overwrite history. Field history tables, marketing automation activity logs, product event data and archived intent exports usually get you close enough. The diagnose-before-you-build playbook covers how to do this read-only, without touching production. Check: you can show what any account looked like on the snapshot date.
Measure lift for each candidate signal
Compare the outcome rate of accounts with and without each signal against the overall base rate. Include the signals you already score and the ones you suspect matter. Drop anything that shows no lift or appears on too few accounts to trust. Check: you have a ranked list of signals with lift and coverage beside each.
Set weights, gates and decay
Turn lift into weights, apply fit as a gate, and choose a decay rate based on how quickly lift fell off with signal age in your data. A suggested starting point, not a benchmark, is a half-life of 30 days for engagement and intent, adjusted once you see your own curve. Check: every weight traces back to a number from step three.
Validate on held-out history
Score the accounts you held back and measure how much of their later revenue the top tier captured. We hold every system to the same bar: tested on around 20 of the client's own past cases, and 85 percent correct or it does not ship. Check: the model clears your bar on data it was not built from, and every miss has a documented reason.
Put it in the rep's workflow and schedule the re-test
Write the score and its components to the account record, show reps why an account scored high, and route on it. Then set a quarterly re-backtest, because buyers, products and markets change. Check: a rep can see the top three reasons behind any score, and the next re-test is on the calendar.
Workflow: what each score band triggers
A score earns its keep only when it changes what people do. The bands below are a suggested structure; set the cut-offs from your own validation results so that each band has a clear owner, action and time limit.
Profile: in-ICP, recent intent surge and several engaged contacts across functions, all within the decay window.
Action: routed to the owning AE or SDR with the reasons attached, first touch within one business day, multi-threaded outreach to the roles already engaging.
Owner: sales. The SLA is tracked, and missed SLAs are reviewed weekly.
Profile: in-ICP with one strong signal, for example intent without engagement or a single engaged contact.
Action: signal-triggered outreach that references what changed, plus targeted marketing to widen engagement across the buying group.
Owner: SDR team with marketing. Accounts that add a second signal move up automatically.
Profile: good fit, no current signals, or signals that have decayed.
Action: marketing programs only, no rep time. The account stays monitored so that a new signal lifts it immediately.
Owner: marketing.
Profile: fails the fit gate, is a current customer handled by CS, or is in an active opportunity already.
Action: no outbound. Engagement from these accounts is logged and passed to the right owner rather than scored as new demand.
Owner: RevOps maintains the rules.
The board narrative
Three statements usually carry the conversation.
We stopped ranking accounts by how much they look like our best customers and started ranking them by the signals that preceded our past wins. The weights come from our own closed-won history, not from opinion, and we re-test them every quarter.
Before rollout, we scored accounts from a period the model had never seen and measured how much of that period's revenue the top tier captured. We report that capture rate each quarter alongside pipeline, so the board can see whether the model is still earning its place.
Rep time now goes first to accounts showing active buying behavior, and accounts without signals are handled by marketing programs. The measures to watch are pipeline per rep hour, conversion from top-band accounts to opportunities, and time from first signal to first touch.
Cross-domain: how account scoring connects to the other systems
Account scoring is not a standalone project. It is the prioritization layer that several revenue systems depend on. The Signal-Based Outbound Engine uses the same intent, engagement and recency inputs to decide which accounts get outreach and what that outreach says, so a validated score directly improves who it contacts and when. Speed-to-Lead uses the score to decide how fast and to whom an inbound request goes; a hand-raiser from a Band 1 account deserves a different response than one from an excluded account.
Further down the funnel, the Pipeline Hygiene Sentinel and the Forecast Assistant benefit from knowing whether an opportunity began in a high-signal account or was forced into the pipeline from a cold list, because those two kinds of deals behave differently. And the same four-signal logic, pointed at existing customers, is the foundation of expansion and risk scoring in the Churn Signal Watchtower. The wider picture of how these fit together is on the GTM Operations page.
If you are deciding who should own the model, the GTM engineer vs. RevOps manager vs. growth engineer decision tree helps with that call. Our own approach is forward-deployed engineering: start with one outcome, backtest on your own history, and put the score in front of reps only after it has proven itself on deals you have already won and lost.
Sources: 6sense, 2024 Buyer Experience Report (global survey of B2B buyers across North America, EMEA and APAC, October 2024). Forrester, The State of Business Buying, 2024 (December 2024). Salesforce, State of Sales, 6th edition (5,500 sales professionals, surveyed March to April 2024, published July 2024).
