Meta Andromeda: which creatives the algorithm treats as different, and how many you need

Andromeda in plain words

Ad delivery at Meta runs in stages. First, from tens of millions of ads, the system has to quickly pick a few thousand that fit one person — that's Andromeda's job. Other models then rank those thousands, and the auction runs.

According to the Meta Engineering blog (Dec 2, 2024), Andromeda delivered a +6% improvement in recall and +8% in ads quality on selected segments. The same post says the system is built for a large volume of creatives: within a month, over a million advertisers used Meta's generative AI tools to create more than 15 million ads, and advertisers adopting Advantage+ creative for the first time saw a 22% increase in ROAS.

What changed for creatives

The buyer's main lever used to be targeting: narrow audiences, interests, split ad sets. With Andromeda, Meta decides who sees which creative, and the creative itself becomes the signal. Jon Loomer, one of the most cited Meta ads experts, describes the shift as moving the main lever from narrow targeting to creative diversity.

In practice: for the algorithm to find different buyers, it needs different ads — for different motives, styles and situations. If every ad in a set says the same thing the same way, there is nothing to choose from.

Which creatives count as different

Based on Meta's guidance as summarized by Jon Loomer and market practice. Meta doesn't publish exact rules for how it judges creatives as similar; agency articles sometimes quote specific “visual difference” percentages, but there's no official confirmation.

So why make technical variants?

Technical edits don't turn a creative into a new idea — and they shouldn't. They have three jobs of their own:

1. Placements. One idea has to look right in Reels (9:16), feed (1:1 or 4:5) and in-stream (16:9). A frame cropped by the system or black bars hurt results. See one source, three placements.
2. Hooks. A different opening of the same clip changes the first impression — already a step toward diversity. See hook testing in the first 3 seconds.
3. Clean stats and rotation. Separate, clearly named files make tests readable and let you swap fatigued versions in time. See signs of creative fatigue.

Uniqueizer Pro does all three in one run: formats 9:16, 1:1 and 16:9, a pack of 5 different openings from one clip and file names for Ads Manager — so your time goes into new ideas, not exports.

Try it free — 5 files a month

How many creatives to launch

There is no single official number, and practitioners' benchmarks vary: Jon Loomer, for example, writes that he has tried ad sets with dozens of ads. For a small team we suggest starting like this (our approach, not a Meta rule):

• 3–5 different concepts;
• 2–3 hook variants for each;
• each in the placement formats you need.

That's 6–15 genuinely different ads instead of 30 copies of one idea. To plan a month, see how many ad creatives you need per month.

Rebuild your test for Andromeda in 5 steps

1. List 3–5 angles. Pain, benefit, social proof, comparison, product demo.
2. 2–3 hooks per angle. Question, number, before/after, an unexpected frame.
3. Build placement formats. 9:16, 1:1 or 4:5, 16:9 — in Uniqueizer Pro in one run.
4. Name files by a scheme, e.g. geo_offer_angle_hook_format, so the report shows what worked. Ready scheme: the file name generator.
5. Read stats by concept, not by file. Compare angles first, then hooks inside the winning angle. See the hooks × bodies testing matrix.

Common mistakes

• 10 crops of one clip instead of ideas. The algorithm has nothing to choose from.
• One format for every placement. Vertical gets cropped in feed; square gets bars in Reels.
• Conclusions after 1–2 days. The algorithm needs time to learn.
• No naming scheme. A week later nobody knows which file tested what.

Need the files for this? You can make Meta Ads creative variants from one master in a single run.