Abraham posts a three-step test for copy versus infra
The test isolates one variable before you tear down a sending setup that may be fine.
Nick Abraham published a three-step test for working out whether spam placement is a copy problem or an infrastructure problem, before anyone starts rebuilding a setup that might be healthy.
Step one is to send a normal B2B email from the same setup, something along the lines of "Hey, what time are we planning to speak?". Step two is to run an inbox placement test on it. If that plain email lands in primary, Abraham says the infrastructure is probably still healthy. Step three is to push the campaign copy through the same setup. If the control lands and the campaign copy goes to spam, the debugging starts with the copy.
If the control lands and the campaign copy goes to spam, start debugging the copy.
Why the control email matters
The whole thing rests on one idea, which Abraham states plainly at the end of the post. You change one variable at a time instead of blindly rebuilding the campaign. The plain email is the control. It uses the same domains, the same mailboxes and the same sending setup as the campaign, so the only difference between the two sends is the words.
For people sending, this is the step most teams skip. Placement drops, and the reflex is to buy new domains, start a fresh warmup cycle and write off weeks of sending reputation, on the assumption that the infrastructure went bad. Abraham's test costs one email and one placement check, and if the control lands in primary it tells you the rebuild would have fixed nothing.
What it does not tell you
The test is a diagnostic, not a diagnosis. It points at copy, it does not say which part of the copy is the problem, and Abraham does not name a link, a phrase or a formatting choice as the usual culprit here. It also only reads the setup at the moment you run it, so a control that lands today is evidence about today.
Abraham has been arguing in this direction for a while. He said earlier this week that 90% of copy split tests are noise, which is the same instinct applied in the other direction. Most changes operators make are not measured against a control, so the result they read is not the result they got. A plain email sent through the same mailboxes is about as cheap as a control gets.