Creative testing on Meta in 2026: why the ad is now the targeting

A few years ago, a good Facebook ad campaign started with the audience. You picked ages, cities and interests, built a lookalike from your customer list, and split everything into neat ad sets. The creative was almost an afterthought.

If you still run ads that way, you've probably noticed it doesn't behave the way it used to. Broad campaigns sometimes beat your carefully built audiences, and the same ad set swings from great to terrible in a week. The reason is that Meta's system has changed what it pays attention to. Today, the ad itself does much of the work your targeting settings used to do.

What changed inside Meta's ad system

When you run an ad, Meta has to decide, in a fraction of a second, which of many possible ads to show each person. In December 2024 Meta's engineering team described the system that does the first round of that sorting, called Andromeda. It narrows tens of millions of candidate ads down to a few thousand for the next stage to rank.

Meta says the new system has 10,000 times the model capacity of what came before, and that it improved recall (finding the relevant ads) by 6% and ads quality by 8% on selected segments. Those are Meta's own figures. The part that matters for you is in the same post: Andromeda is "designed to maximize ads performance by utilizing the exponential growth in volume of eligible ads". In other words, Meta built it to handle many more ads per advertiser, partly because its own tools now help people make more of them. The post says "more than a million advertisers used our generative AI (GenAI) tools to create more than 15 million ads in a month".

At the same time, Meta has pushed its automated campaign type, Advantage+, into the centre of Ads Manager. Advantage+ is Meta's name for settings where its AI chooses the audience, placements, budget split and sometimes creative variations for you. Meta has renamed Advantage+ shopping campaigns to Advantage+ sales campaigns, and its product page says the system can "automatically test your creative (up to 150 creative combinations)".

Put those together and you get a simple idea: when the system picks who sees your ad, the content of the ad tells it who to look for. A video that opens with a new mother struggling with a pram will find different people from one that opens with a cyclist fixing a puncture, even inside the same broad audience. That's why people say "the creative is the targeting".

What this means for how you test

The old habit was to test audiences: the same ad shown to five interest groups. The new habit is to test ideas: different ads shown to one broad audience, letting the system find the right people for each.

That doesn't mean "make 50 random ads". Meta's own guidance on A/B testing is still the core rule: "only one variable should change while the rest of the campaign setup and creative remains identical." If you change the hook, the offer and the format all at once, you won't know which one worked.

A useful way to think about it is four levels, from biggest to smallest:

Test the big levels first. Different angles tend to reach different people; small hook tweaks mostly change how many of the same people stop. Once you have a winning angle, test hooks inside it.

Meta's creative testing tool

Since late 2025, Ads Manager has had a built-in Creative Testing option at the ad level. Jon Loomer, who writes in detail about Meta ads, has a walkthrough of how it works (October 2025):

• It sits in the ad editing screen, just after the Ad Creative section.
• You test two to five ads, each a copy of the ad you're editing, which you then change.
• You set aside a test budget from your campaign; Meta recommends no more than 20%.
• The default test length is 7 days.
• It runs as a true A/B test: "each person will only see one of the ads", so the test ads don't compete with each other.
• It requires the Highest Volume bid strategy.

The benefit over just adding new ads to an ad set: new ads get guaranteed spend. Without a test, Meta often pours budget into one ad early and the others barely run, so you never find out whether they were any good.

A worked example: one test plan

Say you run a home cleaning service in Pune, and you have one ad running: a photo of a clean living room with "Book a deep clean today".

Here's how I'd plan the first test, with angle as the single variable. Everything else stays the same: same format (15-second phone video), same offer, same landing page, same audience.

You run it for the default seven days, judged on cost per booking. Suppose ad C wins clearly. You've learned that "guests are coming" is a reason people book. The next test keeps the guests angle and changes only the hook: three different opening lines about visitors. After that, you test the offer.

Notice what you're doing: each test answers one question, and the winner becomes the new baseline. That's how a small budget learns faster than a big one spread thin.

(The business and the result here are a made-up example to show the method, not a real case.)

Common mistakes

• Too many ad sets with tiny budgets. Meta's system needs enough results to learn. Splitting a small budget across many ad sets means none of them gets enough.
• Editing every day. Ads Manager marks new or heavily edited ad sets as "Learning" while the system works out who to show them to. If you change an ad on Tuesday and judge it on Wednesday, you're judging the system's guesswork, not your ad. Decide your test length up front and wait.
• Judging on the wrong number. A high click-through rate means nothing if nobody buys. Pick the result you actually want (purchases, bookings, leads) and judge on cost per result.
• Making near-identical ads. If your five ads differ only in the shade of the button, the system can't find different people for them. Make the differences meaningful.
• Ignoring fatigue. When the same people see an ad many times, the frequency figure in Ads Manager rises, clicks fall and costs climb. That's the moment to bring in the next winner from your tests, not to raise the budget.