Cold email copywriting is one of those skills everyone assumes is the bottleneck. A campaign underperforms, someone rewrites the opener, reply rate moves half a point, and the team concludes the copy was the problem all along.
Usually it was not. Teams that fix a copy first tend to spend three months fixing the wrong variable, which is why any useful answer to how to write a cold email has to start one layer below the writing.
This piece covers what actually changes reply rates in B2B cold email: five principles that hold across industries, three worked examples, and one technical shift in 2026 that quietly changed what your variations are for.
What you'll learn:
- Where copy actually sits in the deliverability hierarchy
- Why writing about your own solution kills reply rates
- The assumption rule that separates a personalized cold email from an insulting one
- Three cold email examples, before and after
- Why your email bounce rate hides the problem instead of showing it
- What spintax is, what changed in 2026, and what it is still good for
Cold Email Deliverability First, Copy Fourth
Before you rewrite a single line, be honest about the order of operations. Cold email deliverability is decided upstream of the message, and copy cannot compensate for a failure there.
Infrastructure sits at the base: authentication, warmup, and sending volume. Data quality comes next: whether the accounts on your list belong there at all. Then targeting and timing. Copy comes fourth.
This ordering is not a preference, it is arithmetic. A perfect email sent to the wrong company converts at zero. A mediocre email sent to a company with a live, expensive problem converts to something. You cannot write your way out of a bad list.
The upstream question is whether outbound can work for your business at all, which we broke down in how to know if an outbound sales strategy will work for your company. If your offer needs a market education program before it makes sense, no copy principle below will save the campaign.
Everything that follows assumes you cleared that bar.
Principle 1: Write About the Problem, Not About Yourself
Here is the single most common failure we see when we audit an existing campaign: the email is about the sender.
It reads like a compressed company page. Who we are, what we do, how long we have been doing it, which logos we work with, why our approach is different. Every sentence is accurate. None of it is about the reader.
The reader has one question in the first two seconds, and it is not who you are. It is whether this concerns them.
The same failure shows up in a subtler form when teams lead with perks and credentials that matter internally but mean nothing externally. Office culture, team size, funding round, awards. These are things you are proud of. They are not reasons for a stranger to reply.
The test: delete every sentence that would still be true if you sent the email to a completely different industry. If the email survives that deletion intact, it is about you.
Principle 2: Personalized Cold Email Without the Accusation
This one costs more reply rate than any other mistake, and almost nobody notices they are making it.
Cold email advice tells you to lead with a pain point. So teams write pain points. The problem is that most pain points, phrased as an opener, are an accusation.
Write to a CTO that most engineering teams ship AI-generated code without review guidelines, and that CTO reads one thing: you think I am one of them. Write to an agency owner that most outbound agencies do not run enough A/B tests, and the agency owner reads the same thing. Nobody replies to a message that opens by implying they are careless.
The rule that fixes it: personalize on what is visible from the outside, never on what you are guessing about the inside.
An assumption about the world is an observation. An assumption about the reader's internal process is an accusation dressed as insight.
The two on the left invite a reply because they carry information the reader may not have. The two on the right invite defensiveness.
A better structure for the same underlying insight: name a trend you are observing across the market, then ask whether it resonates. That reframes the relationship from diagnosis to comparison, and it is the kind of framing worth testing deliberately when you are still validating message-market fit. You are not telling them what they do wrong, you are telling them what you see and asking where they land.
Principle 3: Cold Email Structure Built Around One Question
The best-performing TAM-wide campaign we have run in the last year was a single question with no pitch attached.
Not a value proposition. Not a case study. A question that the reader can answer in four words, where both possible answers are useful to you.
Take a supplier of specialist contract crews selling into a defined manufacturing niche. The opener does not explain the service, the coverage model, or the SLA. It asks whether they currently have a vendor covering that function. Yes, it tells you they are addressable and already spending. No tells you they are addressable and not spending. Silence tells you nothing, which is fine, because the message cost you two lines.
Why this outperforms a well-written pitch: a pitch asks the reader to evaluate. A question asks the reader to report a fact about their own operation. The second is dramatically cheaper cognitively, and it opens a thread you can develop once they respond.
The constraint is that this only works when your offer is concrete enough to compress into one question. If the reader has to accept four premises before your question makes sense, you do not have a question, you have a pitch with a question mark on the end.
Cold Email Examples: The First Three Principles Applied
Principles are easy to nod at and hard to apply. Below are three rewrites, one per principle so far. Each pair changes a single thing and leaves the rest of the message alone, which is also how you should edit your own copy.
Note the length. Every rewrite lands between 40 and 70 words. That is not a style preference, it is what survives an automated send at volume.
Principle 4: One Sequence Is Not a Campaign
A pattern we see repeatedly when taking over an account from a previous vendor: one sequence, written once, running for a month, with new leads poured in at the top.
Nobody repeated the message. Nobody segmented. Nobody ran a second angle against the same list to see which framing landed. The sequence opened with a name, a company, what the company does, and a request. It ran until the list was gone.
That is not a campaign. It is a broadcast with a queue attached.
A real campaign has:
- Multiple sequences per segment, because the same offer lands differently on a plant manager and a CFO, and differently again when a buying signal triggered the send
- A fixed iteration cadence, typically every two weeks, where underperforming variants get cut
- Per-variant reporting, so "it is not working" becomes "variant three is not working"
- A feedback loop with whoever owns the offer, because the person who sells it daily knows which objection comes up first
Without per-variant data you are not testing, you are guessing with extra steps. Most sequencers expose this now. Very few teams look at it.
Principle 5: Send a Plain Text Email and Nothing Else
A plain text email is the format that survives automated sending at volume. The technical layer is trivial to get right and routinely gets ignored.
For automated outbound at volume:
- No links in the first message. Every link is a signal to filter on, and the reply you want does not require one.
- No images, including in the signature. Tracking pixels and image-heavy signatures are among the oldest bulk-mail markers.
- No HTML formatting. Plain text, no styling, no custom code. The email should look like something a person typed.
- No attachments. Same reasoning, higher penalty.
None of this is clever. It is hygiene. But it interacts directly with the next section, because in 2026 the cost of getting hygiene wrong went up sharply.
What Is Spintax, and What Actually Changed in 2026
Spintax is a syntax that generates multiple versions of one email from a single template, written as {option one|option two|option three}. Your sequencer picks one combination per recipient, so no two messages sent are byte-identical.
The old justification was simple: identical content sent to hundreds of recipients gets fingerprinted as bulk mail, and spinning breaks the fingerprint. Treat it as hygiene, apply a few blocks, move on.
That framing is wrong, and the reason it survived so long is that the failure mode is almost invisible in your reporting.
Email Bounce Rate: The Second Number Nobody Watches
Most teams monitor a single email bounce rate. There are two kinds hiding inside that number, and they mean opposite things.
A hard bounce means the recipient address does not exist. That is a list quality problem, and a verifier fixes it.
A sender bounce means your own inbox provider refused to let the message leave. The send was blocked and soft-bounced back to you. Nothing was wrong with the address. Something was wrong with what you were trying to send. Google appears noticeably more aggressive about this than Outlook.
Eric Nowoslawski of Growth Engine X, who sends millions of emails a month, published campaign data isolating exactly this. Two variants of one enterprise campaign: same list, same inboxes, same subject lines, same signatures, same unsubscribe language. The single difference was body variation. Email A ran roughly 20 spintax options per line. Email B ran three.
Read the hard bounce row first, because it is the control. The two variants differ by 0.08 percentage points, which confirms both lists were equally clean. The verifier did its job.
Now read the sender bounce row. Email B was blocked at roughly twenty times the rate of Email A. Nearly a quarter of everything it tried to send never left the building. Reply rate followed at 1.38% against 0.22%, which is largely just the arithmetic of a message that never arrived.
The campaign had been running normally against enterprise accounts before the bounce alert fired. Sender bounces build quietly and then spike.
The tell is the relationship between the two bounce types. When hard bounces stay flat on a list you have already verified but total bounces climb, the list is not the problem. The copy is.
Cold Email vs Spam: Why This Effect Is Hard to Prove
The data above is unusually legible. Most of the time you will not get a chart like that, and you should be sceptical of anyone who claims otherwise, including us.
The measurement problem is structural. When variation is too thin, the most common outcome is not a visible block. It is that the message is accepted, delivered, and filed into spam. Delivery reporting shows it as sent. Bounce reporting shows nothing. Reply rate drops, and there is no line item anywhere explaining why.
That means the metric most teams would use to prove the effect is exactly the metric that cannot see it. You only get a legible signal in the extreme case above, where the provider stopped accepting the send outright instead of quietly filtering it. That is the visible tip of a much larger effect.
The practical rule: if reply rate falls on a list you have already validated, and bounces look normal, check the variation in your copy before you rewrite the message.
Why this inverts the usual advice
Most published guidance says to apply spintax to two to four blocks and stop, on the grounds that more variation produces unreadable combinations and unmanageable testing. The data above says thin spintax is what gets you blocked.
Both concerns are real. They are just about different things, and conflating them is the mistake.
The two campaign versions described above were two test arms. Inside each arm, density is a separate dial. You want two or three arms so the comparison stays readable, and heavy variation inside each one so neither arm gets throttled.
The constraint on density is not a number of blocks. It is whether every generated combination still reads like something a person wrote. Preview fifteen rendered versions out loud before launch. If they hold up, add more variation, not less.
The compliance and infrastructure floor still applies
None of this substitutes for the layer underneath it. Google and Yahoo introduced bulk sender requirements in October 2023 and Microsoft followed. Through 2024 and most of 2025, failing them meant landing in spam. From November 2025 Gmail began issuing permanent rejections at the SMTP level, so non-compliant mail does not reach the spam folder, it never arrives. Postmaster Tools was rebuilt at the same time, replacing graded reputation with a binary compliance status.
The bulk sender threshold is 5,000 messages per day to Gmail addresses, and complaint rate has to stay under 0.10%. Our guide to email deliverability covers the SPF, DKIM, DMARC, and warmup layer in full.
Consent and legal basis sit alongside this, not after it. What you are permitted to send, to whom, and what you must include differs by jurisdiction, which we cover in cold email and data privacy.
One second-order problem worth naming. AI-generated copy has its own fingerprint. When everyone prompts the same models with similar instructions the output converges, and convergence is itself a detectable pattern. Spinning AI-generated copy with an AI spinner compounds that rather than solving it, which is why the readability check has to be done by a person.
Whatever you use to send, it has to report hard bounces and sender bounces as separate numbers, broken down per variant. If your reporting collapses them into one bounce figure, you cannot see the problem described above, and you will keep blaming the list.
What we found in our own campaigns
We went back through 46 campaigns and 204 message variants in our own sending stack to check how much spintax we were running.
The answer was none. Zero spin blocks across all 204 variants.
That looks like a contradiction until you see what replaced it. Those 204 variants contain 441 merge fields, and 359 of them are not standard fields like first name or company. They are generated per lead: an opening line written for that specific account, a reason-for-reaching-out built from that account's own context. 81% of our variants carry at least one generated field.
The effect on the inbox provider is the same one Email A achieved, arrived at from the other direction. Spintax produces variation by recombining phrases you wrote in advance. Per-lead generation produces variation by writing the line fresh for each recipient. Both break the pattern. The second also happens to be more relevant, which is why our reply rates hold without spinning.
Two more numbers from the same audit, both of which line up with the principles above. 73% of our variants contain a question mark, which is the one-question principle showing up in practice rather than in theory. 8% contain a link and none contain an image, which is the stripped-message rule.
The takeaway is not that spintax is optional. It is that the thing you are actually buying is variation per send, and there are two ways to buy it. If your stack cannot generate copy per lead, spin heavily. If it can, generate.
Pre-Send Checklist: Six Cold Email Tips That Prevent the Usual Mistakes
Six cold email tips, ordered by how much damage the matching mistake does. The most expensive cold email mistakes all sit above the copy layer. Before any sequence goes live:
- Delete every sentence that would survive in a different industry. If the email stays intact, rewrite it.
- Check every assumption. External and verifiable, or internal and guessed? Cut the second kind.
- Try compressing the opener into one question. If it survives, send the question instead.
- Confirm you have at least two sequences per segment, with a date in the calendar to review variant performance.
- Strip the message, then count your variations. No links, no images, no HTML, no attachments. Then check spintax density per line, and check sender bounces separately from hard bounces.
- Verify authentication and complaint rate before you blame the copy. In 2026 those decide whether the copy gets read at all.
The last point is the one that reorders everything above it. Copy is the variable you tune once the others are correct, and the last one worth touching when a campaign underperforms.
FAQs
What is the difference between a hard bounce and a soft bounce?
A hard bounce is permanent: the address does not exist, and the send will never succeed. A soft bounce is temporary, and the mailbox is usually fine. The soft bounce category is where sender bounces live, which is why they hide so well. A verifier removes hard bounces and does nothing about the rest, so a campaign can show a clean hard bounce rate while a quarter of its sends are being refused.
How long should a cold email be?
Between 40 and 70 words for a first touch. Across our own campaigns the median sits at 67. Length is a symptom rather than a target: if the message runs long, it is usually because the reader has to accept several premises before the ask makes sense, and cutting words will not fix that. Cut premises instead.
How should I start a cold email?
With something about the reader's situation that is verifiable from outside their company. A market event, a regulatory change, a published signal. Do not open with an assumption about their internal process, however well phrased, because that reads as an accusation and kills the reply.
Is cold email legal?
In most jurisdictions yes, under conditions that vary by market: what counts as a lawful basis, what you must disclose, and how you must handle opt-outs all differ. Compliance is a precondition rather than an optimisation, and it sits alongside authentication rather than after it. We cover the detail in cold email and data privacy.
Does spintax still work in 2026?
Yes, but not for the reason it was originally sold. It no longer meaningfully defeats content fingerprinting, because filters evaluate meaning rather than exact strings. What it still does is keep you from being refused at the point of sending, and the campaign data above suggests thin spinning is worse than none of the usual scapegoats. If your stack can generate copy per lead instead, that achieves the same variation and is more relevant.
Get a Read on Your Current Copy
If you have a sequence running and cannot tell whether the problem is the message, the list, or the infrastructure, book a 30-minute call. We will look at what you are sending and tell you which of the three it is.
