Turn the ICP into observable criteria

A criterion no field can verify will never reach your list. “Innovative”, “growth-minded” and “takes compliance seriously” are fine in a pitch and useless in a query, so each line of the ICP card needs a field, a source and a rule that returns an answer.

The deliverable is a criteria sheet with one row per criterion. We keep the example from the previous chapter: payroll and scheduling software for French employers with several sites, and invented figures.

Put every ICP line through the script test

Ask whether a script could answer yes, no or unknown from a field available for most accounts. If it can’t, find a proxy. If no proxy holds, the line leaves the list logic and becomes a question for the first sales call.

ICP line	Observable proxy	Field	Source	Verdict
Childcare or care home operator	Declared main activity	NAF code 88.91A or 87.10A	SIRENE	Filter
Several sites	Open establishments per SIREN, 3 or more	Count of active SIRETs	SIRENE establishment file	Filter
20 to 499 employees	Headcount band	trancheEffectifsUniteLegale	SIRENE	Filter
Payroll handled in house	A payroll job ad, recent or open	Job titles	Job boards, LinkedIn	Score
Expanding	New establishments in the last 12 months	Establishment creation date	SIRENE	Score
Fed up with their current tool	None you can trust			Sales question

That last row is common. Pain and dissatisfaction rarely leave a data trace. Keep them for the conversation, where they belong.

Know the four families of criteria

Family	Examples	Typical source	Main weakness
Firmographic	Activity, headcount, age, legal form, location	Registries, LinkedIn	Self-declared or dated
Structural	Number of sites, group membership	Registries, maps	Groups often split across several legal entities
Technographic	CMS, CRM, payment tool, analytics	Website detection	Sees only what a public page loads
Signals	Hiring, funding, new site, new executive	Job boards, news, registries	Expires quickly, covered in the signals chapter

Firmographic and structural criteria decide who is in the TAM. Signals decide who to contact this month, so they belong in scoring and should never shrink the market itself.

Use registry fields with care

In France, the SIRENE files give you the first criteria almost for free. The NAF nomenclature on the INSEE site lists the codes, and INSEE revises it, so check which version your source uses. Headcount comes as a band code:

Code	Employees	Code	Employees
00	0	21	50 to 99
01	1 to 2	22	100 to 199
02	3 to 5	31	200 to 249
03	6 to 9	32	250 to 499
11	10 to 19	41	500 to 999
12	20 to 49	NN or empty	Not provided

Our filter translates to this:

etatAdministratifUniteLegale = 'A'
AND activitePrincipaleUniteLegale IN ('88.91A', '87.10A')
AND trancheEffectifsUniteLegale IN ('12', '21', '22', '31', '32')

Three traps. A NAF code is the main declared activity, so a group can sit under a holding code such as 70.10Z or 64.20Z while its operating subsidiaries carry the real one. Software companies scatter between codes like 62.01Z and 58.29C with little logic. And the headcount band refers to a given year, shown in anneeEffectifsUniteLegale, which can be several years old. Before fixing your NAF list, count the codes held by your best customers:

SELECT naf_code, COUNT(*) AS best_customers
FROM customer_accounts
WHERE segment IN ('childcare_groups', 'care_home_groups')
GROUP BY naf_code
ORDER BY best_customers DESC;

Codes you hadn’t thought of will appear, and the ones you assumed may turn out to be rare.

Give each criterion a role and three possible answers

A hard filter removes an account, so it needs a field that is nearly always filled. Sparse fields go in the score, where a blank costs points instead of a place in the market.

Role	Requirement	Effect	Example
Hard filter	Field present for nearly all accounts	Account in or out	Activity, active status
Score	Field may be incomplete	Adds points	Stack, hiring
Sales question	Not observable	Asked during discovery	Satisfaction with current tool

Treat a missing value as its own state. In a spreadsheet, each criterion column returns one of three words, and a final column combines them.

=IF(F2="","unknown",IF(F2>=3,"pass","fail"))

=IF(COUNTIF(G2:I2,"fail")>0,"out",IF(COUNTIF(G2:I2,"unknown")>0,"to check","in"))

The “to check” group deserves a sample review. If most of it turns out to be good accounts, loosen the rule.

Back-test on customers you know

Run the rules on your best-fit customers and on accounts you would not want again. A useful criterion lets most of the first group through and stops much of the second.

Criterion	Best customers passing (of 60)	Poor-fit accounts passing (of 40)	Decision
NAF in list	54	6	Hard filter, add the missing codes
3+ open sites	52	9	Hard filter
Headcount 20 to 499	41	14	Keep, unknown passes to review
Hiring in the last 90 days	27	18	Light score
Blog on the website	22	15	Drop

The NAF row needs a caveat, because the segments were partly built with that field, so it flatters itself. The six best customers it misses are the ones to read: they show you holding structures and odd codes. Any hard filter losing more than a few of your best accounts is too strict for a market you have not seen yet.

Collect the harder criteria

Technology and hiring take more than a registry. Derrick, which I co-founded, so keep that in mind, covers these from a column of companies.

Criterion	Feature	Input	Output	Credits
Stack of a known company	Website Technologies	Website URL	Detected technologies	2 per site
Companies on a given stack	Find Companies by Tech Stack	Technologies or site type, optional country	Domain, site type, score, country, technologies	1 per company found
Hiring	Company Hiring Signal	LinkedIn company URL	Yes or no, number of open roles, job titles	1 per company
Anything stated on the site	Website Crawler then Ask Claude	URL, then a closed question on the text	Page content, then your classification	2 per row, each

Detection only sees what a site exposes, so a payroll tool used behind a login will never show. That rules out most back-office software as a technographic criterion, and our payroll vendor can’t read its competitors’ installed base this way. Any AI classification should be a closed question (“does this site mention several locations: yes, no, unclear”), and you should read a sample of 50 answers before trusting the column.

Without a tool, browser extensions that detect technologies, a look at the page source and a manual pass over job boards work fine at the scale of a back-test.

What the criteria sheet looks like

Criterion	Role	Field	Source	Rule	If unknown	Back-test
Activity	Filter	activitePrincipaleUniteLegale	SIRENE	In code list	Review	54 of 60
Sites	Filter	Open SIRET count	SIRENE	3 or more	Review	52 of 60
Headcount	Filter	trancheEffectifsUniteLegale	SIRENE	Bands 12 to 32	Review	41 of 60
Hiring	Score	Open roles	LinkedIn	Payroll role open	No points	27 of 60

Your next step: take every line of your ICP card, write its row in this sheet, and run the back-test on your own customers. The source column then tells you what you need to buy or extract, which is what the data sources chapter settles.