The short version
Most ICP documents get written once, printed into a slide, and never touched again, which is strange, because the market that validated them keeps moving. I treat an ICP as a living hypothesis: a best current guess at who buys fastest, pays most and churns least, re-tested every quarter against actual closed-won and closed-lost data.
The test is simple: pull your last 20 closed-won deals and your last 20 closed-lost qualified opportunities, then compare firmographics, buying triggers and objections side by side. If your ICP document doesn’t explain the pattern you see, the document is wrong, not the market. Update it, don’t defend it.
Running the won-versus-lost comparison
Pull 20 closed-won and 20 closed-lost qualified deals into one sheet. Columns: industry, size band, geography, business model, the trigger that started the conversation, number of stakeholders, cycle length, deal value, primary objection, and source. Add one column for the first-year outcome of won deals: renewed, expanded, churned or complained.
Count, rather than eyeball. For each attribute, calculate the win rate by value, such as win rate for firms under 200 staff versus over. Anything with a gap of 15 points or more is a candidate criterion. With only 40 rows, treat such results as leads to test, not conclusions.
Converting adjectives into a score
Replace phrases like “innovative mid-sized lenders” with checkable fields. Example scoring: company type match 0 to 3, business model 0 to 2, product fit 0 to 3, geography 0 to 1, seniority of the contact 0 to 2, buying signal 0 to 3. Assign weights according to the win-rate gaps you found. Anything above a set total becomes priority.
Write one sentence per criterion explaining why it matters, and include a data source for it, such as Apollo, a regulator list or a website. A criterion you cannot populate for at least 80% of accounts is useless in practice and should be dropped or replaced.
Back-testing and versioning
Apply the new score to last quarter’s opportunities, blind to outcomes, then check whether the high scorers won more often. If priority accounts do not win at least 1.3 times as often as the rest, the criteria are noisy. Revise before you roll the score into prospecting lists.
Version the ICP like a product: v1.3, date, what changed, why, and what evidence. A short change log prevents the common situation where different reps work from different copies and cannot explain why the targeting moved.
Listening for drift
Between quarterly reviews, watch for early signs: a new type of buyer appearing in inbound, a segment that stops replying, a competitor winning a niche, or churn clustering in one group. Add a one-line field in CRM for “why we lost” with fixed options, and review the counts.
Also run a small deliberate test of adjacent segments: ten outbound conversations into a segment that your ICP excludes. It is the only way to discover that a rule is unnecessarily narrow. Cap the experiment so it cannot consume more than 10% of effort.
Worked example (illustrative)
A hypothetical vendor re-tests its ICP after reviewing 20 won and 20 lost deals, finding that business model and a buying trigger separate outcomes better than company size.
What changed: A smaller, differently defined list produced faster and better outcomes. Replacing company size with trigger and model criteria explained the pattern in the closed deals, and back-testing confirmed it before the list was rebuilt.
Do it: the steps
Tick them off as you go. Your progress stays on this device.
Traps to avoid
- ×
Defining the ICP by who you would like to sell to. Define it by who actually buys, quickly, and stays.
- ×
Building the ICP from the biggest deals only. Large, slow deals may carry high churn or heavy customisation; include retention and cycle length in the analysis.
- ×
Skipping the lost deals. Winners alone cannot tell you what makes a company a poor fit.
What to track
- Priority-tier win rate
- won divided by closed in the top tier; compare against the rest, looking for a ratio of 1.3 times or more.
- ICP coverage
- share of pipeline created from ICP-qualified accounts per quarter.
- Retention by segment
- twelve-month retention for each ICP tier to confirm the score predicts durability, not just wins.
Benchmarks are typical planning ranges, not guarantees. Calibrate to your own data.
Questions people ask
How many criteria should an ICP score use?
Three to five that demonstrably separate winners from losers, plus a couple of simple data-availability checks. More criteria feel thorough but add noise and slow list building. If two criteria always move together, keep the one that is easier to populate from your data sources.
What if my best customers do not match the ICP I wrote?
Then the document is wrong. Describe what the best customers actually share, test whether lookalikes convert, and replace the old definition. Keep the rewrite narrow; do not stretch the ICP to include everything. Document the change so the team understands that targeting follows evidence.