Short answer: make both, and count them separately. Variations get more out of an ad that is already working. Only new concepts can find your next winner. If most of last week’s batch were variations, make more new concepts.
A variant is the same ad dressed differently: a new hook, different text over it, or a slightly different visual treatment. A new bet (a new concept) changes what the ad says: a different claim about what the app does for you, a different pain the viewer recognises, or a different reason to install today. Variants get more out of something that is already working. Only a new bet can find your next winner.
If you are shipping more AI ads than ever and finding fewer winners, most of your batch is probably not new. Twenty variants of one claim is one test.
Page Contents
Why do AI ad batches drift into variants?
Because AI made variants nearly free. Regenerating a winner with a new opening line, caption or colour grade takes minutes. A different claim takes research, judgment and a willingness to be wrong. So volume flows to the cheap side, and batches fill up with the same idea, dressed differently.
The ad count goes up and every file is technically a different ad. But the auction is asked the same question over and over: does this claim work? It answered that when the original won.
What counts as a creative variant?
A variant keeps the claim, the pain and the reason to install, and changes how they are presented. Take a hypothetical sleep app whose winning ad is a talking head saying that a guided wind-down gets you to sleep faster than scrolling in bed.
- The same ad with a new hook. The original opens with “Put the phone down and press play.” The variant opens with “I used to fall asleep with my phone on my face.” The rest still argues that the wind-down beats scrolling.
- The same ad with different text over it. The on-screen line changes from “Stop scrolling at 1am” to “Asleep before you finish the chapter.” The words are new; the promise is identical.
- The same ad with a slightly different visual treatment. A warmer grade, a darker bedroom, a different caption style, a tighter crop.
Variants have a real job: getting more out of something that is already working. They extend a winner’s life and show which packaging the claim travels best in. That is not the work of finding the next winner.
What counts as a new bet?
A new bet changes one of the three things a variant holds still. Same sleep app:
- A different claim about what the app does for you. Instead of “you fall asleep faster,” the ad argues “you wake up less tired.”
- A different pain the viewer recognises. The winner speaks to people who cannot fall asleep. A new bet speaks to people who fall asleep fine and then wake at 3am with their mind racing.
- A different reason to install today. The winner implies “sleep better, eventually.” A new bet gives a reason for tonight: a big presentation tomorrow, a flight that lands at dawn, the clocks changing this weekend.
The test is simple: if you described the ad in one sentence to someone who had seen the winner, would they say “that’s the same ad”? If yes, it is a variant, however different it looks.
How do you check how many new bets are in your batch?
Write one line per ad: its claim, the pain it names, its reason to install today. Then count the distinct claims, pains and reasons to install the batch actually contains.
That count is often far smaller than the ad count. It is the batch-level version of measuring creative diversity in your Meta account. A batch that names one claim and one pain is a test of one idea, however many ads it holds.
Next, work out the share of new bets: ads whose claim, pain or reason to install has not been tested before, divided by everything you shipped. It tells you how much of the week is looking for a new winner and how much is maintaining the ones you have. Both are legitimate, as long as you know which is which.
How do you plan a week with both variants and new bets?
Plan them as two separate lists.
- Variants of the winners. Each live winner gets variants: new hooks, new on-screen text, new visual treatments. Every variant carries its parent’s name.
- New bets. Each starts from a different claim, pain or reason to install, drawn from what customers actually say in reviews, support tickets and cancellation reasons. Each gets its own name, with no parent.
Keep them separate in naming and in reporting. Mixed together, results get confused: a strong variant makes the batch look inventive, and a new bet gets judged against a winner’s offspring. Every week we make both: variants of the winners, and new bets. We never count one as the other.
How many of each depends on how much is winning. Several fresh winners carrying spend pull the week toward variants; one tiring winner carrying everything pulls it toward new bets. There is no fixed ratio worth copying, but there is one fixed rule: a week with zero new bets cannot find a new winner.
The mechanics of setting up the tests themselves are covered in how to structure creative tests for Meta ads.
How do you judge whether a new bet worked?
On its own terms. A variant is compared to its parent: is this packaging better than the original? A new bet has no parent.
A variant starts with an advantage: its claim is already proven. A new bet starts cold. Judge it against a variant’s cost and it will usually look like a loser, and you will keep cutting the only ads that could replace your winner.
Give each new bet its own read: enough budget and time to reach a decision, judged against what a winner needs to be for your app, not against what this week’s variants cost. Then read why it did what it did. A new bet that loses at the opening but converts the people who watch tells you the claim may be right and the hook wrong. That is not a dead idea. It is a new bet that has earned variants of its own. More on the reading side in how to analyze ad creative performance.
Frequently asked questions
Is a new actor a new concept? Usually not. A different person saying the same claim to the same viewer is a variant: the face changed, the argument did not. It becomes a new bet only when the person changes who the ad is for. If swapping a young professional for a new parent turns “I can’t switch off after work” into “I can’t get back to sleep after the night feed”, the pain and the audience changed, and that is a new bet.
How many new bets should you make each week? It depends on how much is already winning: fresh winners justify more variants, tiring ones or a single ad carrying the spend call for more new bets. The one answer that is always wrong is zero.
Do variants still matter? Yes. They are how you get the most out of a claim that already works. The mistake is not making variants; it is counting them as new ideas.
How should you name variants and new bets in reporting? Make the type visible in the name. Variants carry their parent’s name plus what changed (hook, text or visual); new bets carry their own name and the claim, pain or reason to install they test. Report the two separately.
Can AI generate new bets, or only variants? Both. It makes variants nearly free, and it can draft new claims, pains and reasons to install from real customer language. Deciding which drafts are genuinely different bets, rather than the old claim reworded, is still a human call.
What is an ad concept? The idea an ad is betting on: a claim about what the product does for someone, a pain they recognise, or a reason to act today. Hooks, actors, on-screen text and edits are how a concept is dressed. Change only those and you have a variation of the same concept. More in creative concept vs creative variant.
How do UA teams organise hundreds of ad creative variants? Name every ad by its concept first and its variation second, so reporting rolls variants up to the concept they test. Keep new concepts and variants in separate lists when you plan and separate rows when you report, so a strong variant is never mistaken for a new idea.
