A Filter Is Not an ICP: Deriving the Real One From the People Who Bought

Forwarded to you? I am Heath. I build go-to-market systems and put AI to work in sales, the right way, then I write down exactly what I built, what broke, and what it moved. One story per week, receipts only. This one is about the ICP that was never a filter. It was the people who already bought.

The reflex: describe the account, hand the reps a filter

Ask most teams for their ICP and you get a filter. 50 to 500 employees, B2B SaaS, uses a CRM. It feels like an answer. It is a targeting list, and a list is not an ICP.

I did exactly that. I wrote the ICP as a firmographic filter, dropped it into the targeting, and told the team to work it. It failed the same way every time. The pipeline filled with accounts that matched the filter perfectly and never bought. They looked right on paper, so nobody could say why they were stalling, only that they were. The filter described ten thousand companies and told us nothing about the one person who actually says yes.

The block was never targeting. It was that the filter answered the wrong question. It said what a company looks like. It never said who the buyer is, what their day looks like, what environment they are selling in, or what problem in their actual life we remove. You cannot guess that in a positioning workshop. It only exists in the people who already bought.

The AE feels this as "these accounts fit the ICP and go dark." The CSM feels it as "these renewals looked healthy on the firmographics and churned anyway." The marketer feels it as "we targeted the segment and the message did not land." Same reflex every time: a description of the company, mistaken for an understanding of the buyer.

The reframe: derive the ICP from the people who bought

An ICP is not a filter you declare. It is a receipt you derive. The move is not to describe a company size band. It is to take the loud, useless firmographic and drill it down to the one specific person, in one specific environment, who hired you for one specific problem, in the exact words they used.

A firmographic filter is not an ICP. It is a list. The real one is the person who already bought.

Same move, other seats. A CSM drills "this account fits the profile" down to the human: which role renews, which day-to-day pain we removed, whether the champion is the economic buyer or a single fan. A marketer drills "this segment is our ICP" down to the words: which problem the buyer names first, in their language, not the deck's. The firmographic is never the thing you act on. The person underneath it is.

How the best teams frame it

I am not the first person to argue that the ICP lives in the customers, not the workshop. The operators who have built this at scale mostly agree on where the leverage is, and it is not the size band. It is the person and their language.

SOURCE

Lenny Rachitsky, "How to identify your ideal customer profile (ICP)"

What it argues. Do not invent the ICP. Watch who gets most excited about the product, which roles implement it successfully, what size actually fits, and get comically narrow, at least three narrowing characteristics. The signal that matters comes from behavior, not from the leads your investors send you.

My take. Agree. The narrowing is the whole point. A filter is broad on purpose so nothing gets excluded, which is exactly why it never describes a buyer.

SOURCE

April Dunford, "Obviously Awesome" (positioning starts with your best customers)

What it argues. Positioning starts from a short list of your best customers, the ones who understood the product fast, bought fast, and became fans. The pattern across that list is the context that makes the product obvious. You reverse-engineer the ICP from who already won with you, not from a segment you wish you owned.

My take. Extend. Best customers reveal the ICP, but you have to mine the words they use, not just the firmographics they share. The language is the part that carries into the persona.

SOURCE

Leah Tharin, "Finding your ideal customer profile in B2B: interview questions and guide"

What it argues. The ICP comes out of structured interviews with real customers, and the questions are the craft. Why they came, why they bought, why they stayed. Build it from evidence, in the customer's words, instead of a room full of internal assumptions.

My take. Agree hard. The interview is where the language lives. My first ICP failed because it was assembled in a room. This one held because it was pulled from the calls.

Even a curator has to concede when the field agrees: nobody who has actually built an ICP thinks it comes out of a workshop. It comes out of the people who bought.

The method: Solve, Stack, Split

SOLVE THE CRUX

What is the real problem, framed as work and not a headcount band?

The problem is not "we need an ICP doc." It is "our ICP is a guess, so the pipeline fills with lookalikes that never buy." So the first work is going to the source: interview the people who already bought and read the pattern that repeats. For the AE that is who the deal actually came from, for the CSM who actually renews, for the marketer whose problem the message actually lands on.

STACK THE CONTEXT

What tech and signals turn raw evidence into one defensible profile?

Not a shopping trip. The call transcripts live in Gong, for the words buyers actually use. G2 reviews get mined for voice of the customer, praise and complaint. Twelve months of closed-won sit in Salesforce, backtested for the pattern that repeats. Amplitude shows how the won accounts actually use the product. Claude reconciles the inputs into one profile and writes the personas, so the ICP survives scrutiny instead of being one loud opinion.

SPLIT · CUT THE DRAG

What low-judgment work goes to the system?

Mining G2 and transcripts for language, backtesting twelve months of closed-won, enriching hundreds of converter titles, reconciling the divergences against a real cohort. This is the part that never scaled when a person did it by hand, and the part AI does perfectly because it is mechanical.

SPLIT · KEEP THE JUDGMENT

What stays human?

The interviews and the personas. Talking to a real buyer, hearing the problem in their words, and writing a persona a rep can picture, with the day, the environment, the problem you remove, and the hard disqualifiers. The system finds the pattern. The operator decides what it means and writes the human.

The workflow: the board that runs it

Solve, Stack, Split is the shape. Here is the actual board, lane by lane: what the agents run, what stays human, and the tool at each step. Once it is wired, refreshing the ICP is a rerun, not a quarter-long project.

What it moved

At a company I was at, a product-led SaaS with a big free base and a sales team layered on top, this is the exact build I ran. Interview the buyers, mine their words, backtest the wins, reconcile it into one profile, then write personas about the person and not the headcount band.

The reps did not target harder. They targeted the person the evidence pointed to, with the problem named in the buyer's own words. Across 424 converter titles, 248 enriched clean, and the titles differed while the motion stayed one. That is what "one ICP, two motions" means: the person and the job underneath were the same, so a single profile carried both.

Both of my receipts are revenue-side. Drop your own workflow in. A CSM turns the same build into a retention profile: interview the accounts that renewed, mine why, and write the persona that predicts a save. A marketer turns it into a messaging brief: the exact words buyers use for the problem, pulled from the calls, become the headline instead of the internal jargon.

WHAT I LEARNED

1. A firmographic filter is not an ICP. It is a list. The real one is a person with a problem.

2. The language is the asset. Mine G2 and the calls for the exact words buyers use, not the words your marketing wishes they used.

3. Backtest closed-won for the pattern that repeats, not the logos you wish you had. Reconcile it against a real cohort so it survives scrutiny.

4. One ICP can carry two motions when the person and the job underneath are the same. 248 of 424 titles differed and still bought one thing.

Run this one this week

Do not write the whole ICP. Pull your last ten closed-won accounts. Read the calls or the G2 reviews for the exact words they used for the problem, not the words in your deck. Write one persona: the person, their day, the problem you removed, and one hard disqualifier. Then hold it up against your live pipeline and see how many "perfect fit" accounts it quietly disqualifies. That is the ICP in miniature, and it tells you whether the full build is worth a quarter before you spend one.

Two builds that sit next to this one:

• Scoring the TAM · "Put the entire TAM through one rubric and make reps work the ranked list top-down." The ICP is the fit input that rubric scores against.
• The Win/Loss Backtest · "The signals that really separated the deals we won and lost were behavioral, not firmographic." The proof that a size band was never the ICP.

This is one build from the Build Log. Every week I take one sales or revenue problem, run it through the loop, and show the receipts. If someone forwarded this, the subscribe button is right below. Keep building. Heath.