ICP fit score: how to build one, and how to test an AI fit score

This guide is for RevOps managers, sales leaders and founders who want every account in the CRM or a prospect list rated for fit, and are deciding whether to build the score themselves or let an AI tool do it. It covers what a fit score is, how to build one, the three ways AI tools score fit, and a test you can run on any AI fit score before you route accounts by it. If you haven't written your ICP yet, start with the ICP template. For scores that predict conversion from behavior, see AI lead scoring.

What an ICP fit score is, and isn't

Vendors implement fit differently. On their own pages, checked 30 October 2026: HubSpot's default "Ideal Customer Profile Tier" property is "the categorization of how well a company matches your ideal customer profile, with Tier 1 being the closest". HubSpot's prospecting agent scores accounts against your ICP and shows a "Reason to Reach Out" for each recommendation. Lusha's ICP fit score runs from 0 to 1 against the account and persona criteria in your target profile, costs one credit per record even when the score is 0, and is combined with a signal score into one lead score. PostHog publishes its in-house fit score, which it describes as "definitional, not a revenue prediction".

That last point matters. A fit score says whether an account matches your definition. Whether it will buy this quarter is a different score, and blending the two too early hides a perfect-fit account that simply has no trigger yet.

How to build an ICP fit score

1. Start with disqualifiers. Accounts you'd never sell to (wrong industry, too small, a competitor, an existing customer) get a status of disqualified with a reason, not a low number.
2. Pick four to six criteria from your best and worst customers, each one something you can check from outside the company.
3. Weight them to 100, heaviest where your won deals differ most from your lost ones. The lead scoring template shows how to turn lift into points.
4. Add an "insufficient data" status. PostHog's published method gives no number to profiles with no usable data and warns: "Never read a missing score as 0." In its production data (18,396 organizations, measured 1 September 2026), about 35% were insufficient data. Scoring them as zero would have buried a third of its accounts.
5. Set tiers by capacity. PostHog calls the threshold "a downstream-capacity decision, not an accuracy one": tier A should be roughly the number of accounts your team can work properly.
6. Version it. Stamp each score with the version and date, so you can tell a changed account from a changed formula.

Version: [v1.0, date]   Owner: [name]   Scale: 0 to 100

DISQUALIFIED (no number, reason recorded):
- [e.g. fewer than 10 employees]
- [e.g. not in our industries]
- [existing customer, open opportunity, competitor]

INSUFFICIENT DATA (no number, retry in [30] days):
- [e.g. size and services both unknown]

WEIGHTED CRITERIA (total 100):
- [criterion 1]: [points]  Source: [where we check it]
- [criterion 2]: [points]  Source:
- [criterion 3]: [points]  Source:
- [criterion 4]: [points]  Source:

TIERS: A [70+]  B [45 to 69]  C [20 to 44]  D [below 20]
Tier A target size: about [number] accounts, our capacity for [quarter]

Three ways AI tools score fit

The third is the newest and the one most AI prospecting tools use, because it can judge criteria no database field holds, such as "offers outsourced accounting" or "sells to hospitals". It's also the one that most needs testing. Bebop also starts from a written description: you describe the companies you want in plain language, and it explains why each one it finds fits, so the same tests apply.

Worked example: building and testing a fit score

Illustrative example. The company, deals and test results are invented; the establishment counts are from the Census Bureau.

The seller: Ledgerline (ledgerline.example) sells workflow software to accounting firms. Its fit score v1.0: firms with 20 or more employees get 30 points (10 to 19 get 15); advisory or outsourced accounting services, 25; two or more offices, 15; a desktop practice-management tool named in job posts, 20; hiring accountants in the last six months, 10. Under 10 employees is disqualified; unknown size and services is insufficient data.

How big can each tier be? County Business Patterns counted 55,052 US establishments in NAICS 541211 (offices of certified public accountants) in 2023. The size gate alone removes the 45,984 with fewer than 10 employees, leaving 9,068: 5,089 with 10 to 19, 2,386 with 20 to 49, 807 with 50 to 99 and 786 with 100 or more. Establishments aren't companies (a multi-office firm counts once per office), so the number of firms is lower. Ledgerline's two reps can work about 400 accounts a quarter, so tier A is set at 70 points to land near that.

Backtest on past deals. Ledgerline scored its 178 closed opportunities from the last 18 months as each account looked when the deal opened, using only facts available then.

Win rates fall tier by tier, so the score ranks the way it should. The insufficient-data row wins at about the average, which is exactly why it mustn't be scored as zero: those accounts need research, not a D.

Testing an AI fit score. Ledgerline then tried an AI tool that reads each firm's website against the written ICP and returns fit or not fit with a reason. The sales manager labeled 100 open accounts blind, and the two sets of labels were compared.

They agree on 84 of 100. That sounds good until you notice a scorer that called every account "not fit" would have agreed on 71. Cohen's kappa corrects for this: expected chance agreement is 0.29 × 0.31 + 0.71 × 0.69 = 0.58, so kappa = (0.84 − 0.58) ÷ (1 − 0.58) = 0.62. That's well above chance, but 9 of the AI's 31 "fit" calls were accounts the manager rejected.

Two more checks found why:

• Consistency. Rerun a week later on 50 of the same accounts, the AI changed its answer on 6 (12%). Five were firms of 10 to 19 people, where it estimated headcount differently each time. Fix: pass in the headcount from Ledgerline's own data rather than letting the model estimate it.
• Evidence. Of the 31 "fit" reasons, 27 checked out against the firm's own site. Four cited advisory services the firm didn't offer, taken from articles about the profession. The account research accuracy check describes how to catch these.

The decision: Ledgerline uses the AI only for the two criteria that need reading (services and practice-management tool), feeds its answers into the points score, and has the manager review every tier A account. It reruns the backtest each quarter and records the change as v1.1.

A test you can run on any AI fit score

1. Label 100 accounts yourself first, blind to the AI: fit or not fit, with a one-line reason.
2. Include 10 accounts you lost for fit reasons. The AI should reject at least 8.
3. Compare the labels in a two-by-two table and work out agreement and kappa as above. Aim for a kappa of about 0.6 or more before you route by the score.
4. Rerun the same 50 accounts a week later. More than about 1 in 10 changing answer means a criterion is too vague or depends on an estimate.
5. Check the reasons on every account the AI calls fit and you don't. Each reason should cite a page about this company.
6. Backtest the tiers on closed deals, as in the worked example.

What to ask a vendor

• Is the score rules on data fields, a model trained on customers, or a language model reading each account?
• Which sources does it read for each account, and how fresh are they?
• Does each score come with a reason and a source we can open?
• How does it handle missing data: a low score, or a separate status?
• If we score the same account twice, do we get the same answer?
• What does it cost per record, and are records that score zero charged?

Common mistakes

• Blending fit and intent too early. A strong signal at a poor-fit company still isn't worth a sales cycle.
• Treating unknown as bad. Missing data needs research, not a D.
• Scoring with hindsight. Backtest on what you knew when the deal opened, not what you learned during it.
• Tiers sized by the formula, not the team. If tier A holds 3,000 accounts and you can work 400, it isn't a tier.

Next step: turn tier A into a list with how to build a prospect list.