Aristotle wrote the cold email

Most people treat cold email as a writing problem. Hooks, curiosity subjects, a first line that proves you read the prospect's LinkedIn. Now a thousand AI tools produce all three.

In the first quarter of this year I sent 250,200 cold emails for one B2B account. About a hundred thousand more went out of a second account I'll come back to. The writing never moved the numbers.

The crucial aspect of a cold email campaign is not about the copywriting of a cold email but is about the call to action. And before the CTA, it's about who gets the email at all.

Start with the inbox. Radicati counts 392 billion emails a day in 2026. Yours lands somewhere in there.

The reader runs one check: is this for my job, in my industry, this quarter. It takes seconds, and it happens before anyone judges a sentence.

Belkins counts strict net-new outreach and reports 0.45% replies. Instantly, across a looser mix of senders, says 3.43%. Gong says the average rep sends 344 cold emails per meeting.

My first-quarter account sat at 1.29%. None of these is a writing score. They're the share of readers who passed the email through that check and found the offer worth answering.

Gary Halbert asked seminar rooms what single advantage they'd want before opening a hamburger stand. Not the recipe, not the location. A starving crowd.

Claude Hopkins wrote in 1923 that a headline's job is to filter out the people who aren't prospects. Two men who sold by mail, one conclusion: the list beats the letter. I got there with less style and more spreadsheets.

Takeaway: the published averages disagree by a factor of eight depending on what gets counted, so a benchmark is a story, not a target. Limits: none of the vendors share their definitions in full, and my two rows are one account each.

Reply rate steers the copy, opportunities decide the campaign. Rows: seven campaigns from my account A, January to March 2026. Bars: reply rate on sent. Right column: opportunities logged. The US campaign out-replied the UK agency campaign and produced 3 opportunities against 18. Limits: one account, one quarter, opportunity tagging is manual, "none listed" means none reported in my tracker, not zero.

I read reply rate every Monday because positive rate moves too slowly to steer on. I don't celebrate it. What paid the quarter was the campaign with the lower reply rate and the better segment.

Aristotle wrote the Rhetoric around 350 BC, as lecture notes for Athenians who had to persuade an assembly or a jury. He named three means of persuasion that live in the speech itself.

Ethos, the character of the speaker. Pathos, the frame of mind the audience is put in. Logos, the argument, or as he put it, the proof or apparent proof provided by the words themselves.

The sophists who taught before him had a fourth: kairos, the right moment. The reason this speech, to this audience, today.

Twenty-four centuries later, every cold email of mine that gets a reply has exactly those four parts, in that order. Why you and why now. Why trust me. What it is and how it works. What you get.

Nothing about the reader has changed since Athens. Only the room got smaller.

Four lines, one order, no exceptions. The mapping is mine. Aristotle named ethos, pathos and logos; kairos is the sophists' word and he treats it under timing. The lines in the fourth column are the templates I actually run.

The order is the part people miss, and Aristotle had it right. Open with ethos and you're bragging to a stranger. Open with logos and you're pitching. Open with pathos and you're a fortune teller.

Kairos first, always. Kairos is also what outbound tools try to fake with "I noticed you raised a round". The real one is plainer: I'm writing because you run this function in this industry, and the offer is built for this quarter. That sentence replaces the greeting.

Pathos is the one I got wrong, for years. I wrote it as a line about the prospect's pain. In August I cut it, because it read as filler and nobody wants their problem explained back to them.

But pathos didn't leave the email. It moved into the CTA. The one emotion a stranger will act on is curiosity about their own numbers: their invoices audited, their site rendered, their category's price per click. The ask carries the feeling, and that's why the CTA decides the campaign.

Everything cited in the rest of this piece is Aristotle rewritten with control groups. Authority with a named peer is ethos. The towel study is pathos with a percentage. Foot-in-the-door is kairos, measured. None of it is new. It just hadn't been tested.

The pain shows up once, as a contrast inside the logos, never as a diagnosis of the reader. Limits: one email, one campaign; I have no split test of this exact email against a version in a different order.

Relevance is built in layers, and every layer is a targeting decision. Industry decides which names and numbers you can show. Title decides the pain and the vocabulary.

Pain decides which step of the mechanism gets the weight, and what you give away. One generic layer and the whole email reads as a blast, no matter how well it's written.

So one ICP, one industry plus one title, gets its own campaign and its own email. Two titles in the same company never share one, because they don't share a pain.

I run 8 to 12 ICPs per account. Enough to learn fast, few enough to review every Monday.

Belkins sees the same thing from the outside. Campaigns to 50 or fewer recipients averaged 5.8% replies; campaigns to 500 or more, 2.1%. The teams at 15 to 25% aren't writing better emails. They're sending to different people.

The first three rows are targeting, the last four are copy, and the copy rows can only be as good as the targeting rows allow. This table is my framework, not a finding: the pass and fail conditions are the ones I enforce, and they're calibrated on B2B software and services campaigns.

The lists come from Apollo, Lusha and Hunter. They simply are LinkedIn big scrapers, but the data is there, who does what, paired with email guessing.

Title and company are copied from profiles and lag job changes by weeks. The email address is a guess from the company's naming pattern, pinged against the mail server. On a catch-all domain the ping proves nothing.

So I segment on their data and send only to verified addresses. At 2% bounces I stop. Google and Yahoo have held bulk senders to 0.3% spam complaints since February 2024, and bounces are the first warning.

Segment on what LinkedIn knows, distrust what the guesser adds. These are working rules from my accounts, not vendor documentation; the vendors describe their sources in general terms and don't publish match rates.

Ethos is where most cold emails die. Generic authority is no authority. "Trusted by leading companies" has been read four hundred times by everyone you email, and it says nothing.

The authority in the cold email needs numbers and recognizable names. Two names from the prospect's own industry, one exact figure. Non-round when the real number is non-round: 2,066% reads as true because it is.

The best evidence I know for why the names must come from the reader's world is about hotel towels. Goldstein, Cialdini and Griskevicius tested door hangers in a Phoenix hotel.

The standard "help save the environment" card got 35.1% of guests to reuse. A card saying most other guests reuse got 44.1%.

Then they went one step closer. Most guests who stayed in this room reuse: 49.3%, against 44.0% for the all-guests card and 37.2% for the environmental one.

The reference group did the work, and the closer it sat to the reader's own situation, the harder it pulled. A hotel group wants to see hotel groups. So my subject and authority lines carry a peer name, never the prospect's own company and never mine.

The closer the reference group, the stronger the pull. Bars: towel reuse rate by message in the second Goldstein, Cialdini and Griskevicius experiment (2008). Limits: one hotel, guests staying two or more nights, and a German replication by Bohner and Schlüter found the general-norm effect but a weaker, less reliable same-room effect.

Recognition works on its own. Goldstein and Gigerenzer showed that people asked which of two cities is bigger pick the one they've heard of. In an email it's simpler than that: a name the reader recognizes is a name they can check, which is why I can't fake it.

Numbers, same thing. Mason, Lee, Wiley and Ames found precise first offers read as better informed and pull smaller counteroffers. Xie and Kronrod found sharp numbers in ads read as more factual than round ones.

Hopkins had it in 1923. One razor maker advertised quick shaves, another a 78-second shave. The difference, he wrote, is vast.

Every claim carries a name or a figure, ideally both, and every figure was verified with the client before it shipped. Limits: these are lines from my accounts, chosen because they passed my own rule; I have no controlled test of each line against its vague twin.

Logos is the argument, and in a cold email the argument is a mechanism. A step by step of How it works is the best way to explain something in most situations. A prospect can picture four steps inside their own workflow. Nobody can picture "AI-powered automation".

A mechanism you can follow reads as true. A claim reads as marketing. Three to four steps, one action each, under twelve words, in their words: rate card, goods receipt, cost center, site sign-off. Then one value line with a number, strongest word last.

The evidence on concreteness holds up for a noisy field. Packard and Berger went through a thousand real customer service conversations. One standard deviation more concrete language: satisfaction up about 9%, later spending up at least 13%.

Their explanation is that customers take concrete words as proof someone listened. A prospect takes them the same way.

The 30 Minutes to President's Club report with Gong, 85 million cold emails, found pitch words cut replies by up to 57%. The emails that worked ran under 100 words, three or four sentences.

Boomerang's 40 million emails: third-grade reading level got 36% more responses than college level. None of these three studied B2B outbound. All three say the reader rewards language they can see.

Simpler sentences got more replies, and the drop at college level is a third of the response. Bars: response rate by reading level, Boomerang 2016, 40 million emails. Limits: general email between people who mostly know each other, not cold outbound; the high school bar is derived from Boomerang's "17% higher than high school" statement, not reported directly.

There's a second reason I keep four elements and cut the rest. Weaver, Garcia and Schwarz call it the presenter's paradox. Readers average what they're shown, so a mildly good point next to a strong one drags the total down.

In August I removed pathos from my own framework for that reason. Every "you're probably struggling with X" line was filler that watered down the names and numbers next to it.

Do not school the prospect, do not congratulate the prospect, do not teach the prospect, do not tell the prospect about their problem. Just why we are contacting them, put the authority point about ourselves, put what we do, call to action.

Pathos is the frame of mind you leave the reader in. In an email it's one line, the last one, and it's an offer. Ideally the positive reply unlocks a mini action that still creates a value. The syntax is fixed, "Would you mind if I share...". What gets shared is the whole campaign.

An audit of their data. A sample built from their public material. A benchmark of their peer group. The template their peers use. The closer to their own numbers, the better.

Avoid asking for a call and definitely do not share the calendar booking link at the first cold email.

The outside data is loudest here. Gong Labs studied 304,174 sales emails. On a cold first touch, asking for interest booked about twice the meetings of any other CTA.

The "day and time" ask only works inside a live deal, where it goes from 15% to 37%.

The 85-million report ranks the offer above the interest question. Offer something concrete, plus 28% on replies. Ask for a meeting, minus 44%.

"Quick chat?" scored minus 2%. So nothing.

Give something and the reply rate rises; ask for time and it halves. Bars: reported impact of CTA type on reply rate in the 30MPC and Gong analysis of 85 million cold emails, as tabulated by Real Good GTM from the report's charts. Limits: vendor data, no confidence intervals published, and "offer" mixes free trials, reports and audits.

The psychology is older than email. Freedman and Fraser asked Palo Alto homeowners in 1966 to put a big "Drive Carefully" sign on the lawn.

Cold, 17% said yes. After a small sticker first, 76%.

Regan's 1971 study: an unrequested Coke roughly doubled the raffle tickets people bought from the giver, liked or not. Burger's team later found the effect gone if the ask came a week after the gift.

That last one is an operating rule. When the prospect writes "sure, send it", the asset goes out the same day, nothing bolted on. The next ask is one rung up.

The closer the asset sits to the prospect's own numbers, the higher the intent of the yes. The ranking is my working order, supported in direction by the CTA data above but not measured rung by rung.

Subjects need numbers and name dropping. Every subject I run is an impersonal trade headline about the prospect's peers. "How GXO and 37 other 3PLs catch carrier overbilling before it gets paid."

Never the target, never the sender, never the offer, never "you". Until August my default was the opposite: a bare "Us x Them", the kind of subject that looks like an internal thread.

I switched after watching the headline subjects run across four accounts. That's one operator's judgment, not a measured A/B.

The published data doesn't settle it, and I'd rather show that than pretend. Backlinko's 12 million outreach emails: subjects of 36 to 50 characters got 24.6% more responses than short ones. Good for the headline.

The 85-million report says the opposite. Under four words, lowercase, looks like internal email. Numbers and social proof hurt opens. Boomerang says three to four words.

The corpora aren't comparable. Backlinko is link-building outreach, Gong is sales, Boomerang is people who mostly know each other. They agree on one thing: the empty or gimmick subject loses replies even when it wins opens.

So I judge subjects on replies, never opens, and keep the headline until my own numbers say otherwise.

The evidence splits on length and agrees on specificity. Limits: three of four are vendor datasets measured mostly on opens, and opens are inflated, see below. My rule is a hypothesis under test.

Opens are close to worthless anyway. Apple Mail has pre-fetched tracking pixels since September 2021, and Apple clients carried 48.6% of all opens just before that.

My first-quarter account reported an 80.06% open rate. Nobody should steer on that.

Follow-ups are the least controversial part. Woodpecker's 20 million emails: 4 to 7 touches got 27% replies, 1 to 3 got 9%. Instantly splits replies 58% on the first email, 42% on the follow-ups.

The 85-million report lands on about six emails over 14 to 28 days. Breakup language on the last one nearly doubles that email's replies.

My sequence is the opener plus five follow-ups over 21 days, three threads, alternating. A reply in the same thread adds one fact about the same asset. A new thread gets a new subject, a new peer name and a new asset, a second first impression for someone who ignored the first.

Never "just checking in". A call shows up no earlier than the fourth touch, and only next to the asset, never instead of it.

Six touches, three threads, three assets, one referral ask at the end. Days and thread alternation are my defaults; the six-touch count and the 14 to 28 day window match the 85-million report, the alternation is my own and untested against a single-thread control.

Every campaign is an experiment. Measure on delivered, never on opens. Change one variable a week, or you learn nothing.

The thresholds below are my starting defaults. They get recalibrated after four weeks of a new account's data, and they'll be wrong for some verticals.

Belkins shows why. Founders replied at 0.57%, C-level at 0.42%. Companies of 11 to 50 people at 0.49%, enterprises above 10,000 staff at 0.22%. A 1% positive rate is a good week in one segment and a disaster in the next.

One signal, one change, one week. These thresholds are hypotheses calibrated on my accounts, not findings; the 2% bounce line is the only one with outside support, from the Instantly and Google Postmaster guidance on list health.

Everything above assumes one email per ICP. AI changes the unit: one email per prospect, written from that prospect's data, at the cost of a template.

It works on one condition. The data going in has to be good. That's the whole rule, and it's the part most people skip.

Good means verified. The title is current. The company facts come from their site, not from a guess.

The proof points are the client's real names and numbers, and the model can't draw on anything else.

Feed it a scraped title that's six months stale and it writes a beautiful email to someone who left. Feed it a vague product description and it writes "AI-powered automation" a thousand different ways.

The relevance layers don't go away. They move from the campaign to the row. Industry, title, pain, peer names and asset still have to be true for that one person, and now nobody proofreads.

Lemkin's complaint about AI outbound is that every email sounds the same. He's right about the ones built on bad inputs: same structure, same generic observation, same "I noticed that". Give the model a verified fact and a real peer name and the output stops sounding like the other thousand.

So the AI budget goes to data before it goes to writing. Verification, enrichment, a proof inventory the model is locked to.

The copy was never the bottleneck. Now it's free, and the bottleneck is in plain sight.

The economics come first. Outbound is a cost per meeting, and a meeting has to be worth it.

Glencoco's 2026 SDR numbers put a fully loaded cost per qualified meeting at $917 to $1,800 for a ramped rep. Count ramp and turnover and it's closer to $1,500 to $2,500.

My first-quarter account: $30,752 of outbound cost, 23 meetings, $1,337 each, against $110,703 of delivered revenue. That works.

The second account didn't. 93,721 emails to 42,141 prospects for an AI fashion-photography SaaS, self-serve, low ticket. 0.42% replies, nothing closed.

The copy wasn't the problem. An enterprise motion was pointed at a product whose deal size couldn't pay for one meeting.

Jason Lemkin has said for years that outbound is hard at a $2,000 ACV and comfortable above $50,000. Check that before any table in this piece.

The same method, two accounts, one works and one doesn't, and the difference is the product's deal size, not the email. Limits: two accounts of mine, revenue attribution done by hand, and the Glencoco range is a model built on Bridge Group and Xactly tenure data, not a measured average.

Peer proof can backfire. Schultz and colleagues told households their neighborhood's average consumption. Heavy users cut back, light users went up.

They call it the boomerang effect. Only an added sign of approval stopped it.

In outbound the boomerang is the market leader. Tell the biggest 3PL in Europe what 37 smaller 3PLs do and you've told them who you think their peers are.

The towel effect is also smaller than its fame. Bohner and Schlüter's German replication found the general-norm advantage but a weaker, less reliable same-room effect. I use peer names because they work in my accounts, not because a paper guarantees them.

Precision can backfire too. Pena-Marin and Bhargave found a precise estimate that turns out wrong hurts trust more than a round one would have. Wadhwa and Zhang found round numbers "feel right" when the judgment is emotional.

An exact number I can't defend is worse than none. That's what the proof inventory is for.

Some of my rules run against Gong's data. "Hope all is well" raised their meetings booked by 24%. I ban greetings.

The 85-million report says "you" beats "we", and my emails open with "I'm Luca".

Packard, Moore and McFerran found service agents saying "I" outperform those saying "we", which is some comfort. But the no-greeting rule is a habit, not a measurement. I'd drop it the week a test showed it costing replies.

The personalization camp has a point and a problem. Backlinko found personalized bodies lifted response 32.7%.

Lemkin wants every SDR email opening with "I noticed that..." banned as fake observation. Both are right.

John, Kim and Barasz found ads that revealed inferred or third-party data cut purchase interest by about 24%. Van Doorn and Hoekstra call it the personalization paradox: fit raises intent, intrusiveness kills it, and the two rise together.

My answer is two merge tags, first name and company, and relevance built into the segment rather than the sentence. Minimize the syntax like {{industry}}, they make the email feel unrealistic and fake.

Then the law. In the UK, PECR lets a business email corporate subscribers without consent; sole traders and partnerships count as individuals and need it.

Across the EU the rules are national. A segment you can't email lawfully comes out at list build, not on a complaint.

And reciprocity wears out when everyone gives gifts. Kunz and Woolcott got 20% of strangers to answer a Christmas card in 1974. A 2015 replication got 2%.

Every free audit I offer is a little less surprising than the last. The AI tools now filling every inbox with the same four-step structure will speed that up.

Three rules stand on outside data, three are borrowed from adjacent fields, three are my habits. That's what a method built by one operator looks like, and it's why the weekly loop exists.

One more thing, and it's the reason I still do this.

Email runs on an open protocol. SMTP was written down in 1982. Anyone with a domain can send a message to anyone with an address, and nobody sits in the middle charging for the introduction.

Look at the alternatives. Google Ads: you pay Google to show something you have to people who are looking for it. Meta Ads: you pay Meta to show a banner to people its systems have tagged as interested, from what they like, who they call on WhatsApp and what they linger on in Instagram.

Both work. Both charge rent, and the rent goes up every year.

Cold email needs none of that. A LinkedIn scrape, an email guess, and an open protocol. The email protocol is open, and that allows the last act of freedom.

Two channels rent an audience from a platform. One builds it and speaks to it directly. Limits: the characterization is mine; the CPC figure is one vendor's analysis of B2B SaaS search terms, not a market average.

The lists are noisy, the guesses bounce, and the inbox providers push back with complaint thresholds. Fine. It's still the only channel where a stranger can reach a decision maker without paying a platform for the privilege, and where the decision maker can answer without one either.

That's also why the rules in this piece matter. Every lazy blast spends a little of the commons. Every relevant email, with a real name, a real number and something free at the end, buys some back.

So the next time you receive a cold email, be thankful to the last piece of freedom, and reply. Read and reply, for that's the most free act you can do.

1. Write the four lines in Aristotle's order. Kairos, ethos, logos, pathos: why you, why me, how it works, what you get. Never open with the pitch or the credentials.
2. Decide relevance before you write. One industry, one title, one pain, one campaign. 8 to 12 live, reviewed weekly.
3. Apollo, Lusha and Hunter are LinkedIn scrapers with a guesser attached. Segment on their data, send only to verified addresses, stop at 2% bounce.
4. Proof is two names from their industry and one exact figure. Non-round when the real number is. Verified or it doesn't ship.
5. Explain the product as How it works. Three to four steps in their words, one value line with a number. Cut every line that describes their problem back to them.
6. The CTA is a free asset built for this ICP, never a meeting on touch one. "Would you mind if I share..." When they say yes, send it the same day.
7. Six touches, three threads, 21 days. Each bump adds a fact, each new thread adds a name and an asset, the last one asks for the right person.
8. AI writes one email per prospect. It works only when the data underneath is verified: current title, real company facts, the client's own proof points.
9. Measure replies on delivered, ignore opens, change one variable a week. And check the deal size first. Below a few thousand a year, the method has nothing to work with.
10. Email is the last open channel. Send relevant, and when a good one lands in your inbox, read it and reply.

The method buys a 1 to 3% reply rate in a segment that can afford it. That's the ceiling, and it gets lower every year as the inboxes fill with the same structure from the same tools. What it never buys is a product people didn't want.