Deriving Your ICP from Closed-Won Deals: A Closed-Won Analysis

Most ideal customer profiles are created in a workshop. A whiteboard, three people from sales and marketing, and at the end there’s a slide with industry, headcount, and revenue band. The slide is plausible. It’s also rarely verified.

Yet the answer is already sitting in the CRM. Every closed-won deal is a data point about who actually buys. Closed-won analysis turns that into a profile.

Why the ICP comes from closed deals, not from the workshop

A workshop ICP describes who you’d like to have as a customer. A derived ICP describes who has already bought from you. The difference only becomes apparent once the two diverge, and that’s the normal case.

Typical deviations a derivation reveals:

• The industry named first in the workshop often doesn’t account for the most closed deals.
• The company size in the profile is frequently set too large, because big logos sound more attractive.
• Triggers that actually led to a purchase don’t show up in the workshop profile at all, because nobody captured them systematically.

As long as the profile goes unverified, the error propagates downstream: into the target list, into the sequence, into the week’s prioritization. A wrong ICP isn’t an imprecise document — it’s a multiplier.

The method in four steps

1. Export closed-won deals. Pull all deals from a sensible time period out of the CRM, with deal size and account reference. Export lost deals too — you’ll need them in step four.

2. Enrich every account. For each closed-won account, pull in company data: industry, headcount, location, tech stack, open job postings, funding stage. What matters is that enrichment runs by the same rules for all accounts, otherwise you’ll be comparing apples to oranges later.

3. Find shared attributes and weight them. Now ask: which attributes do the closed-won accounts share? An attribute shared by a large share of the closed deals is a candidate for the ICP. This frequency gets weighted by deal size — more on that below.

4. Check against lost deals. An attribute is only discriminating if it appears more often among won deals than among lost ones. Skip this step, and attributes end up in the profile that simply describe your market, not your buyers.

The result isn’t a slide deck, but a file with attributes, thresholds, and a note on where each number came from. This file is the input for sourcing, scoring, and prioritization. How it turns into a work order is covered in Lead Scoring: Prioritizing ICP Fit with AI.

Weighting by deal size

The step that gets skipped most often. Unweighted, the smallest deal counts exactly as much as the biggest deal of the year. A profile built that way optimizes for easy prey.

An illustration of the mechanic, using made-up placeholder numbers: if twenty small deals come from one industry and five large ones from another, the first industry wins unweighted 4:1. Weighted by revenue, that ratio can flip. Both views are correct — they just answer different questions: one “who buys most often,” the other “who carries the revenue.”

For the ICP in outbound, the weighted view is usually the right one, because the team’s capacity is limited. If you can only properly work two hundred accounts a month, you should work the two hundred with the highest expected value.

Validating before the profile goes into the target list

Two checks that go fast and prevent a lot:

Discriminating power. Does the attribute show up noticeably more often among won deals than lost ones? If not, it describes your market, not your buyer.

Causation or coincidence. Correlation is enough for a target list, but not for an explanation. If an attribute shows up that nobody on the team can justify, mark it as open instead of adopting it. It could be a genuine finding or an artifact of the last campaign that happened to target that industry.

The second point is why this analysis doesn’t replace the workshop. It supplies the factual basis; the team supplies the interpretation. What it does replace is having that discussion without numbers.

What reproducibility depends on

A derivation that runs manually once is a snapshot. After two quarters it’s outdated, and because the effort is known, it doesn’t get repeated.

It becomes reproducible when three things live in one place: the CRM export, the enrichment, and the analysis. If they sit in three separate tools, every re-run is a manual pass with exports in between. Why fragmented stacks lose exactly this knowledge is covered in Fragmented Sales Stacks Lose Your Most Valuable Knowledge.

In a GTM system with a shared data model, the derivation is a repeatable process on the same data: deals come from the CRM, enrichment runs by fixed rules, the profile gets versioned. That makes the ICP no longer a document that ages, but a state that gets updated with every closed deal. This distinction between a tool and a system is explored further in Agentic GTM: System, Not Outbound Tool.

Conclusion

The ICP doesn’t have to be guessed. It’s sitting in the deals you already have: export closed deals, enrich accounts using the same rules, weight shared attributes by deal size, check them against lost deals for discriminating power. What comes out of that is less comfortable than the workshop slide and considerably more usable, because every line rests on a customer who actually paid.

The fundamentals of a profile’s structure are in ICP Definition in B2B. If you want to try this derivation on your own closed deals: try it free for 4 weeks, no credit card required.