Deliverability before copy. That order is not optional.

Three separate campaigns this year came in looking like the same problem: replies had dropped, and the request was to rewrite the message. None of the three needed a rewrite. All three needed a different diagnosis first.

That order matters more than it sounds like it should. Copy and infrastructure produce the same visible symptom, fewer replies, but they need opposite fixes, and guessing wrong wastes the one resource a struggling campaign can least afford: time before the account owner loses confidence in the channel.

The symptom is identical either way

A cold email campaign that stops converting looks the same from the sender's side whether the message is landing in spam, landing in a broken inbox, rendering as a wall of unformatted text, or landing perfectly and just not persuading anyone. The reply count drops in all four cases. Nothing in a reply-rate chart tells you which one you are looking at.

That ambiguity is what sends people straight to the copy. It is the visible, editable part. A domain's sender reputation is not visible from inside the sending tool. A rendering bug in one email client is invisible unless you specifically open the email in that client. So copy gets blamed by default, not because it is usually the problem, but because it is the only variable most people know how to check without extra tools.

What to check first, in order

1. Inbox placement. Is the email landing in the primary inbox, promotions, or spam, across the actual providers your list uses, not just one test inbox. A message that never reaches the inbox cannot be read, no matter how it is written.
2. Domain and mailbox health. Warmup status, bounce rate, and blacklist status on every sending domain. A domain that is still warming up, or has picked up a blacklist flag, suppresses delivery regardless of what the email says.
3. Authentication records. SPF, DKIM, and DMARC configured correctly on every sending domain. Misconfigured or missing records are one of the fastest ways an otherwise well-written campaign gets filtered before a human ever sees it.
4. Rendering. Open the actual email in the mail clients your list is likely to use, not just the sending tool's preview. Broken HTML, missing line breaks, and images that fail to load all read as unprofessional or suspicious, and none of that is a copy problem.

Only once all four check out clean does a copy review actually isolate the message as the variable being tested. Skipping ahead to copy without ruling these out means the next test is not really testing the copy, it is testing copy plus three unknowns.

Why this keeps happening

Infrastructure work is less visible and less satisfying than copy work. Rewriting a message feels like progress, you get a new draft to review in an afternoon. Diagnosing mailbox health means checking dashboards, warmup scores, and authentication records that most people setting up a campaign never learn to read. It is not that copy diagnosis is wrong to consider, it is that it is the easy hypothesis to reach for, and the easy hypothesis is not always the correct one.

This is also why running cold email at real volume needs the same infrastructure discipline applied continuously, not just when something breaks. Dedicated lookalike domains keep a company's primary domain from carrying any of this risk. Daily health checks on warmup scores, bounces, and blacklist status catch a degrading domain before it silently kills reply rates. Multi-inbox setups spread volume so no single mailbox absorbs enough send pressure to damage its own reputation.

The fix depends on what actually broke

A campaign failing on infrastructure needs domain and inbox work, not a new draft. Rewriting the message changes nothing measurable, because the new version has the exact same delivery problem the old one did, it just has not been tested long enough to prove it.

A campaign failing on copy, once infrastructure is confirmed clean, is the case where a message rewrite is the right move, and now it is a rewrite aimed at a target that has actually been isolated, not guessed at.

Check the infrastructure first. It is not the more interesting fix, and it is usually the real one.