From the blog

ICP vs buyer persona: which one is broken

An ICP is a claim about which companies can buy from you. A buyer persona is a claim about which person inside them will answer, and why. They are two separate bets riding on the same send, they fail independently, and the reply that tells you which one failed is not the one most founders read.

By Rahul Bageria, co-founder · Updated August 2026 · 12 min read

Share

The spec

The two objects, as an operating spec

Not what each one is, which every page already tells you. What each is made of, what moves it, and what breaks when it is wrong.

Property	Ideal customer profile	Buyer persona
What it is made of	Conditions that have to hold inside a company.	A seat, and the one problem sentence it answers.
Built from what evidence	Customers you closed, deals you lost, where support piles up.	Replies, call recordings, job posts, the public org shape.
The unit of evidence	An account.	A conversation.
Where it enters the campaign	The account filters that decide who is on the list.	The title filter, and the first line.
Who should own it	Whoever owns the market decision.	Whoever reads the replies.
What makes it go stale	Something moving on your side: price, product, who started closing.	Something moving on theirs: a title dies, a function centralizes.
What breaks when it is wrong	The whole list, and a quarter of sending capacity with it.	The message. The list underneath it is still fine.
The failure tell	Silence, plus the occasional flat no from several desks.	Activity that goes nowhere: forwards, redirects, polite dead ends.

Sources: the two filter rows come from the tools' own documentation, checked August 2026: LinkedIn Sales Navigator's lead and account filter definitions, Clay's list-builder sources, and Apollo's People and Organization endpoints. No dataset measures how these two objects fail; the rest of the table is how we run the motion.

Two objects that came from two different places

The confusion has a cause. One term has a named author and a date. The other has neither.

Definition

A buyer persona is an interaction-design artifact, a described stand-in for a user, proposed by Alan Cooper in The Inmates Are Running the Asylum in 1999. An ideal customer profile is an ordinary business phrase that accreted a job, and nobody coined it. Both are in our GTM glossary.

In the campaign

Where each one actually enters a campaign

Whatever the strategy document says, both claims get made in the same two places: a saved search and a first line.

The step	What the ICP decides	What the persona decides
Building the account list	Everything. Which companies enter at all.	Nothing. The account table holds no person fields.
Sourcing contacts inside them	Nothing. The companies are already picked.	Which seats you pull, and how many per account.
Writing the first line	The situation you are allowed to assume is true.	Whose week the problem ruins, in their words.
Reading what comes back	Whether that account could ever sign.	Whether that seat could ever care.

None of that split is our convention. Sales Navigator ships two filter tables, Lead and Account, with no person attribute in the account one. Clay has Find Companies and Find People. Apollo separates People from Organization endpoints.

The same page defines the tool's own Persona object as function, seniority, job title and geography: four fields, all filters. It also describes the seniority filter, a person-level control, as a way to find leads matching your ideal customer profile.

The diagnostic

Which layer is broken, and how the replies tell you

Start with what cannot help you. Silence is not a verdict on either document, and more volume never makes it one.

A slice that returned nothing might have held the wrong companies, the wrong seats, or neither: the mail may never have arrived. Those three look identical on a dashboard. Read the replies instead.

Four reply signatures

1

Company

Two desks at the same account say the same no

"We don't do that" from an ops lead and a finance lead at one company is the account answering, not a person. The condition you assumed does not hold there. Change the fit condition, not the title.

2

Person

"Not me, talk to X", repeated across accounts

The only unambiguous verdict outbound hands you. A company that cannot use you does not know who inside it would own the problem, so a misroute cannot be a company error. Change the role filter, keep the list.

3

Person, solved

One seat answers and the others do not

Same companies, different desks, one pattern of replies. The company is held still, so the difference has to be the seat. You found the persona rather than lost it. Open with that seat.

4

Company

Warm interest that never reaches a decision

Everyone is friendly, the calls happen, and nobody can name what would have to be true to sign. Interest is not fit. Ask whether any account in this slice could ever buy.

There is a third branch, and it is the most common. Right company, right seat, wrong moment or wrong ask. Its signature is a positive reply that dies after the yes, and neither document is at fault.

Silence is where the two failures look identical. The replies are where they separate.

The instrument

Why one contact per account cannot answer the question

Every signature in that ledger needed more than one seat at the same company. One contact per account welds the two claims into one variable.

One person per company, one result. A no is consistent with a wrong company, a wrong seat, or both, and no arithmetic separates them later.

The mechanism

Contact two or three seats at the same accounts and the company is held constant. A difference between seats cannot then be a company effect, and one answer from every seat cannot be a seat effect. It works at fifty accounts.

Multithreading is sold as a way to lift response rates. It is also the only instrument a seed team has for separating these two failures. The mechanics are in our multithreading play.

Operator note

The first question

When a founder tells me their ICP is wrong, my first question is how many people they emailed per company. If the answer is one, they do not have an ICP problem yet. They have a campaign that cannot answer the question.

RB

Rahul Bageria

Co-founder, Real Good GTM

How we would run it

A worked diagnosis: one campaign, read twice

An illustrative walkthrough of the method, not a specific client result. We report real numbers only when they are real.

1

Week three · The result

What the campaign returned

One slice, two seats sourced at every account, three weeks of sending. A handful of replies and a lot of nothing. No meetings.
2

The instinct · The wrong read

"The segment is too narrow"

Widening feels like progress. It doubles the accounts, halves the density, and makes the next result harder to read. Nobody has opened the replies yet.
3

The read · By layer

The same replies, sorted

Most of them are one sentence from ops leads: not my call, ask the finance side. Different companies, same redirect. Signature two, and it indicts the seat.
4

Week four · The repair

What the read changed

Same account list. New title filter, one rewritten problem sentence for the seat that owns the budget. The profile was never touched, and the quarter was not spent proving it was fine.

Want a second pair of eyes on a campaign that is not working?

Book a Fit Check

What we do

What to do once you know which one is broken

Three branches, three repairs. Move one object at a time, so the next result is still readable.

1

The persona is wrong: change the seat, hold the list

Leave the account list exactly as it is. Swap the title filter, rewrite the problem sentence for the new seat, send to the same companies. A week of work, and the result reads directly against the last run.
2

The ICP is wrong: change the list, hold the seat

Keep the seat and the copy. Move the fit condition, rebuild the account list against it, send the same message. Changing both at once buys you a new number and no way to attribute it.
3

Neither is wrong: it is the ask or the moment

Right company, right seat, an ask nobody can afford to say yes to. Leave both documents alone and rewrite what you are asking for.

Operator note

Cheaper repair first

When the read is ambiguous, change the seat. A wrong seat costs a rewrite and a re-source. A wrongly widened profile costs the quarter and the readability of everything inside it, and nobody notices until it is over.

KM

Kshitij Maheshwari

Co-founder, Real Good GTM

What collapsing them into one document costs

One document holding both claims fails in two directions. It describes nobody, or it describes everybody.

The common advice

"Write one ICP document, with the persona inside it, so the whole team works off a single source of truth."

✕Every attribute added multiplies against the last one
✕Nobody ever counts how many real accounts still match
✕When it fails you cannot tell which half failed

What actually works

"Two documents, each short enough to be wrong on its own, each tested by a different part of the campaign."

✓The conditions get checked against the wins and losses
✓The seat gets checked against the replies, weekly
✓A failure names the half you have to change

The opposite failure is more common at seed. The document softens every time it excludes an account somebody liked, until it describes any B2B company with a website. Both versions are unfalsifiable, so neither gets thrown out.

A merged document also cannot be entered into any tool, because the tools take two filter sets. Whoever builds the list splits it by guessing, and the guess becomes your targeting.

The measured arithmetic behind the left column, and what a persona built from evidence contains, is in our guide to buyer personas.

Two clocks

Two clocks, and two different reasons

Both objects go stale, and not for the same reason, so one refresh habit cannot catch both.

The ICP clock

•It moves when your side moves, never on a calendar
•A price change can put half the band out of reach
•A shipped feature makes an unservable segment servable
•The tell: a customer type starts closing that the document says is out

The persona clock

•It moves when their side moves, and you hear about it late
•A title stops existing after somebody reorganizes
•A function gets centralized into one shared team
•The tell: your best seat starts bouncing across a whole segment

How fast each layer of a list rots underneath you is in our lead list building guide. What matters here is the cause: two clocks started by different events cannot share one review.

Which to build first, and the one exception

Company first, and not because it matters more. Because of where the evidence comes from.

The build order

The evidence supply decides the order

Start here

The company claim. You can draft one tonight from customers you already have, deals you lost, and where the support load piles up.

Then this

The person claim. It can only be learned from conversations, and conversations only exist after you have picked companies and sent something.

The exception

A horizontal product exactly one role ever buys. The seat is doing the fit work, and the company filter is only a proxy for whether that role exists there in numbers.

Writing those conditions down, testing them against real sends and knowing when to revise them is our ideal customer profile guide.

Key takeaways

4 points

1 The ICP is the company bet. The persona is the person bet.
2 Silence indicts neither of them. Only replies carry a verdict.
3 Two seats at one account is what makes the comparison readable.
4 Repair the cheaper object first. A widened profile costs a quarter.

Failure modes

Getting the diagnosis wrong

Four misreads that all look like decisive action at the time, and one repair that costs more than the rest.

Reading silence as a verdict

A slice that returned nothing has told you about your coverage and your delivery. It has told you nothing about either document, and editing one on that basis is guessing with extra steps.

Reading a misroute as rejection

"Talk to ops, not me" gets filed with the no's. It is the most useful reply in the folder: a correct company handing you the correct seat, for free, in week one.

Rewriting the persona when the ask was wrong

Replies but no meetings usually means the seat is right and the ask is too expensive. Rewriting the persona there throws away the one part that was working.

One seat per account, then expecting to learn

Two claims, one variable, one number. The result stays unreadable no matter how many accounts you add, and adding accounts is what everyone does next.

!

Caution

Widening the ICP to fix a persona problem

The one repair you cannot cheaply undo. The same sending capacity spreads across twice the market, every result afterwards stops being readable, and the person who widened it is never the one who notices.

Do this instead

Change the title filter first. It is a week, and it is reversible.

FAQ

Questions founders ask

What is the difference between an ICP and a buyer persona?

An ICP is a claim about which companies can buy from you. A buyer persona is a claim about which seat inside them will answer, and why. The practical difference is that they are built from different evidence and they break independently, so a campaign that is not working has usually failed at one of them and not both.

How do I know whether my ICP or my persona is wrong?

Read the replies, not the response rate. The same refusal from two different desks at one company indicts the company. A repeated "not me, talk to X" across different companies indicts the seat, and it also hands you the right one. You need more than one contact per account for either read to work.

What if neither the ICP test nor the persona test gets a reply?

Then neither test has run: all you have tested is delivery. Check the mechanical causes first: bounces, catch-alls, whether the mail is landing anywhere at all. Our guide to turning outbound into GTM learning covers that read in full. After that, check whether anything has actually changed at those companies recently, because a correct list with no timing layer produces exactly this.

Can I keep both in one document?

You can write it, but you cannot run it. Every list tool takes two separate filter sets, one for companies and one for people, so somebody splits your document at build time by guessing. That guess becomes your real targeting, and nobody can point at the moment it diverged from the plan.

Which should I build first?

The company one, because of where the evidence lives. A first profile can be drafted from customers and losses you already have. A persona can only be learned from conversations, and conversations start after you have picked companies and sent something to them.

How many personas should one campaign have?

As many as you can write a genuinely different problem sentence for. Two title strings running the same body copy is one persona wearing two hats, and the result it produces cannot be read either way. If the sentence does not change, the seat has not changed.

About the author

Rahul Bageria

Co-founder of Real Good GTM. He has been the first business hire and Chief of Staff at seed-stage B2B startups, building outbound pipeline before any playbook existed. This post is the read he runs when a campaign is three weeks old and going nowhere, and somebody has to say which of the two documents to change first.

Connect on LinkedIn

Keep going