How to build an ICP from win-loss data, not a persona workshop

Most ICP work produces a document nobody opens. It happens in a workshop, it contains a persona with a stock photograph and a name like Marketing Mary, and within a quarter it is a file in a drive that no rep has ever consulted before a call.

The problem is not the format. It is that the input was opinion. Sales believes one segment, marketing believes another, product believes a third, and the workshop resolves the disagreement by averaging it. What comes out is a profile nobody actually wins with.

Your own closed-won and closed-lost record already contains the answer. It just has never been read as a dataset.

The method

Pull your last twenty closed-won and your last twenty closed-lost. If you do not have twenty wins yet, use every win you have and be honest that the pattern is provisional.

For each one, record what you already know: industry, revenue band, employee count, the buyer's title, how the deal started, the cycle length, the deal size, and whether they are still a customer today. That last column matters more than most people expect. A closed-won that churned in eleven months belongs in the loss column, whatever the CRM says.

Then find the attributes where the two groups genuinely diverge.

That distinction is where most of this work goes wrong. Every one of your customers is a company with employees and a budget. Those shared traits tell you nothing. The useful attributes are the ones present in most wins and absent in most losses.

Expect to find three. Occasionally four. If you find nine, you have found noise, not a profile.

What the attributes usually turn out to be

They are rarely firmographic, which is why filters and lists disappoint people.

In practice the separating attribute is usually structural or situational. A specific role existing in the org. A particular system already in place. A trigger event in the last twelve months. A business model detail, like whether locations are corporate-owned or franchised, that changes who can actually sign.

I have seen a company discover that every win had a director-level operations hire in the previous year and no loss did. Their targeting had been built on company size, which was irrelevant. The real signal was a person having recently been given a mandate.

You cannot guess this. You can only find it by reading the two lists next to each other.

Turn it into something a rep can use

A profile that requires interpretation will not survive contact with a busy team. Convert it into a score.

Two axes. Fit is structural and slow-moving. Signal is situational and time-sensitive. You need both, and most companies track only the first.

A workable weighting looks like this. Firmographic fit, which is the crude filter, 25 percent. Structural pain, meaning does the problem you solve actually exist at their stage and model, 30 percent, because it is the strongest single predictor. Access, meaning can you reach the economic buyer, 15 percent. Technographic fit, 15 percent. Timing signal, 15 percent.

The test for the rubric is not sophistication. It is whether a rep can score an unfamiliar account in under two minutes. If it takes longer, they will not do it, and an unused rubric is worse than none because it creates the appearance of discipline.

Tier the output, then behave differently

Scoring is pointless unless it changes what people do.

Tier 1, roughly the top ten percent, gets full account research and multi-threaded, human-led pursuit. Around twenty-five to forty accounts per rep, assigned by name.

Tier 2, the next thirty percent, gets insight-level personalisation and a lighter touch.

Tier 3 exists to catch timing, not to force it.

Here is the discipline that matters, and most teams invert it. Tier 1 accounts get worked whether or not there is a signal. Tier 3 accounts get worked only when a signal fires. Most companies do the opposite, chasing whoever moved most recently, and then wonder why their best-fit accounts went untouched for three quarters.

The three mistakes

Building it from aspiration. Your ICP is not who you wish you sold to. It is the pattern in the deals you won, kept, and expanded. If the enterprise logo you want is not in the won column, it is not in the profile.

Refusing to cut the list. The first honest application of a scoring rubric usually removes a large share of the target list, and that feels like losing ground. It is not. Those accounts were never going to close; they were absorbing effort while looking like coverage.

Treating it as permanent. Re-run it every two quarters against the deals closed since. If the attributes have shifted, your market moved and you want to know that from data rather than from a bad quarter.

What changes when this is right

Three reps asked who your best-fit customer is give roughly the same answer, and can say why. Win rates become explainable instead of mysterious. Territory arguments get shorter, because the split is defensible. And marketing and sales stop debating segments, because the disagreement was never about opinion, it was about nobody having read the record.

Where to start

Pull the forty deals. Put them in one sheet. Read the two groups side by side for an hour.

You will find at least one attribute you did not expect, and that single finding is usually worth more than the workshop you were considering instead.