Meta Andromeda Explained: Creative Strategy for 2026
Andromeda decides which of your ads even reach Meta's ranking stage. Here is what Meta has actually said about it, what practitioners only infer, and a weekly concept and launch cadence built for how retrieval works.
What is Meta Andromeda?
Meta Andromeda is the retrieval stage of Meta's ads system. For each person, it narrows tens of millions of eligible ads to the few thousand worth ranking, and only then do the ranking models and the auction decide what shows. The practical upshot: genuinely different ads give retrieval more distinct options to match to more kinds of people, while small variations of one idea tend to reach the same people, which Meta itself warns about.
Meta's engineering team described the system in a December 2, 2024 post, which reported a 6% recall improvement and an 8% ads quality improvement on selected segments. It also explained why the system was needed: with generative AI creative tools, the number of ad creatives in Meta's systems was "expected to grow significantly."
Where it sits in the pipeline
- Retrieval (Andromeda): builds a shortlist of candidate ads for one person in milliseconds.
- Ranking: heavier models predict how likely that person is to act on each shortlisted ad.
- Auction: your bid, the estimated action rate and ad quality decide which ad wins.
An ad that never makes the shortlist for a person cannot win that impression, however good it is. Andromeda is not a setting, a campaign type or a bid option. You cannot switch it off, so the only levers you hold are what you feed it: your creative, your conversion data and your account structure.
What changed in 2026, and is Andromeda 2.0 real?
On October 28, 2026, Meta's engineering blog described the next generation of Andromeda, and Meta for Business published an advertiser summary the same day. The changes, in Meta's own terms:
- Specialized models. Instead of one general model, Andromeda now routes each ad to specialized "expert" models, which Meta says lets retrieval judge ads almost as well as the ranking stage does.
- Every eligible ad gets scored. The system "surfaces the top candidates at over one billion ads per second."
- Reported results. A 2.4% increase in conversions on Facebook since 2025. In Q3 2026, a tenfold increase in model complexity lifted conversions 1.37% on Facebook and 0.24% on Instagram Reels.
- Nothing to opt into. "There's no setting to switch on, no new campaign type to adopt, no migration to plan."
"Andromeda 2.0" is not Meta's name. It is a practitioner label, and Meta's posts never use it. Some people also stretch it to cover GEM, the foundation model behind Meta's ads recommendations, and GenRecs, a generative recommendation model covered in the same announcement. When someone blames "Andromeda 2.0" for a bad week, ask which system they mean, then check your own account first: a budget jump, a learning reset or a tired creative is the more common culprit.
Meta's advice in the announcement is short: results still depend on "what you give the system to work with," meaning the outcome you optimize for, the quality of the data you share and how much optimization you hand to the models. Its companion explainer on how ads are matched to people adds one creative instruction: supply "more creative variety so the system has more to test and learn from."
What Meta has said vs. what practitioners infer
Most Andromeda content blends Meta's statements with agency folklore. Here is the split.
| Claim | Source | How to treat it |
|---|---|---|
| Andromeda is the retrieval stage, narrowing tens of millions of ads to a few thousand per person | Meta engineering and business posts | Fact |
| The 2026 version scores every eligible ad with specialized expert models | Meta, October 2026 | Fact |
| Supply more creative variety so the system has more to test and learn from | Meta, October 2026 | Official guidance |
| Launch new ads weekly, keep at least 20 diversified ads in Advantage+ shopping campaigns, and put 20 to 30% of budget into testing new ads | Meta, November 2024 | Official guidance, written before the rename to Advantage+ sales campaigns |
| Too many ads at once means fewer exit learning, so cut ads per ad set but keep diverse assets | Meta Business Help Center | Official guidance |
| Creative that is too similar may be delivered to the same audiences | Meta small business guidance | Official guidance |
| Meta groups similar ads under one "entity ID" | Practitioners | Plausible, not documented by Meta |
| Ads must be 70% visually different, or you need exactly N ads per ad set | Practitioners | No Meta source, a heuristic at best |
| Creative is the new targeting | Practitioners | Directionally right, overstated |
Two of Meta's statements look like they conflict. One says feed the system more creative variety. The other, Meta's Help Center page on managing ad volume, says "too many ads can result in worse performance" and tells you to "decrease ads per ad set, but maintain diverse creative assets per ad set." They only conflict if variety means ad count. Read together, they say: more distinct ideas, fewer redundant ad objects. That distinction is the whole strategy.
Why diversity beats volume at the retrieval stage
Retrieval matches ads to people. If ten of your ads are the same idea in ten outfits, they look like one candidate to the people that idea suits. Meta's small business guidance on creative differentiation says as much: "When creative is too similar, our system may not recognize it as different and could deliver it to the same sets of audiences." How retrieval scores similarity is not public, so anything beyond that is inference, but the pattern buyers report fits it: spend piles onto one or two ads, reach stalls and frequency climbs.
The same guidance says that overlap leads to a Creative limited or Creative fatigue status, and Meta's fatigue help page defines both by cost: limited when cost per result is above your past ads but under double, fatigue at double or more. The recommended fix is a new image or video that is "materially different from the original creative." A new headline on the same video does not meet that bar.
Concept versus variation
A concept is a distinct answer to three questions: who is this for, what problem or desire does it speak to, and what does the viewer see. A variation keeps all three and changes something cosmetic. The fastest test: if two ads fit the same one sentence brief, they are one concept.
- Usually materially different: a new persona, a new problem or motivation, a new format (founder talking head versus UGC demo versus static comparison), a new visual world, a new type of proof, a different creator.
- Usually not: headline swaps, button or color changes, a new first frame on the same edit, the same video cropped to another ratio, a price callout added to the same image.
Variations still have a job: once a concept wins, they make good iteration tests. They just do not count toward diversity, so never let 30 variations convince you that you launched 30 ideas. They also burn out together when the underlying concept tires, which is how creative fatigue catches accounts that thought they had a deep bench.
How many distinct concepts to ship each week
Meta does not publish a concept count per ad set, so size it from budget rather than folklore. Meta's November 2024 guidance, three steps to optimize performance with creative diversification, recommends putting 20 to 30% of your overall ad budget toward testing new ads. Pair that with a working rule of thumb (ours, not Meta's): a concept needs roughly two to three times your target CPA in spend before you can call it.
Concepts per week = weekly testing budget ÷ (2.5 × target CPA)
- Example: a $300 a day account with a $40 target CPA spends $2,100 a week, so testing gets $420 to $630. At $100 per concept, that funds 4 to 6 new concepts a week.
- Example: a $3,000 a day account with a $60 target CPA has $4,200 to $6,300 a week for testing. At $150 per concept, that funds 28 to 42 concepts, more than most teams can produce. Production is the cap, not budget.
- Example: an $80 a day account with a $50 target CPA has $112 to $168 a week for testing, enough for about one concept. Run one new concept a week rather than three starved ones.
These are minimums for reading a concept, not quotas. A concept test at this spend is a directional read on cost and early signals such as hook rate, not a statistically clean verdict. The same Meta guidance also recommended keeping at least 20 diversified ads in Advantage+ shopping campaigns (now Advantage+ sales campaigns). Read that as distinct concepts, and treat 20 as a roster to grow toward rather than a day one requirement. More sizing detail lives in how many creatives to test per ad set.
What this means for consolidation
Concept diversity only helps if each concept can collect delivery, which pushes structure toward fewer, broader ad sets: one consolidated scaling campaign where proven concepts live, plus a testing lane where new concepts get planned spend. Package the ratio versions and copy options of one concept inside a single ad (Meta's ad volume page notes one ad can hold up to 10 creative assets) rather than publishing each separately. The full blueprint, with budget tables, is in our Meta ads account structure guide for 2026.
A weekly creative production and launch cadence
Meta's 2024 guidance says the best creative diversification leaders "routinely launch new ads on a weekly basis," with a weekly meeting between buyers and the creative team to review what is working. Here is that loop made concrete, built so every batch gets seven full days before anyone judges it.
| Day | Job | Output |
|---|---|---|
| Monday morning | Read the batch that has now run seven days, at the concept level, then switch on the batch built last Friday. Check the Delivery column for Creative limited or Creative fatigue on scaling ads. | Winners, losers and a list of tired concepts |
| Monday afternoon | Move winners into scaling. Brief next week's batch. | A brief with one line per concept: persona, problem, angle, format, proof |
| Tuesday to Thursday | Produce, delivering every concept in the ratios you run, usually 9:16 and 4:5. | Finished files, named by concept |
| Friday | Diversity audit, naming, QA, then launch the batch paused into the testing lane. | A paused batch ready for Monday |
The brief mix
A batch of only new ideas learns fast but wins rarely. A batch of only iterations wins more often but stops learning. A split that works for most accounts (again our rule of thumb) is about 60% net new concepts and 40% new executions of proven angles, where a new execution means the same angle in a different format, persona or creator, not a new headline. Write the brief at the concept level so production cannot quietly turn five ideas into five color swaps. Test design itself is covered in our creative testing framework.
The diversity audit
Before launch, lay the batch out by angle and format. If 18 files collapse into four angles, you are launching four concepts, and the brief failed, not the media buying. Most teams skip this step because it is tedious by hand, and it is exactly where retrieval rewards discipline.
Why launching becomes the bottleneck
Diversity multiplies operational work. Take a modest weekly batch: six concepts, each with a 9:16 and a 4:5 version, launched into two test ad sets. Packaged properly, that is 12 ads, each needing placement assets, a concept name, UTMs, the right ad set and deliberate enhancement settings. Built by hand in Ads Manager at a few minutes per ad, that is roughly an hour of identical clicking every week, which is where naming drifts and toggles get missed. Add a second brand and the cadence quietly dies.
That is the job Volume Creatives was built for. AI creative grouping clusters a batch by product, angle, hook, format, creator and style, so the diversity audit takes seconds: you see at once whether 18 files are 18 ideas or four. Groups arrive with suggested names, nothing is saved until you approve, and a whole angle can go straight to the bulk launcher, which applies your naming and UTM rules, opts out of enhancements and publishes the batch paused (one recorded launch took 44 seconds, start to publish). The bulk uploader also groups the ratio versions of one concept into a single ad with placement asset customization instead of three competing ads, which is the packaging Meta's ad volume guidance asks for.
One caution: any combination engine will happily multiply cosmetic variants into hundreds of ads. Run the numbers in the ad combination calculator first and build the matrix from concepts, not from every headline you could write.
The Andromeda checklist
Before your next batch goes live
- Every ad maps to a distinct concept with its own one sentence brief.
- Ratio versions and copy options of one concept live inside one ad, not as separate ads.
- New concepts launch into a testing lane, so stable scaling ad sets are not reset by new ads.
- Each concept has planned spend of roughly two to three times target CPA before anyone judges it.
- Advantage+ creative settings are deliberate, so the concept you tested is the concept that ran.
- Names encode concept, angle and format, so results roll up by idea rather than by ad ID.
None of this requires guessing what Andromeda does internally. Meta has said enough: retrieval decides who your ads can reach, distinct creative gives it more to work with, and redundant ads slow learning. The edge goes to teams that turn that into a weekly habit. If the weekly launch is what slows you down, group and launch your next batch with Volume Creatives: your first 10 ad uploads are free with no card, and pricing starts at $49 a month when you are ready to run the cadence every week.
FAQ
What is Meta Andromeda in simple terms?
It is the first stage of Meta's ad delivery. For each person, Andromeda picks a shortlist of a few thousand ads from tens of millions of eligible ones, and only those move on to ranking and the auction. If your ad does not make the shortlist, it cannot win that impression.
Is Andromeda 2.0 an official Meta update?
No. Andromeda 2.0 is a practitioner label that Meta does not use. Meta's October 28, 2026 posts describe the next generation of Andromeda, which routes ads to specialized expert models and scores every eligible ad. Meta says it reaches accounts automatically, with no setting to turn on.
How many creatives should I run after the Andromeda update?
Meta does not set a number per ad set. Its 2024 guidance recommended at least 20 diversified ads in Advantage+ shopping campaigns and 20 to 30% of budget for testing new ads. A practical way to size your weekly batch is testing budget divided by about 2.5 times target CPA, counted in distinct concepts rather than variations.
Does Andromeda mean interest targeting no longer matters?
It matters less. With broad targeting and Advantage+ audience, retrieval finds people for each concept, so creative carries much of the targeting signal. Keep hard limits such as location, minimum age and exclusions, and let distinct concepts do the work interests used to do.
Do headline and color variations count as creative diversification?
Not in a meaningful way. Meta's fatigue guidance asks for a new image or video that is materially different, and its creative guidance warns that similar creative may reach the same audiences. Headline and color swaps are useful iteration tests on a proven concept, but they do not add new concepts.
Media buyers who run client accounts on Meta every week
Written by the team behind Volume Creatives, the bulk Meta ad launcher. We build the tooling we use to run client ad accounts every day.
Launch your next test in one click.
Volume Creatives bulk launches hundreds of Meta ads with enhancements off, naming and tracking applied on every one.
Start freeFirst 10 uploads free. No card required.