Ideal customer profile for B2B startups
An ideal customer profile is a live hypothesis about which companies you can win, written so it can be proven wrong. Three customers cannot give you a pattern, but they can give you a prediction worth testing. This guide covers company-level fit only.
By Kshitij Maheshwari, co-founder · Updated August 2026 · 25 min read
Written by operators who run outbound for seed-stage teams, not by a vendor selling you the field list.
What an ideal customer profile actually is
It is a company-level claim about who you can win, and the useful version is a prediction rather than a portrait.
An ideal customer profile is the set of conditions that make a company winnable for you: what has to be true inside it before your product is the obvious answer.
Ask three sources what belongs in one and you get three different objects back:
A firmographic filter set you paste into a data platform. Honest about its job, which is building a list.
A description of the company that gets the most value from you. Broad, and it drifts into personas within a paragraph.
The segment where retention and enthusiasm concentrate, found by surveying a user base you do not have yet.
Read your ICP and ask what would have to happen for it to be wrong. If nothing could falsify it, you have written a description of your market, not an ICP.
ICP, buyer persona and market size are three different objects
Three separate questions get answered by one document on most sites, which is why the document ends up true of everyone and useful to nobody.
| The object | The question it answers | Where it sits |
|---|---|---|
| Ideal customer profile | Which companies can we win, and what has to be true inside them? | This guide, company level, start here. |
| Buyer persona | Who inside the account do we message, and what do they care about? | Out of scope here. It is the person layer, and it comes second. |
| Market size | How many of those companies exist, and in what order do we work them? | A counting question. TAM enrichment is the done-for-you version. |
Get the company right first. A perfectly researched persona inside a company that cannot buy is a dead end, and it is the more expensive mistake because it feels like progress the entire time you are making it.
Why the worksheet fails a founder with three customers
Every ICP template assumes a corpus of closed accounts, and the assumption stays invisible until you sit down to fill one in.
- ✕Your last twenty closed-won accounts
- ✕Your last twenty churned accounts, for contrast
- ✕Twenty to thirty customer interviews
- ✕A running survey program to segment
- ✓Three customers, one of whom bought as a favor
- ✓Eleven losses, most of them never coded
- ✓Forty conversations that went nowhere
- ✓An onboarding and a support inbox you ran yourself
Both columns are real. Only the second one is yours, and the rest of this guide is written for it.
Your ICP is a hypothesis, not a document
Steve Blank settled this twenty years ago, and the template industry un-settled it.
A hypothesis is a claim specific enough to be wrong. "Mid-market SaaS companies" cannot be wrong. "Runs its own support queue, with a named owner, past 500 tickets a month" can.
Blank's framing, on his own site in 2010 and in The Four Steps to the Epiphany five years earlier: a startup is an organization formed to search for a repeatable and scalable business model. Who the customers are is the first hypothesis in that search, not a decision you defend.
Write the profile, then write the sentence that would kill it. If you cannot write the second sentence, the first is a description. Every line gets a falsifier beside it, or it gets marked assumption.
The document is not the deliverable. You are building a claim you can hold up against next month's replies. We have watched founders spend a week on the formatting and never once compare the profile to who actually answered.
The firmographic trap
Company size and industry are how you find a company in a database, and almost never why it buys.
| The attribute you can buy | The threshold underneath it | Why the proxy leaks |
|---|---|---|
| 50 to 200 employees | Processes more than 500 invoices a month | Headcount per invoice varies by an order of magnitude across industries. |
| Industry: SaaS | Sells to buyers who run a security review | Half the category never sees one. The other half sees one per deal. |
| Uses a given CRM | Has a named owner for the pipeline data | The tool is installed everywhere and owned in about a third of places. |
| Raised a Series A | Someone can sign for this line item this quarter | Stage predicts that budget exists, never that anyone owns it. |
- !A proxy nobody labeled as a proxy becomes a fact in six weeks
- !Firmographics are the lookup key, applied last, after the conditions
- !The companies that sell those fields wrote most of the ICP guides online
- !If every criterion is purchasable, you have described a list
The fit criteria that survive contact with reality
The criteria that predict a win are situational, and Steve Blank wrote the usable list of them down twenty years ago.
The five conditions
They have the problem
Not a problem your category solves. This one, inside this company, this quarter. If you cannot name the hour of the week where it shows up, you are guessing.
They know they have it
Awareness is the line between a sale and an education program. A company that has not noticed the problem is a two-year project you cannot fund at this stage.
They are actively looking, with a timetable
Somebody owns finding an answer, and a date is attached to it. Without the date the deal has no reason to close this quarter, or any other quarter.
They have already cobbled together a fix
The strongest of the five and the easiest to observe. A company running a spreadsheet to do your job badly beats a perfect firmographic profile with no spreadsheet in it.
They have budget, or can get it quickly
Not necessarily a line item. A named person who can move money without assembling a committee, inside the length of your sales cycle.
Not one of the five is a firmographic. Blank published the list on his own site in 2010, and the concept comes out of The Four Steps to the Epiphany in 2005, which makes it older than the templates that ignore it.
States, not events
A state stays true for months: they run their own support queue, they cross the volume threshold, someone owns the problem. An event has a decay clock: they raised, someone moved, they launched.
States belong in the ICP: they decide whether you can win the account at all. Events belong in the signal layer: they decide whether it answers this month. Collapsing the two is how a good profile produces a dead campaign.
The three-condition segment test
Bill Aulet's test for whether a segment is actually a segment runs at three customers and takes about ten minutes.
-
1
They buy the same product
No forks, no special versions, no one-off integration to get a signature. If closing two customers meant building two slightly different products, you have two segments.
-
2
They buy through the same sales process
Same demo, same objections, same security review, roughly the same length. If the motion changes from one customer to the next, the segment is too broad and you are running both halves at half strength.
-
3
There is word of mouth between them
Did any of your first customers introduce you to someone in the same role somewhere else? If none did, either the segment has no internal network or the product is not yet worth risking a reputation on.
The third condition is the one nobody runs, and it is the only one that costs nothing. If your customers do not talk to each other, you picked a list, not a segment.
The three conditions come from Step 2 of Aulet's Disciplined Entrepreneurship (2013), as published on the book's own companion site. It was written for founders before product-market fit, which is exactly why it still runs at three customers.
Which of your customers count as evidence
Before you mine three accounts for a pattern, throw some of them out.
Count every logo as a data point
All three are 30-person fintechs, so that is the ICP.
- ✕The favor deal an investor asked for
- ✕The pilot nobody opened after week two
- ✕The deal carried by a champion who has left
Run the admissibility test first
Could we win this account again, at list price, with nobody we know inside it?
- ✓Yes: it is evidence, use it
- ✓No: it is a story, keep it out of the sample
- ✓Write down what you excluded, and why
- !Two of your three came out of the same Slack community
- !Or the same investor introduction, or the same conference
- !Write how each one found you next to what they have in common
- !If the two columns match, you learned about your channel, not your buyer
None of this argues against founder-led selling, which is how the first accounts get closed. It is also what makes them unrepresentative: you win on relationship what the next cohort has to be sold on merit, and only the second kind of win locates the segment.
What three customers can actually tell you
Three admissible accounts cannot support a pattern, but they can support five specific readings, in this order of signal strength.
Closed-lost, coded by reason
The largest sample you own. Separate "never a fit" from "a real evaluation that went the other way", because only the second one is telling you something about fit.
Time to first value, per account
The account that took eleven weeks is telling you something the account that took four days is not. Usually it is telling you where the segment stops.
Where your own hours go
The account eating disproportionate founder time is either outside the segment or telling you the product is unfinished for it. Both are findings, and both change what you do next.
Who introduced you to whom
Aulet's word-of-mouth condition, checkable today and free. One introduction to the same role at another company is worth more than a page of firmographic overlap.
The disappointment question, at account level
Which companies would be very disappointed to lose you? Rahul Vohra's account of building Superhuman's product-market-fit engine, in First Round Review, puts directional results at around forty respondents.
Your losses are the bigger dataset
Before product-market fit the honest picture is forty conversations and three purchases, and the thirty-seven are the dataset. What matters is the reason coded at the moment of loss, in the prospect's own words: no budget, no owner, solved it another way, not painful enough yet, wrong shape of company. Four of those five are fit findings.
A disqualification is an account that was never a fit and should never have entered the pipeline. A closed-lost is a real evaluation that went the other way. Only the second one is about your profile.
What outbound replies add that customers cannot
Customers arrive when they arrive, and outbound is the only ICP evidence you can generate on demand.
- ✕Define the ICP properly first
- ✕Then build the list against it
- ✕Then send, once the profile is settled
- ✕Learn about the market later, from customers
- !A reply is interest, not fit, and not budget
- !A silent slice may have had a weak offer, not the wrong companies
- !Two weeks of sending is a read, never a verdict
- !Positive replies from outside the profile are the most interesting rows and the easiest to over-read
The order is backwards. A batch of outbound reads a named slice inside two weeks, faster than any customer will ever arrive, and that is the real reason a pre-product-market-fit founder should be sending at all. The pipeline pays for the exercise. The read on the segment is what you keep.
Writing it down: the one-page ICP
One page, four columns, and the fourth is the one nobody writes and the only one that makes the document worth keeping.
| The line | What goes in it | The evidence behind it | What would falsify it |
|---|---|---|---|
| The situation | The condition that has to be true inside the company | 2 of 2 admissible customers, 7 of 11 losses | A segment without it replies and buys anyway |
| The workaround | What they built to cope before you existed | Named in every onboarding call so far | A closed-won that had no workaround at all |
| The threshold | The volume or complexity point where it starts to hurt | Assumption, nothing observed yet | Wins clustering below the line you drew |
| The lookup key | The purchasable attributes that find companies in that state | Chosen, not discovered. A proxy, and labeled one | The key returns accounts that fail the conditions on inspection |
| The motion | One demo, one process, one onboarding path | True of both admissible customers, no forks yet | A win that needed a different sales process |
Lines with no evidence still get written down, marked assumption. Marking them is the whole discipline: six weeks later, an unmarked assumption is indistinguishable from a finding, and nobody in the room remembers which was which.
Narrow beats broad, and the honest reason why
You narrow so the experiment works, and the focus everybody talks about is a side effect.
Narrow is a learning-rate decision
Two slices at 150 accounts each tell you something. Six slices at 50 each tell you nothing and cost the same money. You narrow to make the result readable.
Big enough to matter, small enough to lead
Geoffrey Moore's rule for a first segment, in his own words in a 2024 interview: big enough to matter, small enough to lead, and a good fit with your crown jewels.
Getting the market wrong is what kills companies
CB Insights coded 385 shutdowns in its March 2026 report: 43% cite poor product-market fit, most early-stage. That is evidence the stakes are real, not proof that narrow beats broad.
Nobody has proven narrow beats broad
There is no controlled study. There is a mechanism, Aulet's three conditions, Moore's sizing rule and the failure data. That is an argument, not a proof.
What the wrong hypothesis costs
A quarter on the wrong segment costs the quarter, the sending reputation of every domain that touched it, and the sequence you did not write for the right segment.
Blank and Aulet wrote for the stage you are at, before product-market fit, with almost no data. Moore did not. Crossing the Chasm starts once early adopters are already yours and the pragmatic majority is next. Take the beachhead sizing rule; leave the whole-product apparatus until you have a chasm in front of you. Applying it at seed is the most common misread of the book.
The instinct at seed is that a wider profile is a safer one, because it feels like more shots on goal. In the campaigns we run it works the other way: a wide list returns a result nobody can read, so the next decision gets made on exactly the same guesswork as the last one.
Fit is not timing
A perfectly fitting company that changed nothing this quarter will not reply, and that is not an ICP problem.
Fit decides who goes on the list. Timing decides who answers this month. Fit conditions are states that hold for months. Buying signals are events with a decay clock: a raise, a new owner for the problem, a hiring push for the pain you solve. A good profile with no timing layer produces a correct list and an empty calendar.
The evidence for why timing carries so much weight, including Gartner's finding on organizational change and the 95-5 rule from the Ehrenberg-Bass Institute, sits on our signal-based selling guide. It is not repeated here, because one page should own a number.
Not sure whether your ICP is wrong or your timing is?
Book a Fit CheckTurning the ICP into something outbound can test
A hypothesis you cannot settle inside one batch is not a hypothesis yet.
-
1
Write the prediction, with a number in it
Name the slice, the size of the batch and the reply rate you expect. A prediction with no number cannot be disconfirmed, which is the entire reason for writing it down.
-
2
Pick a batch you can actually read
Enough accounts that one extra reply does not flip the answer. At the rates a seed team sees, that is closer to 150 accounts per slice than to 50.
-
3
Change one thing between slices
Same offer, same sequence, one differing fit condition. Move two variables and neither slice can tell you which one moved the number.
-
4
Freeze the hypothesis for the full cycle
Define the window before you send: the send cycle plus the reply tail. Changing the profile mid-cycle guarantees you never learn whether it was right, and it is the most common way six months pass with no readable result.
When one hypothesis is not enough
Four customers in four unrelated segments is common and honest, and every template assumes it cannot happen. Running two named hypotheses against each other, each with its own predicted reply rate, beats forcing one description to cover both. Four is not workable, because you cannot fund a readable test of four.
When to revise, and when you are just flinching
Change the profile when a prediction is disconfirmed, never on a calendar and never after a bad fortnight.
Review it every quarter
Q3 ICP review: added "and mid-market" to the size band.
- ✕A date cannot know when you learned something
- ✕Cosmetic edits feel like progress
- ✕It widens a little every quarter until it fits everyone
Set the trigger before you send
If Slice B replies under 2% across two batches, we drop the volume threshold.
- ✓Written before the data arrives
- ✓One disconfirmation, two independent batches
- ✓You know in advance what would change your mind
A prediction missed or beaten, twice
A slice misses or beats its predicted reply rate by a margin bigger than one extra reply could produce, across two independent batches. Once is weather.
A win you did not predict, twice
You close a company the profile said was out. Once is luck. Twice is a condition you have not written down yet, and the interview is worth an hour.
Churn or support load concentrating
Both piling up on one attribute. That is the profile telling you what you can serve, and what you can serve outranks what you want.
Before you touch anything, ask how many individual replies would have to change to erase the difference. At a 4% reply rate a 50-account batch yields two replies, so two becoming four looks like a doubling and is one person having a quiet Tuesday. If the answer is one or two, you learned nothing this month.
Easy to list is not evidence of fit
A new tool makes a new segment cheap to build a list for, so the segment quietly goes in. Nobody decides to widen the profile, and by the end of the quarter it covers twice the market with the same sending capacity and no readable result anywhere.
A worked example, start to finish
Illustrative, not a client result. A hypothetical seed company, written to show the method end to end. Real Good GTM publishes no case studies and we do not invent numbers.
Two customers, not three
- Three logos, eleven losses on the board
- One logo bought as an investor favor
- It fails the list-price test, so it comes out
The exclusion gets written down with its reason
A workaround and an owner
- Both admissible customers ran a spreadsheet for it
- Seven of eleven losses had nobody who owned the problem
- Neither condition is on any company record
The condition writes itself out of the losses
Five lines, four columns
- Each line carries its evidence and its falsifier
- The volume threshold is marked assumption
- Headcount goes in as the lookup key, labeled a proxy
Nothing here is claimed as a finding yet
One condition apart
- Two slices of 150 accounts, same offer, same sequence
- Slice B differs only on the assumed threshold
- Predicted reply rates written down before the first send
The abandon condition is set on day one
The trigger does not fire, so nothing changes
- Slice A lands close to its prediction
- Slice B lands two replies below, which is inside the noise
- One or two replies would erase the gap, so the profile stays put
- The threshold line stays an assumption, and the next batch is designed to test it
The discipline is the quarter you do not spend rewriting
Where ICP work goes wrong
Four failures, and none of them look like failures while they are happening.
"50 to 200 employees" starts as shorthand for a volume threshold and ends up as the criterion itself, unlabeled, six weeks later.
At three accounts, one atypical logo is a third of the evidence. Written up warmly, it becomes a description of a relationship.
Three customers out of the same community share a distribution path, not a buying condition. The profile ends up describing your network.
A tool makes a segment listable, one loud prospect complains, a quarter goes badly. Every edit is small and nobody ever decided to broaden it.
Where the common advice is wrong
Most ICP content is written by the companies that sell the fields it tells you to fill in, and it shows.
"Take your last twenty closed-won accounts, find the firmographics they share, and review the profile every quarter."
- ✕Start with firmographics, add the situation later
- ✕Mine twenty closed-won and twenty churned accounts
- ✕Review it on a quarterly cadence
- ✕A strong ICP wins you 68% more accounts
"Write the situation first, mine the losses, and change the profile only when a written prediction fails twice."
- ✓Situational conditions first, firmographics last as the lookup key
- ✓Forty accounts is not a method for a founder who has three
- ✓Revise on a disconfirmed prediction, not on a date
- ✓We read the eighteen pages ranking for these queries in August 2026. Not one cites a benefit statistic
The number is real and traceable, which is what makes it worse. It comes from TOPO's 2019 account-based benchmark, a survey of more than 150 practitioners at enterprise account-based organizations, where both the strength of the profile and the success were self-reported. TOPO was acquired in 2020, so the report survives on vendor mirrors and nobody can check it. Blogs credit the figure to four different research houses. None of them names TOPO.
What to realistically expect
Expect a slower answer than the templates imply, and a more useful one.
Months, not a workshop. The reframe that helps: you are not waiting for the profile to be right, you are shortening the loop that tells you it is wrong. Write the first version this week and treat it as disposable.
A profile does not make a weak offer land. It decides which companies you spend the offer on. If a well-built slice returns nothing at all, look hard at the offer before you rewrite a single line of the profile.
Who you sell to sets the ceiling. ChartMogul's retention report, 2023, puts 2.7% of businesses under 10 dollars in average revenue per account (ARPA) above 100% net revenue retention, against 41.1% of businesses above 500 dollars. Vendor data, and correlation rather than cause.
The ICP you can serve beats the ICP you want. A segment you can win and cannot keep is worse than one you never entered, because it eats exactly the founder hours that would have found the right one.
What works now, August 2026
Current practice as of August 2026, labeled current rather than eternal, and worth re-checking before you copy it.
The fit criterion no longer has to be a field you can buy
Clay and its peers now ship research agents that answer a written question across a few hundred accounts. That matters because the conditions that predict a win are almost never available as a filter.
An agent verdict is a research lead, not proof
Require a source URL per answer and spot-check a sample by hand before anything sends. This is a capability claim about the tools, not a performance claim about results.
Layered ICP scoring is where you are going, not where you start
Every 2026 vendor guide argues for a scored model. Fitting one needs closed-won and closed-lost history you do not have, and a model fitted to eleven data points is decoration.
Coding closed-lost by reason is the fastest evidence you can instrument
A structured reason code plus the prospect's own words, captured at the moment of loss. "We lost" is not evidence, and almost nobody at seed stage sets this up.
Slices tested in outbound batches
This page owns the hypothesis design. The batch sizes, what to track and the kill rules are in the ICP slice experiments play.
The 2026 tooling makes it cheap to check a condition across three hundred accounts, and that has changed what can go into a profile. It has not changed where the condition comes from: your losses, your onboarding calls, and the support queue you answer yourself.
- 1 An ICP is a prediction you can disprove, not a description.
- 2 Situational conditions first, firmographics last as the lookup key.
- 3 Drop any account you could not win again at list price.
- 4 Your losses outnumber your wins and carry more fit signal.
- 5 Every line gets its evidence and its falsifier, or it is marked assumption.
- 6 Narrow enough that one batch of outbound returns a readable result.
- 7 Revise on a disconfirmed prediction, never on a date.
Questions founders ask
How do I define an ICP when I only have three customers?
Where does the ICP stop and the buyer persona start?
How many customers do I need before a pattern is real?
My three customers are in three different industries. What do I do?
One of my customers bought as a favor. Does that count?
How often should I update my ICP?
Can I use AI to define my ICP?
Co-founder of Real Good GTM. He has been the first business hire and Chief of Staff at seed-stage B2B startups, which mostly meant writing a customer profile on a Monday and finding out it was wrong by Friday. This guide is the version of that loop he wishes someone had handed him.
Connect on LinkedInFrom the hypothesis to the test
You have the profile. These three take you into the work: the batch that tests it, the loop that revises it, and the timing layer that decides who answers.
ICP slice experiments
How to run two slices against each other: batch sizes, what to track per slice, and the rules for killing one early.
Read the playTurning outbound into GTM learning
The loop that revises this profile: what a campaign tells you about your market, and how to write it down so it compounds.
Read the guideSignal-based selling
The timing layer this guide hands off to: which events open a window, how long each one stays open, and the evidence behind it.
Read the guideWant a second read on your ICP?
Book a fit check. We'll go through the customers you have, the losses you have not coded, and whether there is a slice worth testing this quarter. If outbound is the wrong motion for your stage, we'll tell you that instead.
Book a Fit CheckNo hard sell. No fake numbers. Real good work speaks for itself.