Meta Andromeda Explained: The 2026 Creative Strategy for Facebook and Instagram Ads

Meta Andromeda is the machine learning system Meta uses for ad retrieval, the first stage of ad delivery, where tens of millions of candidate ads are narrowed to a few thousand for each person before ranking. Meta announced it on its engineering blog in December 2024 and says it was built to handle the growing volume of creatives produced by Advantage+ and generative AI tools. The practical takeaway for 2026: your creative now acts as your targeting signal, so the winning strategy is a steady supply of genuinely different concepts inside simpler, consolidated campaigns, not dozens of near-identical variations spread across many ad sets.

This guide separates what Meta has officially published from what experienced media buyers believe is happening, then turns both into a working plan: how to define a concept, how to structure the account, how to test, and how much creative to plan for each month. If you want the wider picture of the year, start with our Meta ads trends for 2026. For creative mix and refresh cadence basics, see our Meta ads creative strategy framework. This post goes deeper on the Andromeda mechanics and the planning that follows from them.

What Meta has officially said about Andromeda

Meta's engineering team described Andromeda in a post titled "Meta Andromeda: Supercharging Advantage+ automation with the next-gen personalized ads retrieval engine," published December 2, 2024. The confirmed points are:

• Its job is retrieval. Meta describes retrieval as the stage that narrows tens of millions of ad candidates into a few thousand relevant candidates, which later ranking models then score.
• It was built for creative volume. Meta links the system directly to Advantage+ automation and generative AI ad tools, which increase the number of eligible ads. The post says Andromeda introduces a hierarchical index designed to scale to a large volume of ad creatives.
• Reported gains. Meta reported a 6% recall improvement in the retrieval system and an 8% ads quality improvement on selected segments, along with large efficiency gains in model capacity and inference.
• Hardware. Andromeda runs on NVIDIA Grace Hopper Superchips, and Meta said it plans to integrate it with its own MTIA chips and future GPUs.

In a February 6, 2025 Meta for Business announcement, Meta repeated that since Andromeda launched, it has seen an 8% improvement in ads quality, "better personalizing the ads we show from Advantage+ campaigns." That same post is where Meta introduced the streamlined Advantage+ setup, covered below.

What Meta has not published is a rulebook for advertisers. There is no official Andromeda checklist, no published number of creatives you need, and no official definition of how similar two ads must be before the system treats them as the same. Keep that in mind when you read confident claims online.

GEM: the ranking side of the same shift

Andromeda decides which ads are even in the running. Ranking decides which of those wins. In November 2025, Meta's engineering blog described GEM, its Generative Ads Recommendation Model, as its most advanced ads foundation model, built on an LLM-inspired approach and trained across thousands of GPUs.

Confirmed details from that post:

• GEM launched earlier in 2025, and Meta reported a 5% increase in ad conversions on Instagram and a 3% increase on Facebook Feed in Q2.
• GEM works mainly as a "teacher." It transfers what it learns to hundreds of smaller user-facing models through techniques like knowledge distillation and parameter sharing.
• Creative formats and creative representations are among the inputs it uses, though Meta reported no creative-specific results.

For advertisers, the combined message is consistent: Meta's systems are getting better at reading what an ad is about and matching it to the right person. Your manual audience settings matter less. What the ad itself says, shows, and promises matters more.

Advantage+ changes that shape your account structure

The February 2025 announcement made several structural changes that are now part of everyday account management:

• Streamlined setup. Meta began testing a single setup for sales, app, and leads campaigns, so advertisers no longer choose between a manual campaign and an Advantage+ shopping campaign. Campaigns show an Advantage+ "on" label when Advantage+ audience, placements, and campaign budget are used.
• Renaming. Advantage+ shopping campaigns became Advantage+ sales campaigns.
• Advantage+ leads campaigns. Meta reported that in early testing, leads campaigns with Advantage+ on delivered on average 10% lower cost per qualified lead than those with it off.
• Opportunity score. A 0 to 100 score in Ads Manager with recommendations. Meta reported a 5% median decrease in cost per result for advertisers who adopted its recommendations in early tests.

In October 2026, at Advertising Week New York, Meta announced that Ads Creative Studio is rolling out to more advertisers after summer tests, that generating video from existing static assets in Advantage+ creative is now generally available, and that catalog-to-product-video generation is in beta for eligible advertisers. Rollouts are gradual, so availability can differ by account and country. Controls are also narrowing in other areas. Our post on Meta removing placement controls covers what that means for Q4 planning.

What practitioners believe (and how much weight to give it)

The following ideas are widely shared by experienced Meta media buyers. They are practitioner opinion, not Meta documentation. They are useful working models, but treat them as hypotheses to check in your own account.

• Similar ads get grouped. Many buyers believe that ads which look nearly identical, such as the same creator, same setting, and same opening with only a headline change, are treated as one creative for delivery purposes. You will often see this called "Entity ID" clustering. Meta has not published this term or mechanism for advertisers.
• Diversity widens reach. Because retrieval matches ads to people, the belief is that distinct concepts reach distinct pockets of buyers, while near-duplicates compete for the same pocket.
• Fragmentation hurts learning. Splitting budget across many small ad sets with overlapping broad audiences is widely seen as slowing learning and producing noisy results.
• One ad taking most spend is normal. Buyers report that delivery often concentrates on a few ads. This is usually read as the system picking a winner, not as a bug.

These beliefs line up with what Meta has said about matching ads to people at scale. They go further than Meta's own statements, though, so be careful with any specific threshold, such as "you need exactly 15 concepts."

Concepts vs variations: the distinction that matters most

The most useful habit under Andromeda is to separate concepts from variations before you brief any creative.

A simple test: if a stranger saw both ads back to back, would they say "that's the same ad"? If yes, it is a variation. You need both. Concepts find new demand, and variations stretch the life of winners. But in most accounts the gap is too few true concepts, not too few variations.

Five concept axes for ecommerce brands

1. Motivation: save money, save time, look better, feel safer, give a better gift.
2. Persona: first-time buyer, repeat buyer of a competitor, gift buyer, professional user.
3. Proof type: customer review, expert, before and after, side-by-side comparison, data from your own product specs.
4. Format: creator UGC, founder to camera, product demo, static image, carousel, catalog.
5. Funnel moment: problem-aware, solution-aware, ready to buy with an offer.

Changing at least two axes at once is a reliable way to make sure a new ad is a real concept. Our creative studio plans briefs around these axes so production effort goes into new ideas, not reskins.

Account consolidation: a practical structure

Consolidation means fewer campaigns and ad sets, each with enough budget and conversions to learn, with creative doing the segmentation. A common structure for an ecommerce brand in 2026:

1. One main prospecting sales campaign with Advantage+ on, broad audience, Advantage+ placements, and campaign budget. Most of your spend goes here.
2. One testing campaign or ad set with a fixed share of budget, where new concepts get a fair read before graduating.
3. Optional separate campaigns only for real business differences, such as a different country, a different margin product line, or a different conversion goal. Do not split just to test audiences.
4. Customer handling through settings, not structure. Use existing customer definitions and Meta's acquisition settings, such as Customer Lifecycle Strategy, which Meta made available to all advertisers in October 2026, rather than building many retargeting ad sets.

Consolidation works best when tracking is solid. If your Conversions API setup is weak, the system has less signal to learn from no matter how good the creative is. Our guide on reducing Meta ads CPA covers the signal and structure fixes that usually come first.

A testing framework built for Andromeda

Testing under Andromeda is less about proving which audience works and more about learning which ideas the system can scale.

Step 1: Write a hypothesis per concept

Example: "Busy parents will respond to a time-saving angle shown in a real kitchen by a creator." One concept, one hypothesis. This stops tests from turning into random uploads.

Step 2: Launch concepts in small batches

Put three to five new concepts into the testing campaign at once, each with one to three variations. This gives delivery a choice without burying any idea under dozens of near-duplicates.

Step 3: Read early signals, then conversion signals

• Early: thumb-stop or hook rate for video, click-through rate, and cost per landing page view.
• Decision: cost per purchase or qualified lead, and conversion value where you track it.

Set a spend or conversion threshold before launch so you are not making calls on noise. Our creative testing framework covers thresholds and budgets in detail.

Step 4: Graduate, iterate, or retire

• Graduate: move proven concepts into the main campaign, ideally as new ads alongside existing winners, not as replacements.
• Iterate: for concepts with strong early signals but weak conversion, test a new hook, offer, or landing page match.
• Retire: cut concepts that miss both early and conversion thresholds. Record why in a simple creative log.

Step 5: Use Meta's own creative reporting

Where you have access, Ads Creative Studio is Meta's workspace for comparing creative performance and spotting what is working by objective. Use it alongside your own naming conventions so you can report results by concept, not just by ad.

Creative volume planning

Meta has not published a required number of creatives. The right volume depends on budget, average order value, and how quickly your audience tires of an ad. A practical way to plan:

1. Start from budget per concept. Decide how much spend a concept needs to get a readable result in your account, based on your typical cost per purchase. Divide your monthly testing budget by that figure. That is your realistic monthly concept count.
2. Add variations for winners. Plan two to four variations for each concept that graduates, such as new hooks, cutdowns, and aspect ratios.
3. Plan refresh around fatigue, not the calendar alone. Watch rising frequency with falling click-through rate. Those two signals together are the clearest sign an ad is wearing out.
4. Use AI tools for variations, people for concepts. Meta's generative tools can now turn static assets into video and produce variations quickly. They are well suited to extending a concept. The insight behind a new concept still comes from customer research, reviews, and sales conversations.
5. Cover formats and placements. Make sure each concept exists in vertical video for Reels and Stories and in a feed-friendly version. As placement controls narrow, assets that work across placements matter more.

Common mistakes

• Treating 30 headline swaps as 30 tests. If the visuals and opening are the same, you most likely have one idea, not 30.
• Rebuilding audience-split structures. Splitting broad audiences across many ad sets usually spreads budget too thin for the system to learn.
• Turning off a winner too early because spend concentrates. Uneven spend across ads is normal. Judge each ad on cost per result, not share of spend.
• Quoting practitioner rules as Meta policy. Concepts like Entity ID are useful mental models, but they are not published Meta specifications.
• Ignoring signal quality. Weak pixel and Conversions API data limit any creative strategy.
• Letting AI replace research. Generative tools increase output. They do not tell you why customers buy.

Andromeda readiness checklist

• Main prospecting campaign uses Advantage+ audience, placements, and campaign budget
• A separate, fixed-budget testing lane exists for new concepts
• Every new ad is labeled as a concept or a variation in its name
• Each new concept changes at least two concept axes
• Pass and fail thresholds are set before each test launches
• Monthly concept count is based on testing budget divided by cost per test
• Every concept exists in vertical video and a feed format
• Conversions API is set up and deduplicated with the pixel
• A creative log records winners, losers, and the reason for each

If you want a team to plan concepts, produce creative, and run the testing system for you, our Meta ads management service covers strategy, creative production with our studio, and weekly testing for brands in the US and India.

Sources

• Meta Engineering: Meta Andromeda, confirms Andromeda's retrieval role, hierarchical index for creative volume, 6% recall and 8% ads quality gains, and hardware.
• Meta Engineering: Meta's Generative Ads Model (GEM), confirms GEM's role as a foundation model and the reported 5% Instagram and 3% Facebook Feed conversion gains.
• Meta for Business: new Advantage+ AI features (Feb 6, 2025), confirms the streamlined setup, the Advantage+ sales campaigns rename, Advantage+ leads campaigns, and opportunity score.
• Meta for Business: Advertising Week New York 2026, confirms the Ads Creative Studio rollout, generative video updates, and Customer Lifecycle Strategy availability.
• Meta for Business: Ads Creative Studio, confirms the tool's creative performance insights and comparison features.

Frequently asked questions

Is Meta Andromeda an algorithm update I need to opt into?

No. Andromeda is Meta's ad retrieval system, which Meta announced in December 2024 and runs across its ad delivery. There is no setting to turn it on or off; you adapt through creative and account structure.

Did Meta say how many creatives I need for Andromeda?

No. Meta has not published a required number of creatives or concepts. Plan volume from your testing budget and typical cost per result instead of following a fixed rule.

What is an Entity ID in Meta ads?

It is a term used by practitioners to describe the belief that very similar ads are grouped and treated as one creative in delivery. Meta has not published this term or mechanism for advertisers, so treat it as a working model.

How is GEM different from Andromeda?

Andromeda handles retrieval, narrowing millions of candidate ads to a few thousand per person. GEM is Meta's ads foundation model that improves the ranking models that choose among those candidates.

Should small accounts still consolidate campaigns?

Usually yes. Smaller budgets benefit most from fewer ad sets because each one gets more conversions to learn from, with new concepts tested in one fixed-budget testing lane.