Launch a dozen versions of one winning ad and Meta will often spend on only a few of them. That is how delivery is designed to behave, not a bug. Andromeda and "creative diversification" became one of 2026's most discussed creative topics, and a lot of folklore grew up around them. This guide separates what Meta has said about Andromeda, creative variety and ad volume from what practitioners report, then turns both into a checklist and a testing structure for accounts spending $100k+ a month. Claims about how Meta works come from Meta's engineering blog, Help Center and developer docs; everything else is labelled as practice.
What Andromeda is, according to Meta
Meta describes Andromeda as its machine learning system for personalized ads retrieval, the first stage of the ads recommendation pipeline. Meta's December 2024 engineering post says this stage selects from tens of millions of ad candidates down to a few thousand relevant ones. Larger ranking models then predict value for the person and the advertiser to choose which ads are finally shown.
The 2024 post says Andromeda runs on NVIDIA Grace Hopper Superchips and Meta's MTIA accelerator, and reports a 6% recall improvement and an 8% ads quality improvement on selected segments. Meta ties the redesign to creative volume: Advantage+ automation and generative AI tools keep increasing the number of eligible ads. Meta's October 2026 engineering post says those tools have helped advertisers create hundreds of millions of new ads, and that Andromeda has delivered a 2.4% increase in conversions on Facebook since 2025. These are Meta's own figures, not independent measurements.
How creative diversity relates to retrieval and delivery
Meta has not published a rule that says "similar ads are grouped and suppressed." The 2024 Andromeda post looks ahead to a model change that, in Meta's words, will deliver "a more diverse set of ad candidates," and says increased ad diversity can improve people's experience and advertiser outcomes. That is a stated direction for retrieval, not a documented similarity penalty. The October 2026 engineering post says nothing about ad similarity.
The Help Center is more concrete about repetition. Meta's creative fatigue article, which covers ad sets with one creative, says that when it assesses whether an audience has seen an ad too many times, it considers all recent exposures of the ad's image or video, including exposures from other campaigns from your Page. When cost per result is above your past ads but under twice as much, the Delivery column can show Creative limited. At twice or more, it shows Creative fatigue. Its first recommendation is to create a new ad with an image or video that is "materially different" from the original. A near-duplicate of a tired asset is not what Meta asks for. Our frequency and fatigue guide covers the diagnostics in more detail.
Ad volume is the other half. Meta's ad volume guidance says that when an advertiser runs too many ads at once, each ad delivers less often, fewer ads exit the learning phase, and more budget is spent before delivery can optimize. Twenty variations of one idea split delivery twenty ways without giving the system more distinct options.
What Meta itself recommends about creative diversification
Meta's Performance 5 page lists creative diversification as one of five pillars and tells advertisers to "develop a variety of ad creative and enable AI" to reach the people most likely to be interested. Meta's October 2026 business post on how its ads system matches ads to people says to supply more creative variety so the system has more to test and learn from. The Marketing API's performance recommendations include MIXED_FORMATS: use multiple creative formats instead of relying on only one, such as a single image or a single video. We cover the wider framework in our Performance 5 guide.
Meta pairs variety with restraint. The ad volume article says to decrease ads per ad set while keeping diverse creative assets in each ad set, and notes that one ad can contain multiple creative assets, up to 10. Meta's learning phase article says to avoid high ad volumes, because the system learns less about each ad when you create many. The same article says testing new creative is essential for improving performance over time. Read together: more distinct ideas, packed into fewer ads, given enough delivery to learn.
What practitioners say, and how much weight to give it
Much of the 2026 discussion uses language Meta has not published. Some agencies describe a "Creative Entity ID": a label they say Meta assigns to look-alike ads so they are treated as one creative. We found no mention of Entity IDs in Meta's Help Center, developer docs or engineering posts. Practitioners also circulate specific similarity percentages above which ads are supposedly grouped; Meta has published no such threshold. Some advertisers report a Creative diversity rating column in Ads Manager, but we found no Meta Help Center page documenting it, so treat it as an unofficial signal.
The practitioner framing is consistent with Meta asking for a materially different image or video: an ad that changes only the headline, color or crop still reuses the same visual. Use it as a heuristic, and use Meta's documented statuses (Creative limited, Creative fatigue, Learning limited) as your actual signals.
A practical diversification checklist
This checklist is media-buying practice, not a Meta rule. The test for each new ad: would a viewer call it a different ad, or the same ad with a new caption? Log which axes each live ad covers so gaps are visible. Our creative volume planning guide covers how many new concepts each spend level needs.
- Concepts: different reasons to buy, such as the problem solved, the outcome, social proof, comparison, price or offer, and founder story
- Formats: single image, video, carousel and collection, plus vertical video sized for Reels and Stories
- Hooks: change the first seconds and the opening line, such as a question, a bold claim, a demonstration or a customer quote
- People: different creators, ages, genders and roles, such as customer, expert or staff, rather than one face across every ad
- Settings: home, outdoors, studio, workplace and in-use shots instead of one backdrop repeated
- Production style: UGC-style phone footage, polished studio work, motion graphics and static design
- Audience angle: messages written for distinct buyer types, since broad targeting relies on the creative to find them
- Asset packing: group related variants inside one ad where the format allows, rather than launching each as its own ad
How this changes your testing structure
Two pieces of Meta guidance shape the structure. Meta's significant-edits article lists adding a new ad to an ad set, and any change to ad creative, as significant edits that send the ad set back into learning. And an ad set usually exits learning after about 50 results in the week after its last significant edit. Drip near-duplicates into a scaling ad set every few days and it keeps resetting, while delivery is split across ads that add no new options.
A structure that fits both, in our practice: keep scaling ad sets stable and add creative in planned batches. Test new concepts against each other, not small variations of one concept. Judge on cost per result over enough conversions to separate signal from noise, and iterate on hooks and edits only after a concept has proven itself. Our creative testing framework covers how to size and judge tests, and our learning phase guide shows how to budget for about 50 results per ad set per week.
When fatigue shows up, Meta's recommendations are to add a new ad with a materially different image or video (keeping the original active may maximize results, Meta notes), expand the audience, or try Advantage+ creative. Meta says Advantage+ creative builds variations from a single image or video and shows each person the version they are most likely to respond to; it is only available for traffic or sales campaigns with a website destination.
What not to do
These habits either repeat the same idea or work against how Meta says delivery learns.
- Don't launch many near-identical variants of one winner and call it a test
- Don't duplicate the same asset into new campaigns to restart delivery, since Meta counts recent exposures of an image or video across your Page's campaigns
- Don't add new ads to a scaling ad set every day, because each new ad is a significant edit
- Don't plan around unpublished similarity thresholds or Entity ID scores
- Don't read early results during learning as final, since Meta says they aren't necessarily indicative of future performance
- Don't try to make reused or rejected creative look different to get around review; changing pixels to dodge policy enforcement is evasion, not diversification
A note on Power Ads
Power Ads supplies Meta agency ad accounts only. We don't run campaigns, build creative or manage testing. The guidance above is for teams that do that work themselves.
Key takeaways
- Per Meta, Andromeda is the retrieval stage that narrows tens of millions of ad candidates to a few thousand before ranking
- Meta recommends varied creative and formats in fewer ads; too many ads means fewer exit learning
- Meta's fatigue fix is a materially different image or video, not a near-duplicate
- Entity IDs and similarity thresholds are practitioner claims Meta has not documented
- Diversify by concept, format, hook, people and setting, added in planned batches
FAQ
What is Meta Andromeda?
Meta describes Andromeda as its machine learning system for personalized ads retrieval, the first stage of its ads recommendation pipeline. It narrows tens of millions of ad candidates to a few thousand, which larger ranking models then score.
Does Meta penalize similar ads?
Meta has published no similarity penalty. It does say repeated exposure to the same image or video, including across your Page's other campaigns, drives Creative limited and Creative fatigue statuses, and that too many ads at once means each delivers less often. Its fatigue fix is a materially different image or video.
Is the Entity ID real?
It is practitioner shorthand for the idea that Meta treats visually similar ads as one creative. We found no mention of it in Meta's Help Center, developer docs or engineering posts, and Meta has published no similarity threshold. Treat it as a heuristic, not a Meta feature.
How many ads should an ad set have?
Meta gives no fixed number. It says to decrease ads per ad set while keeping diverse creative assets, notes one ad can hold up to 10 assets, and warns that too many ads means fewer exit learning. Size the count to what your budget can feed toward about 50 results per ad set per week.
Does Power Ads manage creative or campaigns?
No. Power Ads supplies Meta agency ad accounts only; we don't run campaigns or produce creative.
Sources
- engineering.fb.com/2024/12/02/production-engineering/meta-andromeda-advantage-automation-next-gen-personalized-ads-retrieval-engine
- engineering.fb.com/2026/10/28/ml-applications/meta-andromeda-next-gen-ads-retrieval-gpu-co-design
- facebook.com/business/news/how-metas-ads-system-matches-your-ad-to-the-right-person
- facebook.com/business/ads/performance-marketing
- facebook.com/business/help/2720085414702598
- facebook.com/business/help/1346816142327858
- facebook.com/business/help/112167992830700
- facebook.com/business/help/316478108955072
- developers.facebook.com/documentation/ads-commerce/marketing-api/overview/performance-recommendations
