ICP & Account-Scoring Playbook

An ideal customer profile (ICP) describes the companies most likely to buy from you, succeed with your product and stay. Build it from the deals you have already won and lost, not from a strategy offsite, and you can rank the accounts worth your limited time before anyone spends an hour on them.

Many founder-led teams sell to whoever shows up. An ICP replaces that with a short list of companies, described by traits you can see from outside: size, the tools a company runs, the events that start a search. It describes companies, not the people inside them, and it is narrower than your total addressable market: the slice where you win fastest, keep customers longest and expand most easily. It comes after you have chosen the segment to win first, which the Customer Segmentation Playbook covers, and before you score the people who respond, which the Lead Management Playbook covers.

Vendor and agency write-ups circulate win-rate lifts for a defined ICP, but they divide by different sets of deals and most disclose no sample, so I do not quote one here. Your own closed deals give you the number that matters.

Build it from evidence, not vibes. The evidence is every deal that closed in the last 12 to 24 months, won or lost, and every customer that churned. Losses and churn matter as much as wins: a trait that is just as common among losses tells you nothing, and the disqualifiers are as informative as the matches.

Record the same handful of facts for each deal, all of them things you can see from outside: size, industry, the tools the company runs, whether it had just raised money or was hiring for a relevant role, who signed, and, for losses, the reason. Fill the gaps from the company’s website and job posts, not from memory.

For each trait, compare the win rate of deals that had it with the win rate of deals that did not. A trait most of your customers share is not enough: if it is just as common among losses, it describes your market, not your best accounts. Keep the three to five traits with the widest gap, and read the churned customers separately, because what they share often becomes a disqualifier in Step 3.

Small numbers need care. With 20 deals on each side of a trait, one deal moves a win rate by 5 points, so trust wide gaps only and expect Step 6 to test them again.

Firmographics alone describe a company; the other layers predict whether it will buy. Write the traits from Step 2 into five layers, and leave a layer empty rather than fill it with a guess:

1. Firmographic fit. Industry, employee count, revenue or stage, geography. Be specific: “B2B SaaS, Series A–B, 50–500 employees” beats “tech companies.”
2. Technographic signals. The tools they already run. A company using your direct competitor already knows the category and has a budget for it, so the conversation is about switching, not about why the category matters. Tools can define a profile on their own: Klaviyo, which sells a customer-relationship platform to consumer companies, reports that about 77.9% of its annual recurring revenue at the end of 2025 came from customers who also use Shopify^[1].
3. Buying triggers. Events that start a search: hiring for a relevant role, a funding round, rapid growth. Write down the ones your won deals had in common; Step 5 uses them to time outreach.
4. Organizational readiness. Does the account have the buying process, budget, and the right economic buyer, and can you reach the people who decide? The B2B buying committee maps who they usually are.
5. Negative indicators (the anti-ICP). Red flags that disqualify even a firmographically matching account. Naming these is what stops reps chasing deals that were never going to close.

Then type the profile into a company database. It should return a list you could actually work: not every company in your market, and not a dozen. If it excludes almost no one, it reads like a personality sketch and is too vague to use.

Turn the profile into a number. Give each ICP signal a weight reflecting how strongly it predicts a good deal, score every account against those weights, and convert the result into a 0–100 fit score and an A, B or C tier. The gaps from Step 2 are your starting weights: the wider the gap, the heavier the weight. Disqualifiers sit outside the score: an account that hits one is out, whatever it scores.

You do not need an AI model for this. In my view, while you have dozens of closed deals rather than hundreds, a rules-based score with explicit weights is the better choice: a predictive model has too little history to learn from, and a score your sales team can explain is one it will trust. Graduate to prediction only once you have the closed-won history to train on.

Fit and intent answer two different questions. Fit asks whether you should sell to this account at all. It is built from the profile, and it is stable: recompute it monthly, not hourly. Intent asks whether now is the moment. It is built from what the account is doing, and it is volatile by design: it spikes and decays.

At any time, most accounts that fit are not buying. John Dawes of the Ehrenberg-Bass Institute estimates that if companies replace a supplier about once every five years, only about 20% are in the market in a year and about 5% in a quarter, and he says the figure is not meant to be precise^[2]. Fit tells you the right list. Intent tells you the right week. Start with fit, then layer intent to time the outreach.

Intent comes from three places. Your own data: pricing-page visits, demo requests, several people from one account engaging at once, which Step 6 of the Lead Management Playbook turns into an account-level rule. Public triggers: job posts for a relevant role and funding rounds; in the US, a company raising money under Regulation D must file a Form D with the SEC within 15 days after the first sale, and the filing is public^[3]. And paid third-party data, which reports research done elsewhere: G2’s Buyer Intent, for example, flags companies that viewed your G2 profile, a comparison page that includes your product, or a competitor’s page in your category^[4].

A scoring model is only trustworthy if it agrees with reality. Run your strongest existing customers through it. If they don’t land in tier A, your weights are wrong: adjust until your known-good accounts score the way you’d expect. Then score last year’s closed deals and check that the tiers separate: tier-A deals should have been won clearly more often than tier B, and tier B more often than tier C.

Be honest about what this proves. When the weights came from the same deals, the test flatters them. The real test is next quarter’s deals, scored before anyone knows the outcome, which is the first job of the review in Step 7.

An ICP is a living model, not a one-time artifact, and it degrades two ways. Data decay: the facts behind each score go stale as companies move, hire, re-platform and get acquired, and the people change too: the median US wage and salary worker had been with their current employer 4.1 years in January 2026^[5]. ICP drift: in my view, a team chasing a number slides from urgent, high-fit accounts toward broader ones that are easier to start a conversation with, and cycles lengthen and win rates fall without anyone deciding it.

The fix is a fixed cadence: a quarterly ICP review owned by whoever runs revenue, with input from sales and customer success, re-validated against the quarter’s closed-won and churned accounts. My rule: change a weight only when the evidence holds for more than one quarter.

What changes by stage?

What mistakes should I avoid?

• A profile written in an offsite. Start from closed deals (Step 1).
• Studying only your wins. A trait as common among losses tells you nothing (Step 2).
• A profile that excludes no one. Write the disqualifiers down (Step 3).
• Weights no one can explain. Sales ignores a score it cannot check (Step 4).
• Chasing intent from accounts that do not fit. Let the tier decide what a signal earns (Step 5).
• Treating version one as final. Test it on next quarter’s deals (Steps 6 and 7).