Skip to content
From the blog

How to read your ICP from reply data

Analyzing cold email replies means sorting each sentence by the claim it argues with, counting the separate companies that made that claim, and editing the one line of your profile that count licenses. The words can take a condition out. Only a win you did not predict can put one in.

By Kshitij Maheshwari, co-founder · Updated August 2026 · 13 min read

The mapping

What each reply is actually evidence about

Five shapes cover almost everything a cold campaign gets back. Each argues with a different line of the one-page ICP, and each licenses a different size of edit.

The reply, roughly What it argues with Line of the profile What it can license
"We don't run that here." "That sits with our parent group." That the situation you assumed holds inside the company The situation An exclusion, once enough separate companies say it and you can see the condition before you send.
"We built it in a spreadsheet." "We already use Acme." What they cope with today, and who you actually compete with The workaround A rewritten workaround line and a message change. A profile change only where the workaround was the condition.
"At our size that's a five-minute job." Where the pain starts, and how far above it you have to be The threshold A candidate threshold to go and test. Never a direct edit on words alone.
Wrong stage, a holding company, the wrong business model entirely The purchasable attribute you used as a proxy for the situation The lookup key A filter fix this week. The cheapest and most reversible edit on this page.
"Not me, ask Priya." "Revisit in Q3." "Interesting, send more." Nothing about which companies can buy from you None Nothing. A misroute indicts the seat, a date belongs to your timing layer, and a compliment is not evidence.

Sources: the buckets, the lines they point at and the counts further down are how we run this read, not a published dataset. No dataset measures which reply predicts which profile edit. The one external finding on this page sits in the receipts section, with its source under it.


The framework

Replies can narrow your ICP. Only wins can widen it.

Both edits feel like learning. One costs an afternoon and undoes in a week. The other spends sending capacity you have not got.

Definition

A condition is a line of the profile a company either meets or does not, checked before you send: a size band, a stage, a tool they already run. Narrowing takes companies off the list. Widening puts them on, and every account added takes attention from the rest.

What the words can do
  • Name a condition you can check from outside, before you send
  • Cut a slice that was never going to buy
  • Fix a filter returning companies you never meant
  • Replace your words for the problem with theirs
What the words cannot do
  • Add a segment your list does not cover today
  • Move the condition that defines the whole list
  • Shift the threshold where the pain is supposed to start
  • Turn one enthusiastic stranger into a market

The asymmetry is not caution, it is what each edit can be checked against. An exclusion is verifiable: you can look at a company and see whether the condition holds before a single email goes out. An addition is a bet on companies you have never tried, paid for out of the sending you were already doing.

A reply can tell you who to stop writing to. Only a customer can tell you who to start.


The pass

The reading pass: five buckets and a quoted sentence

Twenty minutes, one sheet, five columns, no tool. The discipline is not the sorting. It is that the sentence goes in as the person wrote it.

The whole sheet
1
Column one

The sentence, quoted

Copy it out of the inbox, typos and all. A retold reply is your own opinion wearing somebody else's words, and it is the point where the evidence quietly dies.

2
Column two

The company, not the person

The company is the thing you count later, so it needs its own column. The name of the person who answered is useful for the thread and useless for the read.

3
Column three

The bucket, chosen on the claim

Not on the tone. A warm, apologetic no about a parent group is a situation finding. A cheerful "send more" is not a finding at all.

4
Column four

The line it points at

The situation, the workaround, the threshold, the lookup key, or none. This is the column that turns a folder of replies into a decision about a document.

5
Column five

The date it landed

File on the day. A week later the sentence has smoothed out, and you are reading your memory instead of your inbox.

The habit that makes it work

Write the five buckets down before you open the inbox. The bucket you have to add halfway through the month is the interesting one, and you can only notice it if the first five were on paper.

Operator note
Where evidence dies

The reply that changed who we targeted was shorter than this sentence, and I nearly did not write it down. By the time a reply has been retold twice in Slack it has become the thing we already believed, in their voice.

KM
Kshitij Maheshwari
Co-founder, Real Good GTM

The unit

Count companies, then check which companies answered

A month of sending buys you about ten companies' opinions. Small enough to read every sentence by hand, and far too small to slice by industry or size band.

10
Separate companies, one month

is how many said anything at all. Two hundred companies, two seats each, four hundred contacts, and a reply rate you take from last month.

Arithmetic, computed by us Reader-set inputs

400 contacts at 3% is 12 replies. Take out the accounts where both seats answered and you have 10 or 11 separate companies. Use your own rate; 3% is only there if you have never measured one.

Three counts, in order
One company, one answer

Two seats at one account is one

Sending to two people per company is what holds the company still, so a difference between the two answers is a seat difference. That read is a different job: it tells you which layer is actually broken.

Drop the talkative one

Then count the claim again

One company sends four detailed replies and carries a pattern on its own. Take it out and recount. If the pattern goes with it, you learned something about a company, not about a market.

Read the answerers

Four minutes in a spreadsheet

Line the companies that replied up against the ones you sent to, on the attributes your filter already used. If everyone who answered sits at the small end of your band, the pattern is about small companies.

None of this needs software. It needs a column with the company name in it, and the willingness to count to three twice.


The threshold

How much agreement licenses which edit

The count is the same for every edit on this list. What changes is what the count buys, and the cheap reversible edits come first.

The rule we run

The same claim from three separate companies, out of at least eight companies that answered. Under eight, you have met a few people, not read a market. Then drop the most talkative company and count again.

Cheapest edit first
The filter
3 companies, this week

Your list was returning the wrong state

Three companies turned out never to have been in the situation your filter implied. Fix the query, not the profile. You see the result in the next batch, and you can put it back.

The role string
3 companies, this week

The wrong seat opened the email

Three misroutes pointing at the same job title is a sourcing fix. Change the titles you pull, and leave the companies you target exactly where they are.

Their words
3 companies, this week

They named the workaround, and you did not

Three companies used the same phrase for the problem and it was not your phrase. Take theirs. A named workaround also changes what the offer has to beat.

An exclusion
3 companies, plus the gates

A structural fact you can see from outside

Three companies named the same condition and it is visible in your list data before you send. Take the candidate to the five gates an exclusion has to pass.

The condition itself
Words never license this

Moving what defines the list, or the threshold

Move either one and nothing after it compares to anything before it. Sentences buy you a candidate and a prediction here, never the edit.

The last row is the one founders reach for first, because the profile is the only document you can edit without asking anyone. It takes a written prediction tested in a second batch, not a good afternoon of reading.

Write the sentence that would make you delete a condition before the first email goes out. Ours reads: if three separate companies tell us this sits with their parent group, we add that exclusion. Written in advance, it survives your own enthusiasm.

Want the replies read properly, by the two people who sent them?

Book a Fit Check

The receipts

Why the pattern arrives early, and the rare one never pays

The usual instinct is to wait for more replies. The opposite is true: the claims that repeat show up almost immediately, and the rare one costs most of the sending.

The common advice

"Ten replies is not enough to conclude anything. Keep sending and look again next quarter."

  • Never says how many would be enough
  • Postpones a filter fix that costs one afternoon
  • Treats a rate and a sentence as the same evidence
What the counting shows

"Most of what people will ever tell you turns up in the first handful of them. The rest costs most of the sending."

  • Act on the cheap edits at three companies
  • Leave the expensive edit to a second batch
  • Rarity is arithmetic, not importance
Where the findings arrive
1
Early

Nine people carried 91% of it

A 2017 study did the count: 25 people answered at length and named 45 distinct things between them. One person named 53% of those, six named 84%, nine named 91%.

2
Late

The last four cost sixteen more people

The last four turned up only after the ninth person: 8.9% of the findings from 64% of the people. The rare finding is expensive by construction.

3
Deeper

Spotting and understanding are two budgets

Seeing the range of things people said took nine of them. Understanding them well enough to act on the meaning took sixteen to twenty-four.

Those were hour-long conversations and a two-line cold reply carries a fraction of one. So the counts transfer as a floor, not a translation: if nine long conversations were needed to see the range, ten replying companies is the least you should want before touching a document.

The percentages come from a 2017 study of when new findings stop appearing, read in full August 2026. Row two is our arithmetic on its figures.

Waiting for a bigger sample is not rigor when three companies have said the same thing. It is a quarter spent not fixing the filter.

Operator note
Antidote to eloquence

One articulate, detailed, slightly wrong reply moves a roadmap further than nine one-liners that agree with each other. Counting is the only defense I have found, and it works because it is boring.

RB
Rahul Bageria
Co-founder, Real Good GTM

How we would run it

One line of the profile, before and after

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

Three sentences, three companies
  • "Procurement for anything like this sits with our parent group."
  • "We can't sign that here, it goes through the group."
  • "Our holding company runs vendor selection now."

Three separate companies out of nine that answered, all arguing with the same line, and every one of them a subsidiary you could have spotted in the list data. The claim survives dropping the most talkative of the three.

Don't

File it as a note to yourself

Big companies have central procurement, probably avoid them

  • Not a line of any document
  • Undated, so it ages into a belief
  • Nothing could ever prove it wrong
Do

Rewrite the line, with its falsifier

Situation: excluded, subsidiaries whose parent runs vendor selection. Added 14 Aug 2026 on three separate companies. Falsifier: a subsidiary buys anyway, twice.

  • Sits on the line it argues with
  • Checkable before the next send
  • Dated, and it names what would undo it

One line moved. The size band, the stage and the threshold were not touched, so the next batch still compares to the last one.


The ops

Running the read with two people

One owner, one sheet, three passes. Only the monthly one is a meeting, and it is twenty minutes.

The three passes
Only one of them is a decision
On the day, five minutes
Quote the sentence, name the company, pick the bucket, stamp the date. No discussion.
Monthly, twenty minutes
Count companies per claim, drop the talkative one, check who answered, make the cheap edits.
Quarterly, one hour
Read the dated edits against what happened. Any edit with no falsifier gets one or gets deleted.
Key takeaways
4 points
  • 1 Quote the sentence. A scheme that needs a tool stops happening.
  • 2 Count companies. Two seats at one account is one answer.
  • 3 Three of eight answering buys the cheap edits, not the profile.
  • 4 Every edit carries its date and the sentence that would undo it.

The quarter's edits belong in one place. Ours go into the memo that holds what a campaign taught us, next to the rate readings they have nothing to do with.

Operator note
If we disagree

When the two of us file the same ten replies differently, we sharpen the bucket definitions instead of arguing about the replies. On your own: file them, wait a week, file them again, and see which ones moved.

KM
Kshitij Maheshwari
Co-founder, Real Good GTM

Failure modes

Where this read goes wrong

Four of these cost you a month. The fifth costs you the quarter, and it is the one that feels like insight.

The paraphrase

"They said the price was a problem" is a summary of your own worry. Their sentence said something narrower, and you no longer have it.

Counting replies

Twelve replies can be eight companies, and one of those eight can be four of the replies. Count the column with the company names in it.

Acting on the exotic reply

The most memorable reply of the month is the rarest thing you heard. Rare arrives late by construction, so being unusual is not evidence of mattering.

Two edits in one week

Change the profile and the message together and the next batch cannot tell you which one moved. Edit the message first; it is the one that undoes in a week.

!
Caution

Widening on one good reply from outside the profile

It is the only edit here you cannot cheaply undo. That reply is the most exciting email of the month, and one company is not a market. Add the segment and the same sending capacity is spread across twice the companies, with no readable result anywhere for a quarter.

Do this instead
Log the reply, write the prediction, test it in its own batch. The bar for adding a condition is a win you did not predict, twice.

FAQ

Questions founders ask

How do you analyze cold email replies?
Sort each one by the claim in your profile it argues with, not by how warm it sounded. Quote the sentence, name the company, date it, and mark which line of the profile it points at. Then count separate companies rather than replies. Twenty minutes a month, one sheet, five columns.
How many replies do you need before you change your ICP?
Count companies, not replies: the same claim from three separate companies, out of at least eight companies that answered. Two seats at one account is one answer, and the pattern has to survive dropping your most talkative company. Even then, words license the cheap reversible edits, not the condition that defines the list.
What do negative replies tell you about your ICP?
More than the positive ones. A refusal that names a structural fact, like procurement sitting with a parent group, is checkable from outside and can shrink your list safely. Enthusiasm cannot grow it, because a stranger being willing to answer a stranger says nothing about who can buy.
A company outside my ICP replied well. Should I widen the profile?
Not on a reply. The bar for adding a condition is a win you did not predict, twice, and a good reply never clears it. Log the sentence, write the prediction it implies, and give that segment its own batch. Widening on enthusiasm is the one edit you cannot cheaply undo.
How is this different from reading reply rate by segment?
A rate needs volume most seed teams never send, and it answers a different question. Rates tell you which slice responded better. The words tell you which claim in your profile is wrong, and they are readable in week one, from ten companies, by hand.
Can I use AI to sort my replies into the buckets?
For filing them into buckets you defined yourself, yes, and it turns an afternoon into minutes. It adds no companies to your count, and a confident summary of six replies is still six replies. Keep the quoted sentences in the sheet, because the summary is the thing you cannot audit later.
Kshitij Maheshwari, co-founder of Real Good GTM
About the author
Kshitij Maheshwari

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 pass he runs on a month of replies, where the sentence somebody actually typed decides which line of the profile gets touched.

Connect on LinkedIn

Keep going

The document, the test, and the memo

Where the lines live, how an expensive edit gets earned, and where the quarter's edits get written down.

Want the pipeline built, and the replies read?

Book a fit check. We'll look at the replies you already have, tell you which line of your profile they argue with, and say straight if outbound is not the right motion for you yet.

Book a Fit Check

No hard sell. No fake numbers. Real good work speaks for itself.