Why build your ICP from closed-won data instead of a workshop?
An ideal customer profile (ICP) is a description of the type of company most likely to buy from you and stay. It’s written as traits you can observe before contact: industry, size, locations, tools used and business model. It works at the account level. A buyer persona describes a person inside that account.
A persona workshop asks your team who they think the best customers are. The answers are usually honest, and usually shaped by the last three memorable deals. Your CRM holds the record of every deal instead, so the ICP can be tested rather than agreed. It’s also the base layer of market mapping: the same traits become the filters that turn a full market list into the accounts worth working.
Why it matters to a sales leader: in the Ebsta x Pavilion 2025 benchmarksSource 1 (655,000 opportunities), top performers were 24% more likely than low performers to disqualify non-ICP deals early. That’s industry data and a relative difference. A team can only disqualify against an ICP it can state.
This guide defines the target. For the backward-looking question of why deals were won and lost, start with the five-step CRM win-loss analysis; its segment step feeds straight into step 3 below.
What data do you need before you start?
- Every closed deal from the last two to three years, won and lost, with the company associated. Losses are the comparison group; without them you can’t tell a winning trait from a common one.
- Retention for each won account: still a customer 12 months after the deal closed, yes or no. Expansion revenue if you track it.
- A market sample: a list of companies in your broad target space from a data provider or your own TAM build, with the same fields. This is what lets you measure lift. (Sizing that market properly is a separate job, covered in the TAM, SAM and SOM guide.)
- Consistent fields across all three: the same industry list and the same employee bands, or the comparison breaks.
Expect missing values. Salesforce’s 2024 State of Sales surveySource 2 of 5,500 sales professionals found only 35% completely trust the accuracy of their organisation’s data (industry data). The respondents were in North America, Latin America, Asia-Pacific and Europe. Enrich the CRM companies from the same source as your market sample, so both sides have the same gaps.
The six-step method
- Define a good win. Won, and still a customer after 12 months. Flag deals that came through the founder’s personal network or needed a one-off discount; you’ll test them separately in step 4.
- List candidate traits you can observe from outside. Industry, employee band, number of sites or branches, region and tools in use (from job ads, websites or technographic data). Add ownership (founder-led, PE-backed, group subsidiary) and the value band your deals fall in. If a data source or a public website can’t tell you the trait, leave it out of the ICP.
- Compare winners with losers and with the market. For each trait, calculate two things. The first is the win rate when the trait is present. The second is lift: the trait’s share of good wins divided by its share of the market. A trait needs both: a higher win rate shows you convert it, and lift shows it’s rarer in the market than in your customer base.
- Apply the noise rules (next section). This is where most traits drop out.
- Write the survivors as tiers. Tier 1 matches all the core traits, Tier 2 most of them, Tier 3 one, and everything else is outside the ICP. HubSpot’s ABM tools include an “Ideal customer profile tier” company propertySource 3 for exactly this (Tier 1 to Tier 3, set by hand or by workflow, on Professional or Enterprise plans). The definition can then live where reps already work.
- Test the tiers on deals they haven’t seen. Build the tiers from older deals, then score the most recent six months. If Tier 1 doesn’t win clearly more often than Tier 3, the definition isn’t ready.
Fictional company, for illustration only. Quorvel Software sells rostering software to South African businesses. It has 88 won and 324 lost deals from the last three years (a 21% win rate, plausible range 18%–26%), and a market sample of 6,000 companies. Each trait below is shown with the win rate when present (95% Wilson interval) and lift against the market. For readability this table uses all 88 wins; the retention check follows under the noise rules.
| Trait | Share of wins | Win rate when present | Share of market | Lift |
|---|---|---|---|---|
| Logistics and transport | 25% | 34% (24%–47%) | 8% | 3.1× |
| 3+ operating sites | 70% | 34% (27%–41%) | 25% | 2.8× |
| Already on a cloud payroll tool | 55% | 33% (26%–41%) | 20% | 2.8× |
| 50–200 employees | 60% | 30% (24%–37%) | 30% | 2.0× |
| Head office in Gauteng | 50% | 21% (16%–27%) | 48% | 1.0× |
| Founded in the last 5 years | 10% | 22% (12%–37%) | 12% | 0.8× |
Reading it: half the wins are in Gauteng, which feels like a pattern, but half the market is too (lift 1.0×) and the win rate is the average. It’s not an ICP trait. Young companies look no different from anyone else. Sites, payroll tooling and industry all raise the win rate and are over-represented among wins.
Lift is a trait's share of good wins divided by its share of the market. Logistics (25% of wins, 8% of market, 3.1×), 3+ sites (2.8×) and cloud payroll (2.8×) are over-represented among wins. Gauteng (50% against 48%, 1.0×) isn't. Example data: Quorvel Software is fictional. About this data: Quorvel Software (fictional), for illustration only.
Which patterns are noise? Seven rules
These rules decide what survives step 4. None of them needs special software. A spreadsheet and the Wilson interval are enough. The Wilson interval is the range for a proportion that the NIST Engineering Statistics HandbookSource 4 favours over the simple “plus or minus” formula. The spreadsheet version is in the win-loss guide.
- Too few deals. If fewer than about 30 closed deals share the trait, treat it as a hypothesis. Check its range: if it overlaps your average win rate, it hasn’t been shown to matter.
- Lift near 1. A trait that’s as common in the market as among your wins tells you nothing about fit, however often it appears in the customer list.
- Won, then churned. Rerun the table on good wins only. SaaS Capital’s 2025 surveySource 5 covered more than 1,000 private B2B SaaS companies. Their median net revenue retention was 98% for companies with an average contract value under $12,000 a year and 106% for those above $250,000 (industry data; US dollars). Retention varies with the kind of customer you sell to, so a segment that signs but leaves inflates the raw win data.
- One whale or one network. If a trait’s wins come mostly from one large deal or from the founder’s contacts, remove those deals and look again.
- A rep in disguise. If one rep owns most deals with the trait, compare that rep’s win rate on deals without it. You may have found a good rep, not a good segment.
- Not visible before the first call. Pain, urgency and budget belong in discovery, where SPICED (Winning by Design’s Situation, Pain, Impact, Critical Event, Decision frameworkSource 6) captures them. The ICP is for choosing accounts; discovery is for qualifying them.
- An old era. If the trait only wins in deals from before a product, pricing or positioning change, it describes the old company. Split the data at the change.
Quorvel Software’s retail wins: 9 won from 58 closed, but 5 of the 9 churned within a year. On good wins only, retail converts 4 of 58 (7%). Retail moves out of the ICP. That matches its weak raw win rate in the win-loss analysis, but the churn is the stronger reason.
Retail won 9 of its 58 closed deals, but 5 of those 9 customers churned within a year. On good wins only, retail converts 4 of 58 (7%) and moves out of the ICP. Example data: Quorvel Software is fictional. About this data: Quorvel Software (fictional), for illustration only.
Lost deals in segments that stay in the ICP are a different matter. The closed-lost deals guide covers which of them are worth reopening.
A sharp ICP tells reps who to call; timing is a separate gap. Take the 3-minute signal gap scorecard to see whether timing is your bigger gap.
How do you know the ICP works?
Test it the way you’d test a forecast: on data it wasn’t built from. Rebuild the tiers using only older deals, then score the most recent six months and compare win rates by tier. That’s the idea behind testing a score on your own sales history.
Quorvel Software’s tiers. Tier 1: 3+ sites, a cloud payroll tool and 50–200 employees. Tier 2: two of the three. Tier 3 and outside: one or none. Logistics has the highest lift, but only a quarter of wins are in logistics, so making it a rule would push three-quarters of good customers out of Tier 1. It orders the Tier 1 list instead. The tiers were built from deals closed before April 2026; a trait table rebuilt on those deals alone kept the same three core traits. They were then tested on the 70 deals closed from April to September 2026:
| Tier | Closed deals | Won | Win rate (range) |
|---|---|---|---|
| Tier 1 | 18 | 8 | 44% (25%–66%) |
| Tier 2 | 22 | 5 | 23% (10%–43%) |
| Tier 3 and outside | 30 | 2 | 7% (2%–21%) |
The order holds and Tier 1 clearly beats Tier 3. Tier 1 and Tier 2 overlap, so Quorvel keeps both tiers and retests next quarter rather than dropping Tier 2.
On 70 deals closed from April to September 2026, Tier 1 won 44% (25%–66%) against 7% (2%–21%) for Tier 3 and outside, with no overlap. Tier 2 (23%, 10%–43%) overlaps Tier 1, so both tiers stay and get retested next quarter. Example data: Quorvel Software is fictional. About this data: Quorvel Software (fictional), for illustration only.
Two checks after the numbers. Ask two reps to tier ten accounts each from the written definition, and see whether they agree; if they don’t, the wording is too loose. Then count how many Tier 1 accounts exist in your market sample. If the answer is 40, the ICP is correct but your headroom question has just become urgent. When reps do start on Tier 1, a sales-ready lead brief gives them the account context in one page.
Fit is not timing
A perfect ICP account still isn’t buying most of the time. Professor John Dawes of the Ehrenberg-Bass Institute estimated in 2021Source 7 that up to 95% of business buyers are out of the market at any one time. Where companies switch provider about every five years, roughly 5% are in-market in a quarter. The ICP narrows who is worth working. Buying signals (hiring, funding, a new leader, a tool change) show which of those accounts is worth calling this month. The guide to ten buying signals SaaS sales teams should track lists the ones to start with, and how the signal layer works shows where they fit.
That is also why the ICP should stay short. Four or five traits that clearly raise the win rate make a better filter than twelve that each add a little and make every account look like a fit.
When this approach won’t work well
- Fewer than about 50 closed deals. The ranges will be too wide. Use the traits of your best ten customers as a working hypothesis, and label it as one.
- A new product or market. Past wins describe what you used to sell. Use them for the parts of the offer that haven’t changed.
- Most wins came through the founder. You’re measuring a network, not a market.
- Thin company data. Check what share of your South African accounts have employee counts, industry and tool data filled in before you rely on a trait. Coverage in global data providers varies by country, so test a sample first.