Define the ICP from real customers

Your ICP should come out of a query on your customer base. A persona slide can serve as a first hypothesis, but a TAM built from it inherits every wishful line it contains, and a list of 8,000 accounts is an expensive way to discover that one line was wrong.

This chapter produces two things: a customer table with one row per account, and a one-page ICP card saying who goes in the list and who stays out. The ICP and positioning audit asks a related question for your marketing. Here the output has to be precise enough to become list filters.

Throughout this part of the guide, say you sell payroll and scheduling software to French employers with several sites. Every figure below is invented for the example.

Build one table with one row per customer account

Take every account that started paying at least 12 months ago, including the ones that left. Removing churned accounts is the mistake to avoid first, because you would end up describing whoever happens to still be around.

Column	Comes from	Used for
account_id, domain, legal_name	CRM	Joining everything, matching to outside data
siren	Invoices	Exact match to the French registry
first_paid_at, churned_at	Billing	Cohorts and retention
first_year_arr, arr_month_12	Billing	Expansion or contraction
discount_pct, failed_payments	Billing, CRM	Revenue quality
support_tickets_12m	Help desk	Cost to serve
sales_cycle_days, source	CRM	Cost to acquire
referrals_made, reference_calls, public_review	CS notes, CRM	Advocacy
sector, headcount_band, open_sites, tech_stack	Enrichment, next section	Cutting into segments

Advocacy data is rarely in one place. Ask customer success and the founders to go through the list by hand and mark who sent you a lead, took a reference call or wrote a public review. It takes an hour, and nothing in your tools will replace it.

Fill the company columns from outside

Your billing tool knows a name and maybe a domain, rarely a headcount or an activity code. Derrick, which I co-founded, so weigh this accordingly, fills those columns from the list you already have, in its web app, a Google Sheets sidebar, its API or an MCP client.

Column to fill	Feature	Input	Output	Credits
Activity, headcount, legal data	SIRET/SIREN Enrichment	SIREN, SIRET or company name, with an optional département	Name, directors, revenue, headcount, address, legal fields, and a confidence score on name searches	1 per company
LinkedIn URL	Search Companies	Company name	LinkedIn company URL	1 per company
LinkedIn industry and size	Enrich Companies	LinkedIn company URL	Data from the company page	1 per company
Stack	Website Technologies	Website URL	Detected CMS, CRM and similar	2 per site

Registry lookups cover French companies only, and a match on a name alone can land on the wrong entity, so review the low-confidence rows by hand. None of this tells you why a customer stayed. Below a hundred customers, you can do the same work without any tool: look up each SIREN in the free SIRENE data from INSEE, and read the company page and the website yourself.

Measure stays, pays and recommends separately

A single “customer health” score hides the trade-offs you need to see. An account can pay well and leave in month 14, and a small account can love you and send you three leads a year. Keep three measures.

Behaviour	Measure	Example of a good result
Stays	Still active at month 12, then month 24	Active at 12 months
Pays	Net ARR retention at 12 months, discount given, failed payments, tickets per account	Net retention at 100% or more, with a normal discount
Recommends	Referrals, reference calls, public reviews	At least one in 24 months

Once sector, open_sites and headcount_band are filled, define a few segments by combining them, then run this. Twenty accounts is a floor below which I would not read a percentage.

SELECT
  segment,
  COUNT(*)                                                AS customers,
  AVG(CASE WHEN active_at_month_12 THEN 1.0 ELSE 0 END)   AS stays,
  SUM(arr_month_12) / NULLIF(SUM(first_year_arr), 0)      AS pays,
  SUM(referrals_made) * 10.0 / COUNT(*)                   AS referrals_per_10,
  AVG(sales_cycle_days)                                   AS sales_cycle_days
FROM customer_accounts
WHERE first_paid_at < CURRENT_DATE - INTERVAL '12 months'
GROUP BY segment
HAVING COUNT(*) >= 20
ORDER BY pays DESC;

Read the output against what sales believes

Segment	Customers	Stays	Pays	Referrals per 10 accounts	Sales cycle (days)
Childcare groups, 3+ sites	38	92%	116%	4.2	35
Care home groups, 3+ sites	22	95%	109%	2.7	60
Restaurant groups, 3+ sites	64	71%	88%	0.9	20
Single-site employers	190	58%	74%	0.5	12

Sales would probably name restaurants as the best segment: plenty of logos, quick deals. The table says they shrink after signature. Childcare and care home groups do all three things, even with a slower cycle on care homes. Single-site employers bring volume and leave, so they go on the exclusion list, however easy they are to close. Feed the same cut into the retention audit if you run one.

The weights are yours to decide. If cash is tight, a short sales cycle counts for more. Write down the reasoning, because someone will challenge it when the list comes out smaller than they hoped.

Interview for what was true before they bought

The numbers name the segment. You now need to know what those companies had in common the month before they started looking, since that is what you will search for in the market. Talk to five to eight accounts from the best segment and three that churned.

Question	What you are after	What it may become
What had changed in the months before you looked for a solution?	A dated event	A signal to monitor
How many people handled this, and with what?	Team shape, existing tools	A size bound, a stack criterion
Who first raised the problem, and who signed?	Job titles	Personas
Which companies like yours would not need this?	Their own exclusion logic	A disqualifier
For churned accounts: when did you know it would not work?	The missing precondition	A disqualifier

Note anything visible from the outside. “We opened our fourth site and the spreadsheets broke” can be found in data. “We wanted to modernize” cannot.

When you have too few customers

Under 20 accounts per segment, the HAVING clause returns nothing, and lowering it gives you percentages built on noise. Widen the evidence instead: add lost deals with their reasons, live opportunities and anyone who agreed to pay before the product was finished. Then mark each line of the card with what supports it, so “seven customers” and “my hypothesis” stop looking the same. Review the card when you reach 50 customers.

Write the ICP card, with the exclusions

Line	In	Out	Evidence
Sector	Childcare and care home operators	Restaurants, for now	60 customers, retention above 100%
Structure	3 or more sites	Single site	Single-site cohort at 74% net retention
Size	20 to 499 employees	Under 20, 500 and above	41 of the 60 customers; larger accounts untested
Geography	France	Other countries	All customers
Buyer	Head of HR, finance director, founder in small groups	IT as sole contact	Interviews, 8 of 11
Trigger	New site opened, payroll hire	None	Interviews, 6 of 11

Your next step: export the customer table this week, fill the company columns and run the segment query. Then write the card, and take each of its lines to the next chapter to turn it into something a data source can answer.