New release: Apogee V2 is live Black Friday: 7 days unlimited for €1 −99% Ends in See the offerSee

More articles 14 min read

Meta Andromeda: What It Changes for Your Ad Creatives

Meta Andromeda, the system that preselects ads: what Meta has published, what it recommends on creative diversification, and what it has never said.

Mathieu Malige Founder, Apogee Updated
Contents
  1. What is Meta Andromeda?
  2. What does Meta recommend to advertisers?
  3. What makes a creative genuinely different?
  4. Where do concepts come from?
  5. How many creatives should you launch?
  6. How do you organize production, step by step?
  7. Which numbers are wrongly credited to Andromeda?
  8. What it changes in how you run an account
  9. What Andromeda does not solve

Since 2025, “Andromeda” has been used to justify almost anything on Meta: launch 50 creatives a week, delete your ad sets, give up on targeting. Some of that advice is sensible. Some of it credits Meta with numbers it has never published.

This page starts from what Meta has written: its engineering blog, its Meta for Business pages and its earnings calls, read on October 12, 2026, and its Help Center page on creative fatigue, read on November 3, 2026. The agency and practitioner benchmarks added on top, on the levels of a creative, batch volume and batch split, are credited to whoever gives them.

What is Meta Andromeda?

Andromeda is Meta’s ads retrieval system. When an ad slot opens up, it selects a few thousand relevant ads out of tens of millions of candidates. Heavier ranking models then pick the ones that get delivered.

Meta Andromeda · delivery

Three stages before the screen

Andromeda is the first: it narrows tens of millions of eligible ads down to a few thousand candidates.

Every eligible ad
  1. 1 Retrieval Andromeda Tens of millions of eligible ads a few thousand
  2. 2 Ranking GEM and the ranking models A few thousand candidates the most relevant
  3. 3 Auction Ad value The top-ranked ads one ad shown
Based on Meta, December 2024 and December 2025. Real ads from Grüns, Hydratis and Lymphoria (Meta Ad Library, via TrendTrack).

Meta introduced it on December 2, 2024 on its engineering blog, and rolled it out in the second half of 2024. The published figures:

What Meta measuredReported result
Retrieval system recall+6%
Ads quality+8%, on selected segments
Model capacity10,000 times larger
HardwareNVIDIA Grace Hopper chips and MTIA, Meta’s in-house accelerator

In July 2025, Meta announced Andromeda’s extension to Facebook Reels, with nearly 4% more conversions on Facebook mobile Feed and Reels.

Meta has kept improving it almost every quarter since. In Q3 2025, a single model combined retrieval and early-stage ranking, with 14% higher ad quality on Facebook. On October 28, 2026, Meta described the 2026 generation, which advertisers call “Andromeda 2.0”: according to Meta, it scores every eligible ad across Facebook and Instagram, at over one billion ads per second. What changed, and the full timeline, are in our Andromeda 2.0 article.

What does Meta recommend to advertisers?

The Meta for Business pages that mention Andromeda all hold the same line:

  • April 22, 2025. Meta writes that the focus has shifted from niche targeting to creative diversification, which it calls the best lever to find the most relevant audiences.
  • March 27, 2025. Meta explains that Andromeda works behind the scenes as businesses upload more and more creatives to support that strategy.
  • December 16, 2025. Meta specifies that ads that look or feel alike are treated as variations of the same creative.

Meta also offers a creative diversity indicator that labels ad sets as high, medium or low. It presents it as an estimate still in development.

For Meta, diversity is not measured in file count. Ten variants of the same video, with a different color or overlay text, are still variations of a single creative.

What makes a creative genuinely different?

If ads that look alike get grouped together, the question becomes: what makes two creatives different? Three levels help you tell.

LevelWhat changesWhat it bringsWhen to use it
ConceptThe angle, the promise, the persona, the awareness level or the format: talking head, demo, testimonial, staticA new creative that Meta can tell apart from the othersIn every batch: this is where diversity comes from
IterationOn a concept that won: the hook, the first frame or the edit, or a new production that keeps the structure and changes the staging, angle, persona, creator or formatThe best odds of success, because it reuses what workedAfter the results, on winners only
Micro-variationA button color, the font, the crop, one word of overlay textNothing new: very likely the same creative to Meta, which has published no thresholdNever as a test

A concept and an iteration can touch the same elements. What separates them is the starting point: a concept starts from an idea never tested on the account, an iteration starts from an ad that already won and changes one thing. The 2 or 3 hooks you launch a new concept with belong to the concept: they are there to test it.

Why do ten micro-variations count as one? Because what Meta writes about ads that look alike applies to them: ten files in Ads Manager, very likely one signal for Meta, and ten times the production cost. The trap shows up in the brief. A brief asking for “10 versions” of a video produces one creative. A brief asking for 4 concepts, each launched with 2 or 3 hooks, produces 4 creatives and 8 to 12 ads.

Meta’s page on creative fatigue points the same way. To spot a tired ad, Meta counts every recent exposure to its image or video, including exposures from other campaigns on your Page. Our inference: a new headline on the same video meets the same audience. For a tired ad, Meta recommends a new ad whose image or video is materially different from the original.

A quick test to tell whether two ads are two concepts: mute the sound and hide the text. If both videos tell the same story in pictures, with the same person in the same setting, they send Meta the same signal. The first seconds have to differ on screen.

Where do concepts come from?

A concept answers three questions: who am I talking to, about what, and at what point in their thinking. Five dimensions are enough to produce dozens:

  • the angle: a reason to buy, an objection to remove, a benefit to promise;
  • the awareness level: Eugene Schwartz described five, from the person unaware of their problem to the one ready to buy, through those who know the problem, the solutions, then your product;
  • the persona: the same product does not sell with the same words to a young parent and to a night-shift worker;
  • the format: talking head, demo, testimonial, street interview, comparison, static, carousel;
  • the person and the setting: who speaks, where, to whom.

Crossing two dimensions already fills a brief. An illustrative example, for an anti-fatigue supplement: the five awareness levels applied to two personas, a young parent and a night-shift worker, give ten distinct concepts. The story of feeling exhausted at 3 p.m. in front of the kids has nothing in common with a comparison against energy drinks, or with a night nurse’s testimonial. None is a micro-variation of another.

Meta has never written that Andromeda sorts ads by awareness level. But a message written for each level produces different creatives, which cover the funnel from cold to warm.

In Mathieu’s practice, concepts come from customer reviews, comments under the ads, questions sent to customer service and competitors’ ads in the Meta Ad Library, and customers’ exact words often make the best hooks. The blind spot he runs into most: everything sits at the bottom of the funnel (testimonials, promos, before/after), for people who already know the problem or the product. Concepts for people who don’t know they have the problem yet are harder to write; in his view, they are the ones that bring in new buyers.

The article on 25 ad creative concepts gives examples of genuinely distinct concepts.

How many creatives should you launch?

Meta has published no number tied to Andromeda.

Meta’s only figure predates the Andromeda announcement. In November 2024, Meta recommended at least 20 diversified ads in Advantage+ shopping campaigns, putting 20 to 30% of the budget toward testing new ads, and launching new ones every week.

The ranges that circulate, such as 10 to 50 creatives per ad set or 5 to 10 new ads per week, come from tools and agencies. They can be useful benchmarks, but they are not Meta’s guidance.

What Meta’s pages do say clearly is that the difference between creatives matters more than their number.

Four kinds of creatives and their odds

With no official number, some agencies think in splits instead. The one from Nilsen Irmouli, CEO of the creative agency Kreads, starts from one observation: creatives do not all have the same odds of working. In a February 2026 video (in French), he gives these hit rates:

Kind of creativeWhat it isHit rate according to Kreads
New conceptAn angle, promise or format never tested on the account10 to 20%
VariantAnother version of that new concept, launched at the same time, for example with another hook20 to 30%
Minor iterationMade after the results, on an ad that already spends: one visible element changes, such as the hook, the first frame or the edit60 to 80%
Major iterationA new production built on what worked: the same structure, with one of five changes (staging, angle, persona, creator or format)60 to 80%

The four kinds map onto the three levels above: a new concept and its variants, which are its launch hooks, sit at the concept level; minor and major iterations at the iteration level. Micro-variations are not counted.

These are one agency’s figures, not Meta’s. An iteration starts from an ad that already won; a new concept starts from zero.

The 50/20/30 split

From those rates, Nilsen Irmouli draws the split he presented at Meta’s offices in May 2026 (in French), for each batch:

  • 50% major iterations of winners;
  • 20% minor iterations, to keep what already spends alive longer;
  • 30% new at most, new concepts and their variants included.

The 30% of new is where next month’s winners come from: without it, results stop growing and the stock of winners to iterate on runs out. The reverse holds too: an account with no winner yet has nothing to iterate on, so its batch is almost all new concepts and variants.

Batch size follows from the test budget. In this calculation, a creative is one ad judged on its own: a new concept, each of its variants and each iteration count as one. In the routine of Mathieu, Apogee’s founder, each is judged once it has spent 3 times the target CPA. Hence:

Creatives a month = monthly test budget ÷ (3 × target CPA)

Example: $45,000 of monthly spend, 20% of it on testing (an example share, in line with the 20 to 30% Meta recommended for Advantage+ shopping campaigns), and a $30 target CPA. That is $9,000 ÷ $90 = 100 creatives a month. Split 50/20/30: 50 major iterations, 20 minor iterations and 30 new, which means 10 new concepts launched with 2 variants each.

Mathieu waits for 3 × the target CPA on accounts above about $500k of ad spend a month. On a smaller account like this one, he reads the cost per add to cart or the CPC first, so the formula gives a conservative count.

With Kreads’ rates, that batch gives 47 to 64 creatives that work: 1 to 2 concepts, 4 to 6 variants and 42 to 56 iterations. The same 100 creatives spent on new ideas only, 25 concepts launched with 3 variants each, give about 18 to 28. The bar may differ by kind: in Mathieu’s routine, an iteration only has to hold its parent’s CPA, while a new concept has to pass his three winner rules. Kreads has not published its own definition of a creative that works.

Your rates will differ. To measure them, add each creative’s type (new concept, variant, minor or major iteration) to the variables of your Meta Ads naming convention, then count the creatives that work each month, type by type.

How many winners do you need in stock? Each winner gives one major iteration per way to build it, and there are five: 50 major iterations need about ten winners. With fewer, raise the share of new.

Five ways to build a major iteration

Keep what was proven and change one thing. A win or a loss then tells you what that change did.

  1. New staging, same angle and persona. A founder interview that works becomes a podcast-style ad.
  2. New angle, same staging, structure and persona. The podcast stays, the argument changes.
  3. New persona, same staging and structure. The parts of the script that speak to one person are rewritten for another, with other problems.
  4. New creator, same script, angle and staging. A different age, look or gender. No new script and no new idea needed, which makes it the quickest of the five.
  5. New format, same angle, message and persona. A video becomes a static, sometimes just a screenshot of the video that worked.

How do you organize production, step by step?

  1. Research. Read reviews, comments and competitors’ ads. Write down the reasons to buy, the objections and customers’ exact phrases.
  2. Brief by concept. For each concept: the angle, the persona, the awareness level, the format and 2 or 3 launch hooks. One concept, one line of the brief. Iterations of last month’s winners go into the same brief, following the split you chose.
  3. Produce in batches. Shoot more footage than you need and file it by type of shot: product alone, real use, hands, details, before/after. A new hook, reordered shots and a new voice, together, are then enough for an iteration without a reshoot. Reordering shots on its own is still a micro-variation.
  4. Make the ratios. Every creative exists in 4:5 and 9:16, grouped in a single Meta ad so each placement gets the right file.
  5. Name, then launch. The concept, the iteration and the format go into the file name, then into the ad name: it is the first of the three points in the section on running an account, below.
  6. Read at concept level. A concept is judged on all of its ads, once they have spent enough: 3 times the target CPA per creative in Mathieu’s routine.
  7. Iterate on winners. Start with the hook, then move on to the five ways to build a major iteration. The next batch starts from those iterations, with 30% new at most. For Mathieu, the time to iterate on a winner shows over 7 or 14 days: its frequency rises and its CPA rises with it.

Which numbers are wrongly credited to Andromeda?

What you readWhat Meta actually published
”Andromeda increased ROAS by 22%“The 22% is about Advantage+, not Andromeda
”Andromeda processes 10,000 times more ads”The 10,000 is about model capacity
”You need 10 to 50 creatives per ad set”No number tied to Andromeda at Meta
”Above 60% similarity, a creative gets dropped”Meta has published no threshold, only a high, medium or low indicator
”Similar ads share a single ID”Meta does not describe this mechanism in its pages

A Meta spokesperson, quoted by Marketing Brew in September 2026, sums up the company’s position: its guidance is deliberately based on outcomes, not on a specific model.

What it changes in how you run an account

If diversification is the lever, the bottleneck moves. The job is no longer finding the right audience, but producing and launching more distinct creatives, every week, without spending your days on it.

Three points become critical:

  1. Name each file by concept and iteration. Otherwise you cannot tell which concepts drive the results. A Meta Ads naming convention exists for exactly that.
  2. Launch fast. Loading fifty creatives one by one in Ads Manager takes half a day. The Meta Ads bulk upload guide explains how to do it in one pass.
  3. Test in a readable structure. How you spread creatives across ad sets decides what you learn: that is what testing structures are about.

Apogee’s bulk upload reads the concept and iteration in each file name, groups the ratios of one creative into a single ad, and launches the batch in the testing structure you pick.

Andromeda does not change the media buyer’s job, it changes how the time is split: less time on targeting, more time producing, naming and launching distinct creatives.

What Andromeda does not solve

More distinct creatives do not save a weak offer. Diversification helps Meta find the right people for each message; it does not make the message better.

It does not replace reading the results either. Launching fifty creatives without being able to tell which concepts win just produces more noise.

Finally, Meta keeps changing its system. In July 2026, it announced it had deployed a first generative model into its ads retrieval system. On October 28, 2026, it described the 2026 generation of Andromeda, known as “Andromeda 2.0”. The diversification principles remain; the technical details will change.

The feature Bulk upload Meta Ads From raw files to live ads, in a single pass. See the feature

Frequently asked questions

What is Meta Andromeda?

Meta Andromeda is Meta's ads retrieval system. For each possible impression, it selects a few thousand relevant ads out of tens of millions of candidates, before ranking models pick the ones that get delivered. Meta introduced it on December 2, 2024 on its engineering blog.

What is Andromeda 2.0?

It is the name advertisers give to the 2026 generation of Andromeda, described by Meta on October 28, 2026. According to Meta, it scores every eligible ad across Facebook and Instagram, at over one billion ads per second, with a model made of several experts. Meta does not call it “2.0” and ties no number of creatives to it.

How many creatives should you launch with Andromeda?

Meta has published no number tied to Andromeda. Its only figure predates it: at least 20 diversified ads per Advantage+ shopping campaign, in November 2024. The 10 to 50 creatives per ad set ranges you see online come from tools and agencies, not from Meta.

What is the difference between a concept and an iteration?

A concept is an idea never tested on the account: a new angle, promise, persona or format. An iteration starts from an ad that already won and changes one element of it: the hook or the edit for a minor iteration; the staging, angle, persona, creator or format for a major one. Changing a color or a word of overlay text is neither: it is a micro-variation, which Meta very likely treats as the same creative.

What is the 50/20/30 rule for Meta creatives?

It is the batch split presented by Nilsen Irmouli, CEO of the creative agency Kreads: 50% major iterations of winners, 20% minor iterations and 30% new at most. It rests on the hit rates he gives: 10 to 20% for a new concept, 20 to 30% for a variant, 60 to 80% for an iteration. These are one agency's figures, not Meta guidance.

Does Andromeda penalize creatives that are too similar?

Meta says ads that look or feel alike are treated as variations of the same creative. It has published no similarity threshold: the percentages quoted online do not come from Meta.

Does Andromeda require broad targeting?

No Meta page about Andromeda asks you to switch to broad targeting. Meta does write, in April 2025, that creative diversification has become the best lever to find relevant audiences, rather than niche targeting.

Book a demo

Pick the time that suits you.

Open in a new tab

Loading the calendar…

Talk to Founder