Account scoring for ABM: set weights from won and lost deals

AI answers agree on criteria and disagree on weights

The engines agree that fit, intent and engagement are the criteria, and disagree on how much each should count. ChatGPT’s answer gives a 6-part table: ICP fit 35%, technographic fit 15%, intent 20%, engagement 15%, buying committee 10% and strategic value 5%. Perplexity suggests fit 40–50%, intent 20–35% and engagement 20–30%. Gemini puts fit at about 40% and intent at about 35%.

We opened the 21 pages that ChatGPT, Gemini, Perplexity and Google’s AI Overview and AI Mode cite for this question. Of the 19 that loaded, 7 recommend percentage weights. Add the engines’ own answers and there are at least 11 different splits. Pages that give fit a single weight put it at 30–60% of the score.^↗↗

None of the 7 pages shows the closed deals behind its numbers. A vendor guide cited by Gemini says its 40/35/25 split “works for most B2B SaaS companies”.^↗ It adds, fairly, that “your first scoring model is a hypothesis”. Another says its weighting holds up across mid-market and enterprise, based on roughly 40 ABM programmes it audited, without showing them.^↗ The most candid page, cited by Google’s AI Overview, calls its own 6-factor split “illustrative, not universal benchmarks”.^↗

The citations don’t always hold, either. ChatGPT links a Pedowitz Group page under its 6-part table. The only split on that page is fit 60% against intent and engagement 40%.^↗

ChatGPT does say to tune the weights against your own won and lost deals every quarter. No answer says how, or how many deals that takes. Where cited pages put a number on it, they range from 25 accounts with a won deal and 25 without^↗ to your last 200 won and 200 lost deals.^↗

Borrowed weights are guesses; fit comes from your own deals

We think a borrowed split is a guess, however precise its percentages look. Fit weights should come from your own closed deals. A template describes someone else’s customers; your won and lost deals describe yours.

Our reasons are practical. A written ICP tends to describe who the team likes to sell to, while won and lost deals show who buys. We also think a strict fit score makes the rest of the model better. Rep time and signal spend are limited, and every marginal account on the list thins both. A shorter list also leaves the score less noise to average over.

This is the fit analysis our Panel Check (GTM audit · 2–3 weeks) runs in check T2:

Each bar does a different job. The 1.5× bar ignores segments that win only slightly more often than average, a gap a small sample can produce by chance. The 10-deal bar stops a single good quarter from rewriting your ICP. Even 10 deals is a small sample, so we treat a segment that passes as a working answer and check it again every quarter.

Here is the arithmetic on an illustrative dataset of 80 closed deals, 20 of them won: a 25% average win rate.

Logistics earns fit points. Healthcare wins more often still, but on only 6 deals. Each deal moves its win rate by about 17 points, so it waits for more data. Manufacturing wins at about the average rate and earns nothing, however much the team enjoys those deals.

At €25k+, engagement means several people in the buying group

At €25k+, engagement should count only when several people in the buying group are active, measured across the whole account. The group is large: the average B2B buying decision involves 13 people.^↗ Gong’s analysis of 1.8M deals found that multi-threading lifted win rates by 130% on deals over $50k.^↗

Newer data points the same way. Factors.ai, which sells ABM software, analysed more than 50,000 closed deals at more than 100 B2B companies.^↗ Accounts with 6 or more contacts engaged before the deal was created had a pipeline win rate 17.1 percentage points higher than accounts with 1.^↗ It’s vendor data, and the release doesn’t define “engaged”, but the direction matches Gong’s.

We think a single engaged contact doesn’t make a qualified account while the rest of the buying group is unknown. A single person’s clicks describe that person. Several people in different roles reading, attending and replying describe a buying process.

In practice, roll contact activity up to the account. Count the distinct people active in the last 90 days and check them against the roles your won deals needed. Cap what each person can add, so a single keen reader can’t carry the score. Our audit tracks the share of target accounts with 3 or more engaged contacts in the last 90 days (check E2). For how big the group gets at your deal size, see how many people a B2B buying group should cover.

Intent moves an account up the queue without qualifying it

Intent tells you which fit accounts to work first; it doesn’t make an account qualified. Most of a market isn’t buying at any moment: the LinkedIn B2B Institute’s 95-5 rule puts it at 95% of potential buyers.^↗ Third-party intent is also relative. Bombora treats a topic score of 60 or more as a statistically significant rise in content consumption against the account’s own baseline.^↗ So a surge says an account is reading more than usual. It says nothing about fit or about who is involved. Forrester warns against letting intent signals replace the qualification process.^↗

Tools now watch more signals than a team can read. HubSpot’s updated Prospecting Agent, announced in September 2026, monitors 40+ buying signals and assembles a buying group.^↗ The more signals a tool watches, the more it matters which ones your own won deals say to trust.

We think intent earns its cost only when it changes what a rep does that day. A fit account with a surge moves to the top of its owner’s list, with the reason attached. An account outside your fit segments doesn’t join the list on a surge alone. Weight each signal by how often it came before won deals, and let it decay so old research stops counting. The signals worth tracking are in which buying signals show a B2B account is in market.

How do you test a score on last year’s deals?

Score the accounts behind last year’s closed deals with only the data you had before each deal opened. Then check that high scores won more often.

1. Take every deal closed in the last 12 months, won and lost, with its account.
2. Score each account as it stood on the day its deal was created. Engagement logged during the deal leaks the result into the test.
3. Sort the accounts into 3 bands by score, or into your tiers.
4. Compare win rates by band. The top band should win clearly more often than the average, and the bottom band clearly less.
5. Read the misses: won deals from low scores and lost deals from high scores, with their lost reasons.
6. Adjust the weights, run the test again and repeat it every quarter as new deals close.

If the top band wins at about the average rate, the score isn’t separating good accounts from the rest. Fix it before anyone routes work on it. The same test works on a tool’s default model, and it’s a quick way to see whether that ranking fits your market.

Clean the fields a score reads before switching it on

Clean the fields a score reads before it ranks a single account. We think a score switched on over dirty data multiplies the damage the bad data already does. In Validity’s 2025 survey, 76% of CRM users said less than half of their CRM data was accurate and complete.^↗ A fit model that reads a blank industry field sends those accounts to the bottom of the list, whether they fit or not.

The whole CRM doesn’t need to be clean first. The fields this score reads do:

Our audit checks how complete the key account and contact fields are on records created in the last 90 days (check C2), and how often lost deals carry a usable reason (check Q4). If you plan to put AI recommendations on the same data, read whether your CRM data is clean enough for AI recommendations.