In short
- 01"Derive the profile from closed-won accounts, then correct for the fact that closed-won only describes accounts you already targeted."
- 02"The 68% higher win rate everyone quotes is a self-reported January 2019 TOPO survey of 150+ account-based practitioners, not a measured result."
- 03"Roughly half the fields on a standard ICP template cannot be checked before contact, so they are discovery questions, not filters."
- 04"Set the width by send capacity. If your list is smaller than the invitations you can send in a quarter, tightening the profile costs you outreach and buys nothing."
- 05"Trigger-based notes accept at 50 to 60 percent against 15 to 25 percent for generic ones (LeadRiver, 50,000+ requests, April 2026)."
Every company has an ideal customer profile somewhere. It is usually a slide, it usually says "B2B SaaS, 50 to 500 employees, Series A to C, in Europe or North America", and it usually describes about four hundred thousand companies. That is not a profile. That is a market.
The test of an ICP is narrow and unforgiving: does it change who gets contacted this week? If two people build a prospect list from the same document and produce different lists, the document is decoration. What follows is the build, not the theory.
Why most ICPs never change a single send
An ICP fails for one of three reasons: it was written from ambition rather than evidence, it uses attributes nobody can verify before a call, or it is so broad that every account passes. Fix those three and the document starts doing work, because it finally begins rejecting things.
The ambition problem is the most common. Someone writes down the customer they want next year, usually larger and better funded than anyone who has actually signed, and the profile silently becomes a wish. The verification problem is subtler. "Struggling with manual prospecting" is a real qualifier, but you cannot check it on a company page, so it belongs in your discovery call, not in your filter set. And the breadth problem is arithmetic: a profile that accepts 90 percent of the companies you look at has not reduced anything.
A useful ICP is a rationing device. You only have so many contacts to spend, and its whole job is to decide where they go.
What the 68 percent win-rate statistic actually says
Nearly every guide on this topic opens with the same figure: companies with a strong ideal customer profile see 68 percent higher win rates. The number is real but it comes from a self-reported survey of about 150 account-based practitioners published in January 2019, and the chain of citation almost never reaches it.
The figure traces to TOPO's Account Based Benchmark Report, announced in a press release dated 29 January 2019, based on a survey of more than 150 practitioners at leading account-based organizations. The exact wording is "organizations with a strong Ideal Customer Profile (ICP) have 68% higher account win rates".
The clause immediately after it is the one nobody quotes: "more than 80% of the most successful account based organizations believe they have a strong ICP, whereas only 42% of other organizations believe they have a strong ICP." Believe. Both the win rate and the ICP quality are self-reported, in the same survey, by the same person. Nobody audited anyone's ICP.
Source: TOPO Account Based Benchmark Report, press release dated 29 January 2019, wording confirmed against two independent outlets carrying the same release. TOPO was later acquired by Gartner.
Follow the chain before you cite it
Trace the 68 percent through the pages that rank for this query and it rarely lands anywhere. One well-known vendor guide sources it to another vendor's blog, which sources it to a third vendor's blog. Three hops and no primary document. That is the normal state of B2B statistics, and it is worth remembering before any number changes a decision you make.
There is also a direction-of-causality problem that survives even if you accept the survey at face value. Building a real ICP costs weeks of analyst time and access to clean CRM data. Companies that already win are the companies that can afford both. "Winning firms have better ICPs" and "better ICPs make firms win" produce identical survey results, and this data cannot separate them.
None of this means skip the exercise. It means stop leading with a borrowed number and start leading with a mechanism you can verify on your own account. The mechanism is rationing, and rationing is arithmetic you can do yourself. The rest of this page is that arithmetic.
Build the profile backwards from closed-won
Pull your last 20 to 30 closed-won accounts, the closed-lost deals that reached a proposal, and everyone who churned inside six months. The profile is whatever separates the first group from the other two. Any attribute shared by all three describes your market, not your ideal customer.
The method is comparison, not description. For each candidate attribute, work out the share of won accounts that have it and the share of lost or churned accounts that have it. Keep the attributes where the gap is wide. Discard anything present in nearly everyone, however true it feels.
That last rule kills most of a typical ICP slide on the first pass. "B2B" appears in 100 percent of won and 100 percent of lost. It is not predictive, it is your address.
Three things worth doing while you are in there.
Weight by revenue and retention, not by logo count. An ICP tuned to predict closing but not renewal fills the pipeline and empties the account list a year later. If a segment closes fast and churns at month five, it belongs in the exclusion list, not the target list.
Ask who actually sponsored the deal. The ICP is a company profile and a person profile at once. If every win was sponsored by a head of sales and every loss was sponsored by a founder acting alone, that is a sharper filter than headcount will ever be.
Do the noise arithmetic before you believe a gap. This is where most closed-won analysis quietly goes wrong. With 20 wins and 20 losses, the standard error on a difference between two proportions is roughly 15 percentage points. An attribute present in 70 percent of wins and 55 percent of losses is one standard error out, which is indistinguishable from chance. The same attribute at 70 against 40 is about two standard errors out and worth acting on. The working rule at this sample size: ignore any gap under roughly 30 points, however good the story around it sounds. If you want the formal version of this, scoring prospects against real buying signals is the same comparison with weights attached.
Why closed-won only describes where you already looked
Closed-won analysis has a blind spot every guide skips. You can only win deals with accounts someone contacted, so the pattern in your wins is partly a pattern in your past targeting. If your team only ever prospected 50 to 200 person software companies, your wins will prove that 50 to 200 person software companies are your ideal customer.
This is selection on the thing you are trying to measure. Your closed-won set is not a sample of the market, it is a sample of the accounts your previous filters let through, filtered again by who happened to buy. Run the analysis, adopt the output, and next quarter's wins come from an even narrower slice. That feels like increasing precision. It is drift, and it compounds quietly, because every quarter the evidence agrees with you more.
Three correctives, in order of how much they are worth.
- Weight inbound and referral wins heavily. Those accounts arrived without passing your outbound filter, so they are the closest thing you own to an unfiltered sample of who wants your product. If your inbound wins look systematically different from your outbound wins, the outbound filter is the thing that is wrong.
- Read the losses for accounts you would not have targeted. Any deal that reached proposal despite failing your current filters is direct evidence against the filter. One of those is interesting. Three is a rewrite.
- Spend a fixed small share of your sends off-profile. Five to ten percent, permanently, on accounts that match a signal but not the firmographics. That budget is the only mechanism by which an ICP can learn something it does not already believe. Without it, the document can only ever confirm itself.
The honest framing is this. An ICP describes where you have proven you win. It is not a map of where you could win, and treating it as one is how a company talks itself into a market smaller than the one it actually has.
Which attributes actually predict fit
Split every candidate attribute into two families. Firmographics describe a state: what the company is, and whether the problem you solve could plausibly be worth money to them. Behavioural signals describe an event: something that just happened and put your problem on someone's desk this month. You need both, and they answer different questions.
Read the two blocks against each other and the point of the split is obvious. Firmographics predict whether an account could ever buy. Behavioural signals predict whether it will buy this quarter. A profile built only from the top half produces a list that is technically correct and commercially dead, because everyone on it is a fine fit and nobody on it is in motion.
The numbers on the second half are the ones worth remembering. LeadRiver's April 2026 review of more than 50,000 connection requests put trigger-based notes at 50 to 60 percent acceptance, against 15 to 25 percent for generic ones. The trigger is a behavioural attribute doing its job, and it roughly doubles the acceptance rate on the same list. For the full taxonomy of what is trackable, see 12 LinkedIn intent signals you can track without Sales Navigator and signal-based selling on LinkedIn. Hiring is usually the highest-quality single signal available for free, covered in LinkedIn hiring signals as buying intent.
One caution on funding, because it is the single most over-weighted attribute in B2B prospecting. A raise tells you money exists. It does not tell you the money is pointed at your category, and a well-funded company with no owner for your problem is a slower deal than a bootstrapped one where somebody is already being paid to fix it.
Write it as filters you can actually search
Every line of a working ICP must be checkable on public information before you contact anyone. If verifying an attribute requires a conversation, it is not a filter, it is a discovery question, and mixing the two is what turns an ICP into a slide nobody uses.
This is where the downloadable templates fall over. Take the field list that appears on almost every ICP template on the web and audit it one row at a time against a single question: can you determine this about a company you have never spoken to?
Count the rows. On a typical template, roughly half the fields cannot be checked before you contact anyone. They are not useless, they are in the wrong document. Move them to the call plan, and what is left is a query you can actually run.
So rewrite each surviving attribute as a query. "Mid-market" becomes a headcount range. "Scaling their outbound" becomes an open job post for an SDR or a growth role. "Cares about our topic" becomes an engagement event on a specific post. Then check that you can genuinely run each one:
- Company-level filters: headcount, industry and geography are checkable through a company search.
- Person-level filters: whether the owning title exists on the team, and how many of them there are, through a people search.
- Timing filters: open roles that reveal an active initiative, through a job post search.
Keep an exclusion list next to the target list, and take it as seriously. Competitors, agencies that resell what you do, companies below the size where your pricing makes sense, industries whose compliance regime you cannot serve. Exclusions are cheaper to apply than qualifications and they protect the sends that matter.
Then tier what survives, rather than gating everything to a single yes or no. Tier A accounts match the firmographics and are showing a behavioural signal right now. Tier B match the firmographics with no signal yet, so they are worth monitoring rather than messaging. Tier C match a signal but not the profile, and that is where the off-profile learning budget from the previous section gets spent. Tiering is how you keep a narrow profile without going blind to the market outside it. If you run a named-account motion, the same tiers slot directly into an account-based outreach plan.
Set the width by your send capacity, not by your taste
Narrowing an ICP has a cost nobody states: it shrinks the list. The right width is the one where your qualifying list runs slightly larger than the number of people you can actually contact in a quarter. Above that line, precision pays for itself. Below it, precision is self-harm.
Start with the ceiling, because it is fixed and everything else negotiates around it. LinkedIn does not publish an official connection-invitation limit, but the working consensus among practitioners sits around 100 invitations a week, roughly 15 to 20 on a working day. Over a 13-week quarter, that is about 1,300 invitations from one account. Now compare that number to the size of the list your filters return.
Read the first row again, because it is the mistake people make in the direction that feels responsible. A profile that leaves a thousand invitation slots idle is not disciplined. The accounts you would add by loosening it are not displacing better accounts, they are displacing nothing at all. Precision only has a price when demand for your sends exceeds supply. When your list is smaller than your capacity, tightening the profile costs you outreach and buys you nothing.
So the sequence is: count capacity first, size the list second, tune the profile third. Most teams do it in exactly the opposite order, then wonder why a beautifully argued ICP produced four meetings.
Ceilings are not targets
Two different numbers are doing two different jobs here, and spending the wrong one is how accounts get restricted. The roughly 100 invitations a week consensus is the rate you plan at. The daily caps are backstops that make a catastrophic burst impossible, not an allowance to use up. BeReach enforces 50 connection invitations a day as a cap no plan lifts, and holds profile visits under 120 an hour so a large batch never reads as a burst. Daily visit and message budgets start from a base of 300 visits and 70 messages and scale with the workspace, so treat them as the shape of a pacing plan rather than a number to chase. Read connection request limits for how the weekly ceiling actually behaves.
Test it on a real list before it decides who gets your sends
Take 100 accounts your filters return and score them by hand against the closed-won pattern. If more than 70 look like accounts you would genuinely want a meeting with, the profile is usable. Below 50, the filters are too loose. If the query returns almost nothing, they are too tight.
This step is cheap and almost universally skipped. It costs an afternoon, it happens before any message exists, and it is the only point where a bad profile is still free to fix. Manual scoring also surfaces the attribute you forgot to write down, because you will find yourself rejecting accounts for a reason that is nowhere in the document. Write that reason down. It is usually the best filter you own.
Run the test on public information only. Finding companies, checking who holds the owning title, and drafting the first message all work on public data, which means you can validate an entire profile before paying for anything or connecting an account. That matters if the alternative you were considering is Sales Navigator Core at $119.99 a month billed monthly, or $1,079.88 a year, which is about $89.99 a month, per LinkedIn's compare-plans page checked in August 2026. Buy the seat once the profile has earned it, not to find out whether it works.
Then price a wasted slot, so the precision argument stops being aesthetic. At roughly 400 invitations a month, LeadRiver's April 2026 analysis of more than 50,000 requests puts typical B2B acceptance at 30 to 37 percent, so call it 120 to 150 new connections. Belkins' 2026 study of 15.1 million touchpoints puts replies at 12.2 percent when you message an existing connection, against 7.9 percent on a cold connect-then-message sequence. Expandi's 2026 data puts the LinkedIn platform average reply rate at 10.3 percent, against 5.1 percent for cold email. Multiply it through and 400 invitations buys you somewhere near 15 conversations in a month. Every account your ICP lets through that should not have passed is one of those slots gone, and there is no way to buy the slot back. Precision is not a quality exercise here. It is the only lever you have that is not capped. Belkins has since restated its cold email benchmark at 0.45% on 7.5 million 2025 sends, measured against total sends rather than opens, so treat 5.1% as the 2026-vintage figure Expandi quoted.
If you want the per-contact version of the same discipline, monitoring intent signals without Sales Navigator is where the account-level profile turns into a person-level decision.
Every viral post is 100+ warm conversations waiting.
Tell your agent who you want to reach. It finds leads, qualifies them, sends personalized outreach, and follows up.
Keep it alive, or it rots
Revisit the profile quarterly, and immediately after the third win it would have rejected. Three surprises is not noise, it is a segment you have not written down yet. The maintenance loop is short: re-run the closed-won comparison on the last quarter, and move anything that stopped separating won from lost.
Watch the reply rate by segment as your early-warning system. A segment whose replies fall while the rest hold steady is usually one where the profile has drifted, not one where the copy went stale. And keep every rejected account in a list rather than deleting it. When the profile widens later, that list is the first place to look, and it is already scored.
The off-profile budget is what makes the loop work rather than just repeat. Five percent of your sends aimed at accounts the profile rejects is the only stream of evidence that can contradict it. A profile that has not been contradicted in a year has usually not been tested in a year.
One last discipline that has nothing to do with the document. A tight ICP makes the message easier to write, because you know what the person is dealing with before you type. It does not write the message for you, and it does not remove the need for a human to read what goes out. In BeReach, research and qualification and drafting all run without a LinkedIn session connected at all, and the account is only needed at the moment something actually sends, with a person approving every message first. The profile decides who. You still decide what.
Frequently asked questions
How narrow should an ideal customer profile be?
Narrow enough that your qualifying list is slightly larger than the number of people you can contact in a quarter, and no narrower. At the practitioner consensus of roughly 100 LinkedIn invitations a week, a quarter is about 1,300 sends from one account, so a list of 1,500 to 2,500 contacts is the right order of magnitude. If your filters return 50,000 companies the profile is not rationing anything. If they return 300, you are leaving about a thousand invitation slots unused every quarter, and tightening further costs you outreach while buying you nothing.
Is the 68 percent higher win rate statistic about ICPs reliable?
Treat it as suggestive, not as evidence. It comes from TOPO's Account Based Benchmark Report, announced 29 January 2019, based on a survey of more than 150 practitioners at leading account-based organizations. Both ICP strength and win rate were self-reported by the same respondent, so nobody audited any actual ICP. The population was mature account-based teams rather than B2B companies at large, and the causality plausibly runs backwards, since firms that already win can afford the analyst time an ICP takes. TOPO was later acquired by Gartner, and what circulates today is the January 2019 press release rather than the report behind it.
What is the difference between an ICP, a buyer persona, and a TAM?
Three different objects. The total addressable market is every company that could theoretically buy, and it is too broad to route a single send. The ideal customer profile is the subset where you demonstrably win, retain and expand, and it is what builds the account list. The persona describes the human inside the account who owns the problem and can sign for it, and it decides who on that account gets contacted and what the message says. Confusing the ICP with the TAM is what produces a profile that accepts nine accounts in ten.
How many closed-won deals do I need before an ICP is credible?
Twenty to thirty gives you a pattern you can lean on, and even then the arithmetic is unforgiving. At 20 wins against 20 losses, the standard error on a difference in proportions is around 15 percentage points, so any gap under roughly 30 points is inside the noise. Under ten deals, treat every attribute as a labelled hypothesis and test it against a real list rather than defending it. With very few deals, weight the accounts that renewed above the ones that merely closed, because early revenue is a noisy signal and retention is a quieter, more honest one.
How do I build an ICP if I have no customers yet?
Build it from the market rather than from your history, then plan to be wrong. Write three or four competing hypotheses about who has the problem, phrase each one entirely as public filters you can search, and pull a sample list for each. Score 100 accounts per hypothesis by hand and see which list you would actually want meetings with. Then run a small deliberate send against the two strongest and let acceptance and reply rates rank them. This is slower than copying a competitor's positioning, and it is the only version that produces a document your own sends will obey.
Reading this in an AI assistant? Hand it the page and let it summarise, so you can ask follow-up questions against the whole argument rather than the part you have read so far.
