A cold email isn't competing for attention. It's passing a half-second sorting test.
Subject lines · personalization · leading with value · sequences · A/B testing·4 moves · 4 diagrams
Nobody reads your email. They sort it first. The entire job of the message is to land in the "written to me" pile — not to be more persuasive.
A cold email has to clear four gates in a row: get sorted as "written to me," be believed when it claims you actually looked, feel cheap to answer, and still exist during the week they happen to have room for this. One move per gate. One, the subject line only does sorting — never curiosity. Two, personalization is worth what faking it would cost, and merge fields cost zero. Three, the first message converts on a reply, not a meeting — shrink the ask to ten seconds. Four, a sequence isn't persistence, it's sampling — new angle each time, and write the ending before you start.
MOVE 01
A subject line doesn't attract. It gets you sorted.
Its only job is to land in the right pile.
self-referencesorting, not readingno curiosity gaps
The principle in one line
What people do in an inbox isn't reading — it's a half-second sorting task: is this written to me, or sent to a batch? So a subject line's job isn't to be interesting or intriguing, it's to land in the first pile, and the reliable way there is one concrete fact that is true of this person and nobody else: something they said last week, the internal name of their project. Curiosity-gap lines like "quick question" buy an open and no reply, and they burn your next message along the way.
Same person, four subject lines, four verdicts
Worked example
Setup: you do freelance audio editing and want to talk to an independent podcaster.
✗ The subject line is about you
Subject: "Professional podcast post-production services." That sentence is true for every podcaster alive, which is exactly why it isn't true for him — what he sees while sorting is "a batch," not "me."
✓ The subject line is about what he just said
Subject: "the 4 hours you mentioned last episode." He knows what those four hours were; nobody else does. Only he can claim that line, so it lands in the "written to me" pile by itself.
Copy this · four rules for subject lines
1. Put a concrete noun in it: their project name, a number they said, the city they work in. Delete every abstraction — upgrade, enable, opportunity.
2. Write it the way a colleague writes one: no title case, no exclamation marks, six to twelve words is about right.
3. No curiosity gaps, no faked context: a "Re:" on a thread that never existed buys one open and permanent distrust — that's the persuasion-manipulation line, not a wording choice.
4. Never score yourself on opens: replies are the only number that counts. High opens with zero replies means the subject line and the body aren't the same email.
Why it works · mechanism and evidence
Why it works
"Relevant to me" isn't a preference, it's a priority channel in attention and memory: information hooked to the self is encoded more deeply and pulled out of background noise more easily. A subject line carrying a fact true only of this person is plugged into that channel; a line true of everyone waits in the noise.
Hard evidence · the self-reference effect: Rogers, Kuiper & Kirker (1977, JPSP) found that adjectives processed under "does this word describe you?" were recalled far better than the same words processed semantically or phonemically. Symons & Johnson (1997, Psychological Bulletin) confirmed the effect's stability in a meta-analysis of 129 studies. Classic and meta-analytically supported — hard.
Hard evidence (brain mechanism) · where self-reference lives: Kelley et al. (2002, Journal of Cognitive Neuroscience) found with fMRI that self-referential judgments specifically recruit medial prefrontal cortex (mPFC); Northoff et al. (2006, NeuroImage) pooled multiple studies and placed it among the cortical midline structures. "About me" travels a locatable route in the brain — that part isn't a metaphor.
Hard evidence · your own name surfaces from the unattended channel: in Moray's (1959, Quarterly Journal of Experimental Psychology) dichotic listening study, about a third of subjects noticed their own name in the ear they had been told to ignore; Wood & Cowan (1995, JEP: LMC) replicated at a similar rate. That is the model of inbox scanning: almost nothing is really read, but the item about me gets pulled out.
Boundary: these experiments measure memory and attentional capture, not reply rates — they give direction, not effect sizes. The reverse risk lives here too: curiosity gaps and faked context also capture attention, at the price of trust; attention is the means, being believed is the end. And note that only about a third of subjects caught their own name even in the lab, so plenty of well-written messages will be ignored anyway.
MOVE 02
Personalization is worth exactly what faking it would cost
Merge fields cost nothing, so they signal nothing.
costly signalsbecause-clausesone over many
The principle in one line
Personalization doesn't work by flattering anyone. It works by proving the message was written for one person. And a signal is credible only in proportion to the cost of faking it: {{first_name}} and {{company}} cost zero seconds and ship inside every mass-mail tool, so today they are a negative signal; restating public news costs five seconds; only an observation you couldn't have made without looking costs enough to count. And it has to carry a causal clause — not a compliment, but "because of this, I'm writing to you."
Three tiers of personalization, ranked by cost to fake
Copy this · how to write the first line
1. Spend ten minutes finding one artifact: something they said in public, a number, a detail inside their product. Include the location or date — which episode, which page, last Wednesday — because checkability is what makes it expensive.
2. Attach a because-clause, not a compliment: "you said last episode it took four hours — that's exactly why I'm writing," not "love your show." A compliment doesn't explain why you appeared; a because-clause does.
3. Use one concrete case, not a statistic: "the person who edits for you loses half a day a week" beats "the industry average saves 40% of post time" — one can be imagined, the other can only be agreed with.
4. If the cost isn't there, change the name: no artifact worth ten minutes means this person doesn't belong in this round's list (how to subtract the list: the prospecting piece).
Why it works · mechanism and evidence
Why it works
A stranger can't verify your sincerity, only the cost you paid — that's the core of signalling theory: a signal carries information only when faking it is expensive enough. The other half is the word "because": giving a reason raises compliance, but whether the content of the reason matters depends on how big the request is.
A theoretical model, not a psychology experiment · costly signals: Spence (1973, Quarterly Journal of Economics) showed in his labour-market signalling model that a signal separates two types reliably only when it costs the weaker type more. A model rather than an experiment — but it explains precisely why merge fields must fail: they cost the diligent sender and the mass sender exactly the same.
Moderate strength · because-clauses raise compliance, but empty ones only for small asks: in the copier study by Langer, Blank & Chanowitz (1978, JPSP), compliance with a queue-jumping request rose sharply when a reason was given (roughly six in ten to over nine in ten), and for a small request an empty reason ("because I have to make copies") worked nearly as well as a real one; but when the request grew (20 pages), the placebic reason stopped working and only a genuine reason still did. Small sample, single study, moderate. The useful half is the second one: your request isn't small, so your reason has to be real.
Hard evidence · one identifiable person beats a statistic: Small, Loewenstein & Slovic (2007, OBHDP) found donations to a single named beneficiary substantially exceeded donations driven by statistics, and that prompting analytical thinking first actually dampened the response to the identified individual. Peer-reviewed, well replicated. Translated to outreach: write a person and a scene that can be pictured, not an industry average.
Boundary: Langer's study measured small face-to-face requests, so discount it crossing over to email — and don't read it backwards. It isn't "adding because works"; it's "the bigger the ask, the more the reason has to be true." As for how much personalization lifts reply rates, the numbers in circulation come almost entirely from outreach-tool vendors reporting on their own customers — self-selected samples, no control groups — so this page gives mechanism and direction only. Measure the magnitude yourself.
MOVE 03
The first email converts on a reply, not a meeting
Shrink the ask until it costs ten seconds.
foot-in-the-doorten-second askgive first
The principle in one line
"Do you have thirty minutes?" is an expensive request from someone you don't know: open a calendar, spend future time, and carry the risk that the call turns out to be a pitch. So don't ask for the meeting — ask for something answerable in ten seconds: a yes or no, a name. Someone who has replied once is far likelier to grant twenty minutes next, because they now file themselves as a person who talks to you. And "leading with value" isn't attaching a brochure — it's putting a small piece of finished work in the body, with nothing asked in exchange.
Same email, three endings, three prices
Worked example
Setup: the same email to that podcaster, changing only the last line.
✗ Asking for a large block up front
"Attached is our services overview and rate card — would a 30-minute call work? Here's my scheduling link." Attachment, link, calendar, thirty minutes: four thresholds stacked, and any one of them alone is enough for him to put the email down.
✓ Give a small piece, then ask for ten seconds
"I marked the three places in last week's episode where the lip sync drifts — timestamps below, take them, no reply needed. One thing I'm curious about: is this edited in-house or outsourced right now?" Something usable, explicitly free, then a two-word question.
Copy this · a four-line email
1. Line one · why you: the observation true only of them, plus a because-clause.
2. Line two · the gift: one small thing you already finished for them — three marked timestamps, a number you looked up, a quote from a peer. Say explicitly that no reply is needed. A gift with an invoice stapled to it isn't a gift.
3. Line three · the ten-second question: something to confirm, not to think about. "Is this in-house or outsourced?" beats "do you have a need in this area?", because the second makes them do your qualifying for you.
4. Line four · the exit: "If it's not a fit, just say so and I'll stop." Knowing a single "no" ends it cleanly is what makes replying cheap.
5. Leave out: attachments, scheduling links, three time options, image signatures. Keep the whole thing inside one phone screen.
Why it works · mechanism and evidence
Why it works
Someone who has granted a small request updates their self-description to "a person who agrees to this kind of thing," which makes the larger request that follows unremarkable — the foot-in-the-door effect, running on self-perception: people infer who they are from what they did. The second mechanism is start-up cost: a vague request makes the other person work out when and how on their own, and that missing step is why plenty of willing people never move.
Hard evidence · foot-in-the-door: Freedman & Fraser (1966, JPSP) had subjects first grant a small request (answering a few questions about household products); weeks later, asked to let surveyors into their homes to inventory their belongings, about 53% of that group agreed versus about 22% of controls. Classic. Burger (1999, Personality and Social Psychology Review) pooled the literature and confirmed the effect is real but small and conditional — pay people, or make them feel pressured into the first agreement, and it vanishes, because they no longer attribute the act to themselves. So don't buy that first reply with an inducement.
Hard evidence (meta-analysis) · pin down the when and how: Gollwitzer & Sheeran (2006, Advances in Experimental Social Psychology) meta-analysed 94 studies and found implementation intentions ("when X happens, I will do Y") produce a medium-to-large effect on goal attainment (d ≈ 0.65). Translated: a specific next step ("reply with one word") beats an open-ended "happy to chat if you're interested."
Contested · more options, less action: Iyengar & Lepper (2000, JPSP) found a tasting booth with 24 jams converted at about 3% against about 30% for six. But the meta-analysis by Scheibehenne et al. (2010, Journal of Consumer Research) put the average effect near zero and highly condition-dependent. So "don't offer three time slots" is a recommendation from mechanism — one fewer decision is one fewer place to stall — not an established fact.
Boundary: foot-in-the-door snaps when the gap is too wide — a one-word reply doesn't imply a signed contract six months later, and real value has to fill the space between. The reverse risk lives here too: using the small ask as a hook to extract a "yes" and then climb it is something people detect, and what you lose then costs more than the deal. The difference isn't in the move, it's in whether the small thing you gave was genuinely useful and genuinely free.
MOVE 04
A sequence isn't persistence. It's sampling their timing.
New angle every message, and write the ending before you start.
timing sampleswidening gapsa fixed ending
The principle in one line
Most replies arrive after the first message, but what makes follow-ups work isn't that persistence is touching — it's sampling: how urgent this is for someone fluctuates over time. A budget lands, an owner changes, the system breaks again. More touches means more draws at the week they happen to be thinking about it. So every message must enter from a new angle — different pain, different evidence, different person, different channel. "Just following up, did you see this?" carries zero information and burns a slot. And the ending is written first: stop when the sequence ends, say so, archive.
Five steps, widening gaps, a fixed ending
Copy this · what's testable and what isn't
Testable (big differences, reachable at your sample size): audience, trigger condition, channel, sequence length (3 vs 5), and deliverability itself — does a message with a link land in spam.
Essentially untestable (effects too small): rewording the subject line, salutations, sign-offs, Tuesday versus Thursday. At your sample size the noise swallows all of it (the variance arithmetic is in the prospecting piece).
Operating rules: batches of at least 30, finish the batch before changing anything, one variable per batch and only from the first list, and record absolute counts — how many real conversations this batch produced.
Test deliverability before anything else: send the message as-is to two or three friends' inboxes and see whether it lands in spam. Ten minutes rules out a whole class of cause, and it's the one thing worth checking first when replies are zero.
Why it works · mechanism and evidence
Why it works
Sequences work through two independent mechanisms. Timing: urgency fluctuates, so repeated touches are repeated draws, and the hit probability rises with the number of draws regardless of how well you write. Spacing: distributed repetition leaves a firmer trace than bunched repetition. And a third thing worth naming out loud — repetition alone makes a claim feel truer, which is exactly why you must not lean on it to push a claim that doesn't stand up.
Hard evidence (meta-analysis) · spacing beats massing: Cepeda et al. (2006, Psychological Bulletin) quantitatively synthesised 254 studies confirming that distributed practice systematically outperforms massed practice, with the optimal gap lengthening as the retention interval grows. Note this is memory research, so applying it to outreach cadence is an analogy rather than direct evidence — it supports "don't cram five messages into one week," not any particular day count.
Hard evidence · repetition makes things feel truer (so be careful): Hasher, Goldstein & Toppino (1977, Journal of Verbal Learning and Verbal Behavior) found repeated statements were judged more likely to be true; Dechêne et al. (2010, PSPR) confirmed the truth effect meta-analytically. This is both a mechanism and a red line: repetition only raises how true something feels, never whether it is true — using it to push a claim you're unsure of is where persuasion becomes manipulation.
Hard evidence · why the ending has to be written first: Arkes & Blumer (1985, OBHDP) demonstrated the sunk cost fallacy across a series of experiments: resources already spent push people to keep investing in things they should abandon. Translated: by message seven your reason for continuing is usually "I've already sent six," not any new evidence. So fix the length before you start.
Boundary: cadence is constrained by channel and industry — email tolerates more than a direct message. And one hard line: the moment someone says they're not interested, the sequence ends, the name comes off the list, and it doesn't restart under a new pretext. That isn't only politeness — it decides whether you can still be recommended inside that circle, and that pool is worth far more than this one deal.
Your Day 42 Action
Forty-five minutes to trade "an email you worked hard on" for "an email that passes the sorting test, plus a sequence with an ending."
One (15 minutes) · dig up one artifact: pick a name and find one thing they said in public or one concrete detail, with its location attached — which episode, which page, which day. Can't find one? Change the name. Not finding one is itself the answer.
Two (15 minutes) · write four lines: why you (observation plus because), the small gift with "no reply needed" said out loud, a question answerable in ten seconds, and an exit. Put the concrete noun in the subject line. Then delete half of it and send.
Three (10 minutes) · lay out the sequence: write what changes on days 3, 8, 17 and 30, and what the closing note will say. Write the ending before you send the first one.
Four (5 minutes) · test deliverability: send it as-is to two friends' inboxes and see whether it lands in spam.
Boundary: if your total addressable set is a few dozen, don't run sequences at all — design each one separately. And whenever someone says they're not interested: end it, delete the name, never write again. No exceptions.
Think It Through
1. I'm sending job-hunting or partnership messages, not B2B cold email. Does this still hold?
It holds, and it works better — your list is only a few dozen people, so you can afford twenty minutes per message, which is precisely where this approach earns its keep.
Three things change. One, the artifact has to be harder: not "your company is doing great," but "I tried the feature you shipped last month and got stuck at step X — my guess is it's because of Y." What's scarce in a job-hunting message isn't enthusiasm, it's evidence that you already went and looked. Two, the ask has to be smaller: don't ask for an interview or a referral in the first message; ask for information ("is this role shorter on front-end or on data right now?"). Information costs them almost nothing; a referral means vouching for you. Three, the sequence gets shorter and softer: three messages maximum, the last one saying clearly that you'll stop.
One counter-intuitive note: for individuals, that closing message often gets the highest reply rate of all — it drops the ask to zero and releases the social obligation. But you have to actually mean it when you write it, otherwise you're just buying one reply.
2. Can I use AI to generate the "personalized" opening line in bulk?
Back to the core claim: a signal is worth what faking it costs. AI has pushed tier two — restating public information — from five seconds to zero, so it is turning the whole of tier two into a negative signal, the same thing that happened to merge fields a decade ago. Readers' thresholds are rising with it: what stops someone today is a sentence that obviously required someone to actually watch, try, or calculate.
So the answer splits. Using AI to search and organize: fine — have it pull out what this person has said publicly, or table thirty companies' public information. That saves your time; it doesn't fake cost. Using AI to write the observation itself: no — it can only work from public information, so the output necessarily lands in tier two, and tier two is now worth zero.
One self-check: could an AI write this sentence without your notes? If yes, it isn't worth much. The expensive information comes from the ten minutes you spent — actually using their product, actually finishing that episode. It's also one of the few parts of outreach automation can't eat.
3. I did all of this and my reply rate is still low. Am I just not writing well enough?
Probably not. Cold outreach reply rates are inherently low; single digits is normal. First separate two different illnesses: zero replies and low replies are not the same problem.
Zero is almost never a writing problem. Check in this order: did the message arrive (deliverability — ten minutes) → is the audience right (do these people have the problem at all) → is the trigger right (they have it, but is it urgent). Writing comes last, because writing moves you between two and four percent, not between zero and two.
Low but non-zero means the machine works, so stop rewording — the effect is too small for your sample size to see. Move the same energy to variables you can actually measure: a more urgent audience, a trigger-event column, a different channel.
One last honest note: the returns here are skewed. The one or two replies in a hundred may be worth more than everything the other ninety-eight cost — so the goal was never to raise the rate, it's to grow the absolute count.