Two Versions, One Difference
Write two variants of a piece of copy that differ in exactly one respect, so the result would actually tell you something. Harder than it sounds and the reason most informal tests are uninformative.
Why this one
Most copy tests compare two versions that differ in five ways at once. When one wins you have learned nothing about why, so you cannot apply it anywhere else. Isolation is the whole value.
Steps
Choose the single variable
Length. Or opening with the problem versus the outcome. Or a specific number versus a vague claim. One thing. Write down what you expect to win and why, before you generate anything.
Generate the pair under a strict constraint
The instruction to hold everything else constant is the one the model will quietly disobey.
PromptWrite two versions of [the copy]. The ONLY difference between them must be: [your one variable]. Hold constant: length (within 10%), tone, structure, the specific claims made, and the call to action. After writing both, list every difference between them. If your list contains anything other than the intended variable, rewrite them.
Audit the pair yourself
Read both and list the differences independently. You will usually find a second uncontrolled variable that the self-audit missed. Fix it by hand.
State what each outcome would mean
Before running anything, write down what you would conclude if A wins, if B wins, and if there is no difference. If any of those three has no useful conclusion, the test is not worth running.
You should end up with
A clean pair of variants and a written interpretation for each possible outcome.
Done when
- Exactly one variable differs, verified by your own audit
- You predicted the winner in advance
- All three outcomes have a stated interpretation
If you want to go further
- Design a second test whose result would change what you do, not just what you know