Account scoring has a bad reputation in sales teams, and the reputation is earned. The typical experience is a number appearing on a record, nobody being able to say what produced it, one high-scoring account turning out to be a student doing research, and the field quietly reverting to their own judgement forever after.
The failure is almost never the model's arithmetic. It is that the model was delivered as an oracle rather than as an argument.
Why scores get ignored
Three specific reasons, in order of how often I see them.
The score cannot be expanded. If a rep cannot click an 87 and see the six things that produced it, the score is asking for trust it has not earned. Every score in a system reps use has to decompose into human-readable events with dates.
It was calibrated on the wrong outcome. Many models are trained to predict conversion to MQL, because that is the data marketing has. Sales does not care about MQLs. A model that predicts "will accept a meeting" or "will open an opportunity" is a different model, and it is the one worth building.
It has no memory. A rep calls an account showing 91 on Tuesday. By the time the call notes are written the score is 74 because a decay function ran overnight, and the rep now looks like they invented the reason for the call. Score history is not a nice-to-have.
Four signals worth weighting
For a paid-led ABM programme, the signals that carry predictive weight are consistent across most B2B categories I have looked at.
Committee breadth — the strongest single signal
Distinct job functions engaged within the window. The move from one function to two is the most informative transition in the whole model, more so than the move from two to three. One person engaging is curiosity; two people in different functions is an internal conversation, and internal conversations are what precede purchases.
Weight it around 30% and count functions, not people. Three engineers is one function.
Depth — how much they took
Touches per engaged contact, weighted by asset commitment. A document ad opened and read is worth several video views; a fifteen-second view of a thirty-second video is worth very little. Around 20%.
Seniority — with a caveat
Director-and-above involvement, roughly 20%. The caveat: do not weight it so heavily that an engaged analyst team scores below a single VP who watched three seconds of a video. In most complex purchases the analyst does the evaluation and the VP arrives late to approve it.
Recency — a curve, not a cliff
Around 15%, decayed smoothly. A hard 30-day cliff creates the Tuesday-to-Wednesday collapse that destroys rep trust. A gradual half-life — something in the region of three to five weeks for most B2B cycles — behaves far better.
The remaining weight goes to bottom-funnel creative engagement and, if you have a site tag, visits to pricing and documentation pages. Both are strong signals with low volume, which is exactly what a small weight is for.
What not to include
Firmographic fit. This is the big one. Company size, industry and technology stack tell you whether an account should buy. Engagement tells you whether it is moving. Blend them into one number and you produce a score where a perfect-fit account that has never engaged outranks an imperfect-fit account whose whole committee is reading your documentation. Keep two axes: fit and engagement. Sales prioritises on the combination, and can see which is which.
Raw impression counts. Impressions are something you bought, not something they did. An account you served forty times because it was cheap to reach is not more interested than one you served eight times. Score engagement; report impressions separately.
Anything you cannot show the rep. If a signal cannot be rendered as a sentence with a date on it, it does not belong in a score that a human is expected to act on.
Calibration
You need three months of history and a defined positive outcome before the model means anything. Pick the outcome that sales recognises — first meeting accepted, or opportunity created — not the one marketing has most data on.
Then do the uncomfortable check: take accounts that produced the outcome and look at what their score was two weeks beforehand. If the distribution is flat, your model is not predictive and no amount of weight-tuning will rescue it; you probably do not have enough engagement volume for scoring to work yet, and should say so rather than shipping a number.
Handing it to sales
Four things to do at handover, all of which are about credibility rather than technology:
- Show the model before the scores. Walk the team through the four signals and the weights. Let them argue. Change one thing they push back on, if it is defensible — joint ownership is worth more than optimal weights.
- Send the expansion, not the number. The CRM field should be accompanied by the three top contributing events. "Score 87 — Head of Ops, VP Eng and a finance director all engaged in the last eleven days" is actionable. "87" is not.
- Publish the threshold flow rate. Tell the team how many accounts to expect per week. Surprises destroy trust faster than inaccuracy.
- Review the misses monthly. Take the five highest-scoring accounts that went nowhere and the five opportunities that opened with low scores. Both lists teach you something, and running the review in front of sales is what turns the model from marketing's tool into a shared one.