A sales ICP is written from the customers you already have

ICP stands for ideal customer profile, and in sales it's the written description of the accounts most likely to buy from you and stay, specific enough that a rep can look at one company and say yes or no. That yes-or-no test is the ICP meaning in sales a rep actually uses.

The ICP describes the company. A persona describes the person inside it who signs the order or uses the product. A persona can also drift into an invented buyer. One test for whether you're describing a real one: "if the buyer's #1 objection isn't obvious in one sentence, you're still selling to a persona, not a customer."

Here's what four narrowed profiles looked like. Each row is one case, and each result is what that person reported.

Read down the "After" column. Every profile picked up something a rep can check against a real account: a headcount band, or a dated event like a raise or a hire.

So how narrow is narrow enough? One co-founder, whose startup got its first 100 customers from outbound, set out the rule they'd follow if they started again: "past 50 people on a list, I treat it as a bad list." Past that point they'd add another filter rather than rewrite the copy. The same plan explains why a dated event beats a title. "A job title is permanent, so it tells you nothing about this week."

A consultant auditing one company's outbound found it spending $50K a month on enterprise accounts, with zero pipeline to show for it. So the consultant asked for a list first: the last 20 closed deals, straight out of the CRM. Not one of those 20 closed deals was an enterprise account; every one was mid-market. The client, in the consultant's words, "had been chasing an aspirational ICP from executive slides instead of looking at the accounts that actually bought." They rebuilt the target list around the mid-market profile those deals described. The consultant reported pipeline moving 6 weeks later.

The steps below are ones people described running. Most come from the co-founder and CEO of a developer-tools company, who credits the ICP exercise with getting the company to $1M ARR.

That CEO warned against settling on one or two traits like "engineers in startups with over 100 staff". They also found that tracking the companies that never bought taught them as much as tracking the ones that did. And they didn't think it should wait: do it "as soon as you start getting paying customers."

Churn is where the profile gets checked. The co-founder of a data-enrichment startup said they'd signed customers outside their ICP, and those customers churned. "Growth hides mistakes for a while. Retention eventually exposes them." The founder of a testimonial software company drew the same line from the other side and decided to stop reviewing sign-ups, churn and activation for users outside the ICP, "people that will probably never love us, or stay."

The record can move the profile too. One founder had a lead who didn't fully fit the ICP call back a few weeks after first contact, and rewrote the ICP because of it. We'd expect every ICP to get rewritten that way now and then. Once the profile exists, checking a single incoming lead against it is a separate job, covered in our guide to lead qualification.

Across the cases we found, nobody published point weights for their ICP. The ones who described how they scored used a few pass-or-fail criteria, a 1 to 10 rank, or a fit pile plus a pile for a person to check. Two of the four leaned on timing.

• A marketing and growth person who used to take every demo that booked started scoring ICP fit first, on company size, stack and one clear job. Demo volume dropped, and by their own account close rate nearly doubled.
• For the devtools agency co-founder in the opening table, an account qualified only when all three of their signals were true at once.
• One person gave an AI assistant their ICP criteria and asked it to rank 50 accounts from 1 to 10. It flagged 3 as urgent from timing signals that person said they'd have missed by hand.
• A builder demonstrating their own qualification product ran 3,000 companies through it. By the builder's count, 1,260 came out as fit and 61 were unclear, and those 61 went to a person for review.

Even in the 1 to 10 rank, the urgent flags came from timing. The outbound co-founder quoted earlier gives the signal a shelf life in their plan for starting over: "A signal is a window, and it closes in about two weeks." It's advice. Still, it's a good reason to put a date next to every signal.

Say a head of sales brings us 4,000 accounts and a slide that says enterprise. First we check the slide against the last 20 closed deals. If those deals are mid-market, so is the list. Each account then gets a yes or no on three or four criteria drawn from those deals, such as size band and whether it has the job the product solves, plus a signal dated within the last two weeks. Anything unclear goes to a person rather than a guess. Accounts that pass with a fresh signal make a short daily list of accounts ready to buy, and the rest wait.

All of it sits on one shared list, and one person owns the definition. Without that, things drift. A GTM person at a sales-software company has watched it happen past about 15 reps: every rep builds their own enrichment logic, and the ICP definition dies in a wiki page no one opens. The sheet beside this section shows the shape with invented accounts: three columns of yes, no or unclear and a verdict of fit, wait for a signal, a person checks, or not a fit.

An ICP usually describes only the accounts you want. The other half is who you won't sell to, and five of the six cases below drew that line only after a paying client had cost them.

• A software developer who builds products for founders fired a $5,000 client and refunded the $2K they'd prepaid, after tens of meetings, endless scope changes and then talk of cheaper agencies. That same day a new client paid $5,000 after a 15-minute call, and the project was done 10 days later. Their lesson: "only do business with action takers."
• An agency owner migrating a no-code app to code fired a client worth $12K a month, two weeks in. The client insisted tokens and commit count were the only way to measure output. The owner's warning works as a screening question: "If they think in commits, tokens, and parallel agents, and you think in shipped features and code quality, you will fight about it eventually."
• A finance professional fired a client worth 20% of their revenue. On paper it was profitable. Counting time, energy and attention, they were subsidizing it. "If you wouldn't sign a client today at today's terms, you're not keeping a customer. You're paying a subsidy."
• The owner of an AI-agents services business fired a client who paid on time and referred them business, but called 11 times a week about things already settled in writing. The client said they were relieved.
• A plastics distributor fired a customer, calling them "100% price shoppers" who "jump suppliers just to save a little." Most of the distributor's quotes never got an answer. Then on the thin-margin orders that did come, the customer was "all over me about every little detail."
• A commercial real estate firm that started a property-management company states its limits: "We don't go out of our pre-set area, we don't take every property, for the residential units, we have a minimum rent threshold." They didn't give the threshold.

These disqualifiers come from the same record as the closed deals and churn above. An account that only looked profitable and a pile of unanswered quotes are closed-deal and churn data too, even if nobody files them that way. So they belong in the document the reps already read, next to the fit criteria. We'd add a line to the won't-sell list the week a client like these shows up.

The founder of an integration-software company almost ran out of cash three times. Each time, the instinct was to cut prices and chase any deal they could find. "The third time, we did the opposite. Raised prices. Narrowed our ICP. Within a year, we were profitable." They raised prices and narrowed in the same move, while the money was running out and their own instinct said to widen, so the profit can't be split between the two.

But narrowing has a floor. The GTM person who watched an ICP die in a wiki page put it in one line: "Too narrow and there's nobody left." One founder hit a different wall. Their first company was growing fast and then wasn't, because the product was "a nice to have and not a priority for our ICP," and that market had a cap. Six months later, reorganised and focused on an adjacent market, they had growth back on track. Another founder knowingly targets an audience that helps them "grow fast but churn a lot," and says there's a reason behind it. A solo founder frames the same trade from what they've seen among founders they know: indie hackers bring easy growth and roughly 10 to 20% churn, other segments grow slower, with about 3% churn. Those figures are their impression. Their conclusion is short: "You just have to choose what works." An investor goes further and calls the ICP "an oversimplification," arguing that "The corner ice cream store probably has at least 5 customer segments of relevance."

The large before-and-after numbers need a second look too. Each one we found either came with other changes or came with nothing to check it against. The agency client in the opening table had its sales process rebuilt in the same months. A consultant reported that a SaaS founder's booking rate went from 1.2% to 6.8% on the same list, but the positioning and the angle were rewritten at the same time as the ICP. And an enterprise seller said win rate went from 8% to 29% when they narrowed the ICP in a past role, a one-line claim with nothing else shown. None of these lifts can be pinned on the ICP alone. We don't doubt the direction. We just wouldn't plan a quarter around the size.

So test it on your own record. Take the next 20 closed deals and the next accounts that churn, and check each one against the profile. If the wins fit and the churned accounts mostly sit outside it, the ICP is doing its job. If a profile outside it keeps buying, rewrite it, the way one founder did after an off-profile lead called back. Twenty deals is a small sample. It's the same number the consultant pulled in the enterprise audit, though, and it's enough to see whether the profile still describes who's buying.