Use this when the question is "who should we sell to?". It produces a one-page ICP that a rep, a marketer and a list builder can all act on the same day.
Inputs
Ask for as much as exists, then work with what you get:
- Closed-won deals for the last 12 to 24 months: company, size, industry, region, deal value, cycle length, who signed, why they bought.
- Closed-lost and churned accounts with the reason.
- The product's core job and price range.
- If there are fewer than 10 customers: the founder's notes on the best five conversations, plus the best five failures.
The play
1. Rank customers by value, not by count
Sort won deals by a blend of revenue, cycle speed and retention. The top 20% is the real ICP; the long tail is noise. A segment with many small, slow, churn-prone deals looks big on a pie chart and loses money.
2. Find the traits that separate winners from losers
For each trait, compare the share among top customers with the share among lost or churned ones. Traits worth keeping show a gap of at least 20 points. Look at four layers:
| Layer | Examples |
|---|
| Firmographic | Industry, headcount band, revenue band, region, business model |
| Technographic | Tools they already run, data maturity, stack that integrates with yours |
| Situational | Hiring for a role, new leader in seat, funding, expansion, a regulation deadline |
| Behavioural | How they buy: committee size, procurement, self-serve trial or demo |
Situational traits usually predict timing better than firmographics predict fit. Read references/icp-scoring.md for the comparison table and the scoring weights.
3. Write the pain in the buyer's words
Pull phrases from call notes, emails and reviews. "We lose two days a month reconciling payouts" beats "operational inefficiency".
4. Set disqualifiers
Hard noes save more time than must-haves. Typical ones: below a headcount that can afford the price, a region you cannot serve, a tech stack you cannot integrate with, a business model where the pain does not exist.
5. Translate it into filters
Every trait becomes a filter a list builder can run: industry codes, headcount ranges, job titles, keywords in job listings, technologies detected. Flag traits that cannot be filtered and must be checked by research.
Output format
ICP: <one-line name, e.g. "Multi-location restaurant groups, 5-50 sites">
Why they buy: <pain in their words> | Trigger: <what makes now the moment>
Must-haves: <3-5 traits>
Nice-to-haves: <2-4 traits that raise priority>
Disqualifiers: <3-5 hard noes>
Buyer: <title that signs> Champion: <title that feels the pain> Users: <who uses it>
Filters: <list-builder ready criteria>
Confidence: <high / medium / low and why: sample size, data gaps>
What good looks like
- Narrow enough that a rep can name 50 target accounts from memory after reading it. "B2B SaaS companies" is not an ICP.
- It includes a trigger. Fit without timing produces polite "not right now" replies.
- Disqualifiers are as specific as the must-haves.
- The confidence note is honest: an ICP built from six customers says so and lists what to test next.
- The common mistake is describing the customers you want instead of the customers who bought, renewed and expanded.
Rules
- MUST base every trait on deal data or explicit notes, and say which.
- MUST include at least three disqualifiers.
- NEVER mix the economic buyer and the end user into one persona.
- NEVER present a trait with fewer than five supporting accounts as proven; mark it as a hypothesis.
Example prompts
- "Here are our last 40 closed-won and 25 closed-lost deals. Build our ICP."
- "We have 7 customers. Help me write a first ICP and tell me what to test."
- "Our reply rate dropped to 1%. Check whether our ICP is too broad."