How to build a B2B lead scoring model that survives contact with your reps

Most lead scoring projects fail the same way. The model is built carefully, it produces numbers, the numbers go into a queue, and within about three months the reps have quietly gone back to working their own list. Nobody announces this. The dashboard still says the model is running.

The failure is almost never the maths. It is that nobody could see why a prospect scored what it did, so nobody could argue with it — and a ranking you cannot argue with is a ranking you eventually ignore.

This is the model we use, the reasoning behind each weight, and the specific ways these projects die.

What scoring is actually for

Not prediction. Ordering.

A scoring model’s job is to answer one question: of the 1,500 rows in this list, which fifty should someone contact this week? It does not need to know who will buy. It needs to be better than the order the list arrived in, which is almost always alphabetical, or worse, the order a scraper found them.

That framing matters because it sets the bar. You are not competing with omniscience. You are competing with a rep skimming the first hundred rows and giving up, which is what happens to every unsorted export.

It also sets the success criterion: the model works when the people using it stop overriding it. Not when it hits an accuracy target.

The nine factors

Nine is a deliberate number. Small enough that one person can hold the whole model in their head and sanity-check a change before making it; large enough to capture what actually varies between prospects.

Factor	Default weight	What it asks
Role / decision authority	20	Can this person say yes, or get to whoever can?
Industry match	15	Is their sector one you actually serve well?
Services-to-needs alignment	15	Does what you sell map onto a need they visibly have?
Company size fit	10	Are they the size you serve profitably?
Geography match	10	Can you serve them where they are?
Technology relevance	10	Does their stack or tooling imply a fit?
Outsourcing probability	10	Do they buy this kind of capability, or build it?
AI / digital transformation fit	5	Are they investing in change you benefit from?
Contact info completeness	5	Can you actually reach the named person?

They total 100 on purpose. A model that sums to an arbitrary number invites arbitrary additions.

Two of these are unusual and worth defending.

Outsourcing probability is absent from most models and does more work than its default weight suggests. A company that builds everything in-house is not a slow prospect, it is a non-prospect, and no amount of seniority or sector fit changes that. For marketing agencies selling to other agencies it carries 22 points rather than 10, because reselling outsourced capability is the entire business model.

Contact info completeness looks like a data-quality metric rather than a fit signal, and in most industries that is roughly true. In healthcare it is not: gatekeeping means a generic inbox almost never reaches the named person, so an incomplete record is not a weaker prospect but an unreachable one.

Setting the weights

Start with the defaults. Run one real list. Then change two or three things, not nine.

The useful exercise is not theoretical. Take the fifty prospects your best rep would have called from a list, score the whole list, and look at where those fifty landed. If they are scattered through the middle, the weights are wrong. If they cluster near the top, the weights are close and you are arguing about ordering within the top decile, which is a much better problem.

Three rules that save a lot of time:

Weight what varies. If every prospect on your list is in the same country, geography is worth zero points however important it feels. Points spent on a constant are points not spent on a discriminator.

Do not weight what you cannot observe. Budget is the classic trap. It is the most predictive thing imaginable and you almost never know it at scoring time. A factor you have to guess at adds noise dressed as signal.

Move points, do not add them. The total stays 100. Raising one factor means lowering another, which forces the question “less important than what?” — the question that makes the model coherent.

The threshold is a separate lever

The most common diagnostic error is conflating the weights with the qualification threshold. They fix different problems, and adjusting the wrong one is how a model accumulates six months of well-meant tweaks and no clearer story.

• Too few prospects qualifying? That is usually the threshold, not the weights. Your market may simply be narrower than the threshold assumes.
• Prospects qualifying but not converting? That is the weights, or the ICP behind them. Lowering the threshold will make it worse.

A 70% default threshold is a reasonable starting point. What matters more is that everything below it is parked rather than deleted, and re-scored automatically when the ICP, the weights or the prospect’s own data change. A prospect who fails in March can qualify in June without anyone re-importing them.

Why every score needs a written reason

This is the part that decides whether the model is still in use next year.

A score of 89 tells a rep nothing they can act on or dispute. “A senior technology decision-maker at a mid-size retail SaaS business in a target country, with a product-modernisation need that maps to work you have delivered before; only the incomplete contact record and a partial technology signal held it below 90” tells them three things: why it is high, what would make it higher, and whether the machine has understood their market.

The second and third matter most. When a rep disagrees with a score and can see the reasoning, one of two useful things happens: they learn something about the model, or the model learns something about the market. When they disagree and cannot see the reasoning, they route around it — and you will not find out for a quarter.

This is why every score Dizenly produces is stored with its explanation, and why the per-factor contribution is visible rather than rolled into a single number.

Five ways scoring models die

1. The blended ICP. One profile quietly covering three buyer types produces mediocre scores for all three. If your ICP contains the word “or” more than once, it is probably three ICPs. Splitting them fixes more scoring problems than re-weighting ever does.

2. Chasing the last deal. A big win closes, and the model is re-weighted to look like that customer. Three of those and the model describes your recent history rather than your market. Changing weights is cheap, which is exactly why “cheap to change” and “wise to change” have to stay separate questions.

3. Scoring on unobservable factors. Budget, urgency, internal politics. All highly predictive, none knowable at scoring time.

4. No threshold discipline. Lowering the threshold whenever the queue looks thin, until it is 40% and means nothing.

5. The black box. Covered above, and the one that actually kills most of them.

Reviewing it properly

Quarterly, not weekly. A model revised every few weeks is a diary, not a description.

At each review, three questions:

• Of the prospects that converted, where did they score? If good outcomes are scattered evenly across the range, the model has no signal and needs rebuilding rather than tuning.
• Of the prospects reps overrode, what did they have in common? That common factor is either a missing weight or a missing disqualifier.
• Has anything changed about what you sell or who you sell it to? If yes, the ICP changes first and the weights follow. Never the other way round.

Where to start

If you are building from nothing:

1. Write the ICP as one paragraph a machine could apply. If it needs three, you have three ICPs.
2. Take the nine factors and the default weights above.
3. Score one real list of at least a few hundred rows.
4. Compare the top fifty against the fifty your best rep would have picked.
5. Change two or three weights. Not nine.
6. Set the threshold so the qualified list is about a fortnight of work.
7. Leave it alone for a quarter.

Most teams get to step five and want to change everything. The discipline of changing two or three things is what lets you tell, next quarter, which change did what.

If you would rather see it running than build it, Dizenly does this on your own list and shows the per-factor arithmetic on every score — and the twelve industry playbooks are the same nine factors re-weighted, with each change explained, which is a reasonable place to steal a starting point from.