Poldermarketing
How do I define the right ICP for an early-stage startup?
Define the early ICP as one situation you can falsify: the problem, the alternative, the trigger, and who can say yes this quarter. Add only the traits that make those accounts findable, plus who you will refuse. That page is the beachhead for this cycle, not the whole market. Rewrite it when the same calls stop clustering.
Start from the situation, not a company box
The right ICP for an early-stage startup is the situation in which a specific account will try an imperfect product now. It is a hypothesis you can falsify, not a demographic and not a later market.
Michael Seibel puts it in a picture: hair on fire, grab a brick, do not wait for a hose. Andreessen’s point there: the market pulls the product out of the startup. “SaaS companies with 50 to 500 employees” has no fire, no current alternative, and no reason to act this quarter.
Kevin Hale says a problem worth a company is popular, growing, urgent, expensive, mandatory, or frequent. At least one of those six is required, and more than one is better. Count a trait only when buyers already show it, not when a slide asserts it. A solution in search of a problem fails his test. Narrow the buyer and the circumstance, not the problem’s existence.
Split the account, the beachhead, and the person
An ICP is the account in a situation, a beachhead is this cycle’s slice, and a persona is a person. The total market is a later ceiling.
| Term | Names | Use it for | Wrong when |
|---|---|---|---|
| ICP | Account in a situation | Pursue or refuse | A job title only |
| Beachhead | This cycle’s slice | List, page, and calls | The whole market |
| Persona | One person | Wording and who to book | Choosing which companies exist |
| Total market | Later ceiling | A line in the deck | This month’s list |
| Refusal list | Who is out | Stopping a bad deal | Hoping volume will teach you |
Michael King at a16z wants answers a list can use: size or a revenue band, business type, geography, industry, titles, the problem, the facts that define it, technologies, and buying behaviors. “Enterprises and midsize companies that use computers” fails. Personas stay out of the ICP. HSBC Innovation Banking, quoting Rob Denton, splits B2B the same way: company, then person. For a consumer product, the person in the situation is the account.
Clarify says not to mix buyer and user, then its FAQ calls the ICP an individual. Ignore the FAQ. Directive Consulting names buyer, champion, IT or security, and procurement as an enterprise committee. Four separate people is the wrong first beachhead unless that pain exists nowhere else you can staff.
Run segments through gates before you score them
Pass or fail promising segments in order. Scores from 1 to 5 for pain, urgency, willingness to pay, reach, and fit hide a fatal miss: a 1 still leaves a large product, a 5 is not five times a 1, and reach can beat a true fit.
- They already act. Eric Migicovsky asks what they did and what they dislike. “Sounds bad” is a fail.
- One shared alternative. A spreadsheet and a vendor are two segments.
- Someone you can book can commit on a cycle you can survive. Directive Consulting wants the shortest cycle, the clearest pain, and the fastest adoption. Amplify Partners cites Hex: a giant firm would not roll a year-old prototype across the company. WarpStream’s cofounder, in that guide, starts where the problem is current and the buyer will take the risk.
- A win transfers: same onboarding, same objection, a peer they would actually introduce, and no custom build you cannot staff. If ten accounts that like the product would not introduce each other, it is a list, not a beachhead.
- You can point at accounts. Amplify’s ceiling: 200 to 300 firms can be a design-partner pool, and 15,000 means you named a market. That is not a contact quota.
Stackmatix calls reach a channel problem and fit the strategy. Two passers that differ by one variable stay two hypotheses. Do not join them with “or”.
With customers, prefer who gets value, buys quickly, and stays over the largest deal. A lost budget or a departed champion is process. Failure at the job is the product. King separates a missing capability from a sales path you could copy. Clarify flags an investor’s company at a different size as the same error.
Keep a seed profile narrow enough to refuse
At seed, narrow means one situation, one role you book first, and one buying path, so a stranger can accept or reject an account. Too narrow means only people who already like you, or one loss retiring a category that is still real.
Daniel Ahmadizadeh, in YC’s first-time founder roundup, says defining the customer early and precisely would have saved time later, at the expense of growth during the batch. Letting anyone sign up left the target unclear and caused the product-market fit trouble. Do not widen the profile so a market slide looks larger.
Amplify starts at about three to five design partners, at least two, and expects revisions. Refusing 30 minutes for a demo is a non-money miss. Peter Reinhardt saw Segment’s pull in small companies on an open-source library, not in banks. Five customers independently loading data into a warehouse was the segment. Professor Miller had been excited about the classroom product. Asked by Paul Graham whether he would use it, he said he probably would not. Excitement without use is a favor.
Stackmatix calls the total market the ceiling and the ICP the front door. “All small businesses” is a wish, not a profile. King assumes customers and a revenue-operations team. At seed, use the calls you have.
A broad ICP shows up as two different sales
The profile is too broad when two accounts inside it need a different problem, a different alternative, a different signer, or a different onboarding.
- The homepage cannot say one thing.
- Wins and losses both fail to cluster.
- “And” joins industries that buy differently.
- A competitor could use the same sentence.
- Anyone who might benefit is in. Directive names that failure.
One or two deals are not a pattern, and one internal loss is not a market verdict. Clarify names both.
Rank industry, size, and role by what they change
Sort industry, then size, then role by what changes the sale.
| Cut | Add it when | Leave it out when | Seed default |
|---|---|---|---|
| Industry | Workflow, rule, or alternative differs | Pain and signer match | One vertical if the words differ |
| Size | Signer or cycle changes | Headcount means serious | Smallest band that can say yes |
| Role | User, champion, buyer, or blocker differ | One founder does all of it | Book who feels it and can commit |
| Geography | A works council, a public-procurement step, or a language the rest of the slice does not use | Added only to lengthen the list | One language, one way of buying |
| Stack | You need the tool, or it is the alternative | Logos are decoration | Required tool or competing tool |
Do not borrow a size band. One enterprise logo, or one team inside it, is an anomaly or a point solution. Denton’s Nested case, as HSBC reports it: people who had already failed to sell a home converted around five times more often. Copy the separator, not the multiple.
Put the bet and the refusals on one page
A seed ICP is finished when a teammate can accept or reject an account from one page, and the page names who is out.
- Situation: who, which problem, which circumstance, which alternative, which consequence they already feel.
- Trigger: name one event a stranger could recognize: a failed attempt, a new owner of the job, or a rule that just started to bind. Fresh means that event is still open.
- Findability: only the industry, size, place, or tool that passed the table. Name the community, list, or tool screen a stranger could open. If the only names are people who already know you, the profile is not findable yet.
- Roles: who you book, who uses it, who can block.
- Three to five refusals: no signer, procurement you cannot staff, a custom build, occasional pain, a cold trigger, no capacity to implement.
- Review every 6 to 8 weeks, and when pricing, the product, the stage, or a new capability changes. Stackmatix sets the interval. King adds the capability. Directive saves disqualifiers for mid-market. Put them on this page.
- Confirmation: the same problem, the same committing person, and a real next step, on more than one account that did not already know you. Validate an ICP before outreach is that test.
Denton wants each opportunity in or out. A maybe waits for the review. This page feeds the first go-to-market strategy. Positioning into campaigns and which experiments to run first start from that sentence.
Leave the profile unlocked in four cases
Do not freeze the ICP when the problem is still in your words, when the yeses are favors, when the segment matches a channel you already bought, or when two segments still pass. An unheard refusal is not a finding.
Denton tells pre-revenue founders to close what they can and get paying customers before a formal profile. Clarify’s pair is the same warning: too few points, and a page treated as permanent.
You earn the lock with Migicovsky’s questions: hardest part, last time, why it was hard, what they tried, what they dislike. Do not pitch, do not ask if they would use a feature, and do not out-talk them. Founders do this. Founder-led GTM before a full team comes before a hire sent to find the ICP from a slide.
I would write the one-page bet with you
I would sit in the conversations and leave one situation, the refusals, and one beachhead.
I am one person, Poldermarketing, fully remote, in Dutch and English. For a freshly funded startup anywhere, that is all-in: one seat for marketing, AI, and automation, from the first message to the first customers, without a separate specialist for each part. Hire me freelance for a project or a few days a week. I am strong in AI, content, automation, and building workflows. Google Ads and Meta Ads are newer for me: I set them up and review them, but do not scale a large media program.
The free growth scan shows how the site reads to a stranger. Positioning and messaging is where the sentence gets sharp enough to refuse an account. How I work is the shape of an engagement.
Questions people ask
How many interviews does a first ICP need?
Do not copy an interview quota from a blog. Migicovsky's bar is that you do not need thousands of people, and that one or two conversations about a real past event already beat a pitch. Amplify puts the first commercial test at about three to five design partners, and expects you to revise the profile. Stop when the same situation shows up on accounts that do not know each other, and when you have heard a refusal you believe.
What if there are no customers yet?
Write the hypothesis and keep closing what you can close. Denton, quoted by HSBC Innovation Banking, tells pre-revenue founders to get a batch of paying customers before a formal profile, because without demand you can hypothesise and you cannot measure. The hypothesis still names the problem, the alternative, the trigger, and who can say yes. It is not yet a wall. Replace a line when repeated behavior disagrees with it.
How do I choose if two segments both look good?
Run both through the same gates and do not average them into one sentence. If both pass, keep two hypotheses that differ by one variable, such as who can sign or what they use today, and hold the next conversations apart. Drop the one that fails a gate, or that only moves because someone likes you. This cycle gets one beachhead. The other waits on the review date.
Should a narrow ICP worry investors?
Say the beachhead and the later market on separate lines. A narrow profile is not a small company. Ahmadizadeh, in YC's first-time founder roundup, says defining the customer early and precisely would have saved time later, at the expense of growth during the batch. Letting anyone sign up left the target unclear and caused the product-market fit trouble. Do not widen the profile so a market slide looks larger.
How do I know early customers are favors?
A favor compliments the idea. A fit has already spent effort on a workaround, will try a rough version, or will introduce a peer. Professor Miller had been excited about the classroom product. Asked by Paul Graham whether he would use it, he said he probably would not. Excitement without use is a favor. Amplify warns that a friendly design partner without the pain does not teach you. Count the behavior only on someone who did not already know you.