Meta Andromeda: how ad retrieval works and what it changes for creative

Quick answer

Meta Andromeda is Meta's retrieval system: out of tens of millions of ads, it narrows the field to a few thousand candidates for each person before ranking. Ads that look very similar are treated as variations of the same creative. So launch genuinely different concepts, in a simple structure, instead of dozens of micro-variations.

What is Meta Andromeda?

It is the machine learning system Meta uses for ad retrieval. Meta's delivery happens in several stages. In the first one, retrieval, the system scans the ad inventory and pulls out a small group of candidates for that person, at that moment. Only those candidates move on to the ranking models, which estimate value for the person and for the advertiser, and from there to the auction.

The Meta engineering post that introduced Andromeda, published on December 2, 2024, describes the scale: retrieval selects a few thousand candidates out of tens of millions of ads for the models that follow. An ad that does not make it through this filter never reaches the auction, however good its bid.

What does Meta say officially, and what is market interpretation?

It is worth separating the two before you change the account, because a lot of the advice about Andromeda blends them together.

Why do similar creatives count as one?

Because, according to Meta itself, when several ads look or sound the same, they are treated as variations of the same creative, and learning and delivery optimization are shared at that level. In practice, ten versions of the same video with a different caption color do not give the system ten options to show ten kinds of people. They give it one option, repeated.

Meta separates two concepts in the same post. Iteration is refining an asset: same image, different copy; same video, different call to action. Diversification is creating distinct pieces, designed for different personas or use cases, that vary in visuals, mood, story and message. Both have their place. The mistake is calling iteration diversification and believing the account has variety.

What counts as a different concept?

A simple test: describe the ad in one sentence without mentioning color, font or edit. If two ads share the same sentence, they are the same concept.

What changes for media buyers?

The lever moved. Before, much of the work was slicing audiences: interests, age, 1% lookalikes. With retrieval choosing candidates per person, creative variety becomes what gives the system options for different audiences inside one broad ad set. Creative has become targeting, too.

• The brief changes. Instead of asking for "5 variations of the video that worked", ask for 3 to 5 different angles: pain, proof, objection, offer, behind the scenes.
• Production changes. One good piece of raw footage yields several iterations, but only one concept. Plan concepts before you plan edits.
• Analysis changes. Compare concepts with each other, not files. A concept with five micro-variations is not being tested five times.
• Hook testing still matters, just later: first find the concept, then swap the first seconds of the winner, as in creative testing on Meta Ads.

How many creatives should you put in a campaign?

Meta does not publish a number. The official post on diversification recommends a broad portfolio of assets, with more creative freedom across text, image and video, built for several personas, and stops there. The figures of 10 to 20 creatives you see around are rules of thumb from agencies and tools.

In our operation, we count concepts, and the number depends on the phase:

The practical limit is not the algorithm, it is the budget: every ad needs impressions to be read. At $100 a day, 20 ads barely get off the ground. The budget math per test ad set is in creative testing on Meta Ads.

Does your account structure need to get simpler?

Generally, yes, and this is where market advice lines up with the direction Meta points to when it recommends Advantage+. Many small ad sets, each with one interest, split budget and conversion signal; each one learns alone and slowly. Fewer ad sets, with broader audiences and more concepts inside, concentrate learning.

• Consolidate ad sets that do the same job. Three ad sets with similar interests for the same product usually become one.
• Keep apart what has different economics. Retargeting and broad prospecting, or two products with very different price points, stay separate.
• Separate testing from scaling. Test in ABO for a clean read; scale in CBO, as in CBO vs ABO.
• Local businesses are a partial exception. A small geographic radius is already a narrow audience; opening up interests there changes little. What changes is creative variety.

What about Advantage+?

Andromeda was introduced alongside Advantage+ automation: in the engineering post, Meta says Advantage+ and generative AI increase the number of creatives in the system, and that retrieval had to keep up. In the diversification post, Meta recommends Advantage+ campaigns and cites an internal figure: advertisers who were not using Advantage+ sales campaigns saw, on average, 22% higher ROAS when they turned on AI targeting.

Read that number for what it is: Meta's own data, with no period, niche or country breakdown disclosed. Advantage+ tends to work better when the account has conversion signal and creative variety for Meta to choose from. Without both, it automates the choice among very few options.

What are the most common mistakes after Andromeda?

1. Confusing volume with diversity. Twenty files, two concepts. The account looks full and the system sees very little.
2. Testing format before testing concept. 9:16 against 4:5 of the same video is an iteration test. Find out which message works first.
3. Broadening the audience without broadening the creative. A broad audience with a single concept shows the same ad to everyone. The two changes go together.
4. Spreading a few concepts across many ad sets. It fragments the learning that consolidation was meant to concentrate.
5. Launching 20 ads with budget for 4. The ones that get no impressions did not lose: they were never tested.
6. Blaming Andromeda for every drop. Check tracking, offer, frequency and CPM before rebuilding the structure.

How do you build creative volume without becoming hostage to production?

Treat creative as a calendar, not an emergency. A routine that works in our operation:

1. List the client's personas and objections. For an online course: people afraid they will not have time, people who bought a course before and never finished it, people who want proof of results. Each one becomes a concept.
2. Pick one format per concept. Testimonial, demo, behind the scenes, static offer, carousel. A different format helps the system tell them apart, but the concept comes first.
3. Produce 3 to 5 concepts per round and test them in ABO with a set budget.
4. Iterate only the winner. Change the opening, the copy and the call to action of the winning concept, and move it to the scaling campaign.
5. Replace before fatigue. The refresh pace by business type (12 days for digital products, 18 for ecommerce, for example) is in hook rate, hold rate and creative fatigue.

How do you launch several concepts at once in Ads Editor?

More concepts per round means more ads to upload, and in Ads Manager each one is a manual, laborious process. With bulk publishing in Ads Editor, the whole batch goes out in one action:

1. In bulk publishing, pick one or more accounts, the campaigns and, if you want, the target ad sets.
2. Drag the creatives for every concept at once, images and videos together. Each creative becomes its own ad, which keeps the read per concept.
3. Set copy, URL and UTMs once for the batch, using the account's UTM default.
4. Leave the status as Paused, check one ad and activate the batch together, so every concept starts at the same time.
5. Name each ad after its concept (pain-time, proof-testimonial), so the report compares concepts, not files.

The full walkthrough, including the limits of Meta's spreadsheet import, is in how to bulk upload Meta ads.