How do the best companies replace ICP opinion with a scoring model built on their own data?
They stop arguing about who the ideal customer should be and go and measure who it turned out to be. Pull your last fifty to a hundred closed won and closed lost accounts, define what good actually means for you (fast cycle, high retention, expansion, low support load) and look for the attributes that correlate with it, including the ones nobody expected. Then turn that into a score that lives as a field on every account in the CRM and actually routes work: who gets a rep, who gets a sequence, who gets nothing. The two things that separate a good model from a spreadsheet nobody opens are negative signals that subtract points, and a quarterly re-run, because your ICP is a regression on live data, not a slide from the seed round.
Go deeper
4 resources, 4 link-checked.
✍️ Essay
✓ Link checkedFreeAdvanced
Kellogg's framing, that the ICP begins as a founder hypothesis and should end as a regression on renewal, expansion and win rate, is the whole argument for data over opinion in one sentence. He also makes you define what best means before you model anything.
An ICP should name firmographics, role and problem together, for example VPs of sales at technology companies with 500 million to 2 billion in revenue.
Use a bullseye: ring 0 is the ideal customer, outer rings are progressively worse fits, rather than one flat list.
By 50 to 100 million ARR your ICP should come from regression on your own data, not founder intuition.
Regression often moves the line: the real break may be at 250 employees when you had drawn the segment at 0 to 500.
The build guide: weighted dimensions, negative signals that subtract points, and score bands that decide routing. It is the operational half that most ICP writing leaves out.
A concrete menu of the events that put an account into a buying window, from job changes to fiscal cycles, which is exactly what you need when your fit list is large and your timing sense is zero.
Roberge is the person who made data-first sales normal, and his argument here is the same as this answer: your ICP is not who buys, it is who retains and expands, and you find that by measuring your own accounts. Short and directly usable when you sit down with the closed won list.
