How to
How to run an ad split test that actually says something
Published · Updated
If both cells contain the same words, the test can't give you a result. Whichever wins, the number is noise — and you paid twice for one ad.
It's the most common way a first campaign wastes its budget, and it's almost invisible: the ads manager accepts it, both cells deliver, and the report shows a winner. The winner is variance.
How do I know if my two ad variants are actually different?
Read only the words a person sees, and ignore every label you gave them. Put the two cells side by side and read only the words a person would see — headline, body, button. Ignore the labels. A test matrix that says "Variant A: control / Variant B: emotional hook" is describing an intention, not a difference. If the visible copy's identical, so is the ad.
Why do AI-generated split tests end up identical?
Because "write me variations" describes paraphrase, and paraphrase is what you get. We measured this in our own output. Across 48 generated campaigns carrying two or more cells, 35 shared one headline across every cell and 15 were identical in both headline and body. The model filled in the axis labels correctly every time and then wrote the same ad twice.
That isn't a quirk of one tool. Asking for "variations" gets you paraphrase, because paraphrase is what the request literally describes. A test needs a different idea, not different adjectives.
How do I write a split test that can actually produce an answer?
Change the angle rather than the wording, vary exactly one thing, and check the audience still clears the platform's floor after you halve it.
- Change the angle, not the wording. If cell A leads on price, cell B should lead on the objection that stops people buying, not on the same price in warmer language.
- Vary one thing. Two cells differing on everything tell you which one won and nothing about why. One axis per cell, named.
- Check the audience survives the split. A 500-person audience split in two is two cells of 250, and LinkedIn will not deliver below 300 per cell. The campaign goes live and nothing happens — you won't be told why.
Does a valid split test mean the ads are good?
No — those are separate questions. Whether your two distinct cells are any good. This is about whether a test can produce an answer at all. A campaign can clear every point here and still be two weak ads.