Most creative tests do not fail because the ideas were bad. They fail because, at the end, nobody can say what the result means. Two ads with different openings, different presenters and different endings were run against each other, one did better, and the only lesson available is "that one."
A test that teaches you something changes one thing at a time, names every version by what changed, and decides in advance what would count as a win. This guide covers how to set one up for UGC-style video, and how to build the variants without remaking the whole ad each time.
The one rule: change one thing
If two things change between variant A and variant B, a difference in results cannot be traced to either. So each round of testing picks one variable, holds everything else still, and makes several versions of just that one.
That sounds slow, and it is slower than throwing ten unrelated ads at the wall. It is also the only way to learn something you can use on the next campaign, rather than something true of one ad on one week.
What to test, in order
Not every variable is worth the same. Test them roughly in this order.
1. Hooks
The first three seconds decide whether anything after them is seen, so the opening is the first thing to test. Write at least three hooks in different patterns — a question, the problem said out loud, a specific number — rather than three wordings of the same idea. Different patterns tell you what kind of opening a product wants; different wordings only tell you which sentence won. Our guide to writing hooks covers the six patterns.
2. Presenters
Who is on screen changes who feels addressed. Vary the things that map to the audiences you are trying to reach — age, region, voice, setting — and keep the script identical. If a brand sells to two quite different groups, this is where a test earns its keep.
If the presenters are AI presenters, they are labelled as AI and they never speak as customers, whichever one wins. See AI presenters and the FTC rule.
3. Calls to action
The last three seconds, and nothing else. "Shop now" against "See the colours", or a price stamp against none. Endings usually move less than openings, which is why they come later.
4. Length and shape
Fifteen seconds against thirty, or 9:16 against 4:5 in the same feed. These are worth testing once the opening and the presenter are settled, because they change everything at once and are hard to read before then.
5. Language
Each language is a new variant, not a setting. Transcreate — rewrite the script so it sounds native, keeping every fact and roughly the same length — rather than translating word for word.
Build the matrix without remaking the ad
A test of three hooks, two presenters and two calls to action is twelve variants. Made naively, that is twelve ads. Made the way UGC ads are actually built, it is far less work, because the middle of the ad does not change.
Split every ad into three parts:
- The opening: the hook, three seconds.
- The body: the problem, the product doing it, and the reason to believe.
- The ending: the call to action.
The body is shared across hooks and endings, so it is made once per presenter (and once per language). The openings and endings are short, and cheap to make in several versions. For three hooks, two presenters and two endings, that is two bodies, six openings and four endings — twelve short pieces that assemble into all twelve variants.
Name every variant by what changed
A name like "Ad v7 final (2)" throws away the one thing a test is for. Name each variant by its parts, in a fixed order, so the names sort and the results read themselves:
| Variant | Hook | Presenter | Ending |
|---|---|---|---|
| H1-P1-E1 | Question | Presenter A | Shop now |
| H2-P1-E1 | The problem, out loud | Presenter A | Shop now |
| H3-P1-E1 | Specific number | Presenter A | Shop now |
| H1-P2-E1 | Question | Presenter B | Shop now |
When H2 beats H1 and H3 with both presenters, you have learned something about openings. When P2 wins with every hook, you have learned something about who the ad is for.
Give each variant a fair chance
The most common testing mistake is calling a winner too early. A handful of results can swing either way by chance, and a variant that looks twice as good on day one is often level by day four.
A few habits help, without any statistics:
- Decide the budget and the stopping point before you start. Write down how much each variant gets and what result would make you stop it. Deciding afterwards invites you to stop the moment your favourite is ahead.
- Split the audience. Meta and TikTok both offer split tests in their ad managers, which divide the audience so that variants do not compete for the same people.
- Run variants at the same time. A variant run on a Tuesday and another on a Saturday are testing the days as much as the ads.
- Change nothing mid-test. Editing a live variant turns it into a different ad.
Read the right number for each variable
Different variables move different numbers, and reading the wrong one hides the answer.
- Hooks: look at how many people keep watching past the opening. Ad platforms report this under names like three-second views or thumb-stop rate; divided by impressions, it is often called the hook rate.
- Presenters and bodies: look at how long people keep watching, and how many click.
- Endings: look at clicks and conversions, since the ending only reaches people who stayed.
Keep the downstream number in view throughout. A hook that stops more thumbs but sells less is not a winner, just a louder loser.
Then change the next thing
Keep the winner, and make it the new control. The next round changes the next variable: if the problem-first hook won, test presenters behind it. Over a few rounds, the ad converges on a combination you understand, which is worth more than one you stumbled on.
Where Vessero fits
On a Vessero board today you can build the pieces of a test side by side: wire one brief into several hook cards, run each card for up to four takes at a time, and see the price of every run before you press it. The free hook generator writes ten hooks across the patterns to test from.
Two parts of this workflow are still being built, and are marked as coming across the site until they ship: a variations card that fans one concept out into hooks, presenters, lengths and formats and prices the whole matrix before it runs, and Cast, the presenters themselves, always labelled as AI.