# Meta Andromeda: An Operational Playbook for Advertisers (2026)

**Short answer:** Meta rebuilt ads delivery around large AI systems — Andromeda rebuilt the retrieval stage that picks a few thousand candidate ads out of tens of millions, and GEM plus the Adaptive Ranking Model rebuilt the ranking stack behind it. None of it is a setting you can toggle. What you control is what the system retrieves *from*: how many genuinely distinct creative concepts exist in your account, how cleanly the account is structured, and how deliberately you rotate fatigued concepts out. Every platform claim below is cited to Meta's own engineering disclosures or documentation — and there is an explicit section on the numbers that do **not** exist (there is no "creative similarity score" in Ads Manager).

This is the operational companion to our creative feature models essay (`/blog/creative-feature-models-beyond-the-winning-ad`), which makes the measurement case — why the "winning ad" is a weak unit of learning once delivery is automated. This post answers the day-to-day question: what do you actually do differently in an account now?

## What Andromeda actually is (and what it is not)

Meta's ads delivery is a multi-stage recommendation system. Retrieval is the first step: in Meta's words, it selects "from tens of millions of ad candidates into a few thousand relevant ad candidates," which the ranking models then score for the specific person and context.

![Schematic of Meta's multi-stage ads delivery: a wall of candidate ad creatives is narrowed at the retrieve stage, ordered at the rank stage, and one ad is matched to a specific person. Counts are illustrative — the real retrieval stage narrows tens of millions of candidates to a few thousand.](/images/blog/articles/andromeda-playbook/andromeda-playbook-article-retrieve-rank-match-pipeline-16x9-2400x1350.webp)

Andromeda is Meta's redesign of that retrieval stage, announced December 2, 2024 and deployed across Facebook and Instagram. Published results: a **+6% recall improvement** in the retrieval system, a **+8% ads quality improvement** on selected segments, and a roughly **10,000× increase in model capacity** for personalization at that stage. (Source: Engineering at Meta, "Meta Andromeda: Supercharging Advantage+ automation with the next-gen personalized ads retrieval engine," December 2, 2024 — engineering.fb.com.)

Two later disclosures fill in the ranking side:

| System | Stage | What Meta disclosed | Reported results |
| --- | --- | --- | --- |
| Andromeda (Dec 2024) | Retrieval | Personalized retrieval engine selecting a few thousand candidates from tens of millions | +6% retrieval recall; +8% ads quality on selected segments; ~10,000× model capacity |
| GEM (Nov 2025) | Ranking foundation | LLM-inspired ads foundation model trained across thousands of GPUs | Meta reports a 5% increase in ad conversions on Instagram and 3% on Facebook Feed in Q2 |
| Adaptive Ranking Model (Mar 2026) | Ranking inference | Request-aware routing serving trillion-parameter-scale ranking under ~100 ms latency bounds | +3% ad conversions and +5% CTR for targeted users after the Instagram launch |

(Sources: Engineering at Meta — "Meta's Generative Ads Model (GEM): The Central Brain Accelerating Ads Recommendation AI Innovation," November 10, 2025; "Meta Adaptive Ranking Model: Bending the Inference Scaling Curve to Serve LLM-Scale Models for Ads," March 31, 2026.)

The direction of travel is explicit: Reuters, citing the Wall Street Journal (June 2, 2025), reported Meta aims to let brands fully create and target ads with AI by the end of 2026 — provide a product image and a budget, and the system generates the creative, picks the audience, and suggests spend.

What Andromeda is **not**: a campaign setting, a bid strategy, a new attribution model, or a metric. There is nothing to enable and nothing to opt out of.

## Why this changes what you feed the system

Two facts from Meta's own disclosure do the work:

1. The retrieval stage is now personalized at enormous capacity — the 10,000× model-capacity increase exists so retrieval can make finer-grained distinctions about which ads suit which person and context.
2. Meta states the goal directly: "Increased ad diversity can improve people's experience with ads and drive better advertiser outcomes." The same post notes that more than a million advertisers used Meta's generative AI tools to create over 15 million ads in a single month.

The operational inference (ours, clearly labeled — not a Meta quote): **a retrieval system built to match distinct ads to distinct people can only exploit distinctiveness you actually give it.** Ten near-identical variations occupy one point in the system's map of your account; ten genuinely different concepts occupy ten.

This inference is backed by Meta's advertiser-facing guidance too: Performance 5 — Meta's five best practices for direct-response campaigns — names creative diversification as a pillar, alongside account simplification, automation, data quality, and results validation (Meta Blueprint, "Maximize campaign results with Performance 5").

## The playbook

### 1. Diversify concepts, not crops

![Side-by-side comparison for one skincare product: five near-identical bottle-on-stone shots on the same backdrop (one concept in five files) versus seven genuinely different creatives — application demo, testimonial-style portrait, moody lifestyle silhouette, results visual, comparison layout, and outdoor lifestyle imagery.](/images/blog/articles/andromeda-playbook/andromeda-playbook-article-variants-vs-concepts-3x2-2400x1600.webp)

The unit that matters is the **concept** — a distinct combination of hook, value proposition, proof mechanism, and messenger — not the asset count. Resizing one video into three placements and swapping two headlines produces five ads and one concept. What counts as actually diverse: a pain-led demo, a curiosity-gap UGC piece, a quantified-outcome static, a founder-story video, an objection-handling comparison — different hooks, proof, and messengers per concept. Meta's Performance 5 frames diversification on exactly these two axes: different creative concepts and different formats. Enumerate the distinct buyer motivations in your funnel and give each at least one concept that speaks to it natively (production side: `/blog/creative-diversification-how-to-create-high-converting-targeted-ad-creative`).

### 2. Keep the account structure boring

Account simplification is the first item in Meta's Performance 5, before any creative advice. Delivery systems learn per optimization unit; a fragmented account splits conversion signal into thin, noisy slices. Consolidation and diversification are the same strategy viewed from two sides: consolidated structure gives the system one strong learning surface, and a diverse concept portfolio gives it meaningful choices within that surface. Express audience hypotheses as concepts, not as duplicated ad sets with the same creative inside.

### 3. Run fatigue on a cadence, not a crisis

Nothing about Andromeda repeals creative fatigue: repeated exposure still decays engagement, and Meta's official Ads Manager statuses — "Creative limited" and "Creative fatigue" — still key off cost-per-result deterioration against your own past ads, with the full-fatigue label firing only once cost per result has doubled (Meta Business Help Center, "Creative fatigue recommendations in Meta Ads Manager"). Those statuses are lagging confirmations. The detection thresholds, worked math, and refresh-cadence framework live in the dedicated fatigue playbook (`/blog/creative-fatigue-in-meta-ads-detection-and-management-strategies`).

What the Andromeda era changes is the replacement discipline:

1. **Manage fatigue at the portfolio level.** The question is not "is this ad tired?" but "which concept is saturating, and what genuinely new concept replaces it?" A re-crop of the fatigued winner hands the retrieval system the same candidate in a new file.
2. **Keep a standing concept pipeline.** Reactive refreshes produced after the status appears ship near-duplicates under deadline pressure. A monthly cadence of net-new concepts makes honest diversity sustainable.
3. **Retire combinations, not ingredients.** When a concept fatigues, ask which of its features (hook, proof, messenger, context) saturated and which still carry — the feature-level ledger from the companion essay.

![Creative rotation as a continuous loop: a tray of five distinct fresh concepts feeds delivery; a worn, faded creative is retired; its durable features — hook, proof, messenger, format — are carried forward as tokens into a renewed concept that re-enters the tray.](/images/blog/articles/andromeda-playbook/andromeda-playbook-article-creative-rotation-loop-3x2-2400x1600.webp)

### 4. Keep your measurement honest while delivery personalizes

As delivery gets smarter, naive readouts get less trustworthy: when the system routes each ad to the pocket of people it predicts will respond, the performance gap between two ads reflects creative quality *and* routing. Tag every concept's features when it ships; when you need a clean causal read, run a deliberate experiment (`/blog/how-to-run-a-meta-creative-test-that-actually-proves-something`) and size it first with the creative testing calculator (`/resources/tools/calculators/creative-testing-calculator`). Performance 5's remaining pillars — data quality and results validation — are the unglamorous half of the playbook: a personalized delivery system optimizing toward a mis-instrumented conversion event personalizes toward the wrong thing at scale.

## The numbers we deliberately did not print

- **There is no "creative similarity score" in Meta Ads Manager.** Meta has shipped no similarity metric, no diversity score, and no Andromeda-branded reporting surface. The documented delivery statuses ("Creative limited," "Creative fatigue") are cost-per-result definitions, not similarity measurements.
- **Meta has published no threshold for "how different" two creatives must be** to count as distinct to the retrieval system, and no target concept count per ad set. Any hard percentage you encounter is an extrapolation.
- **The cited platform numbers are Meta's own aggregate claims** (+6% recall, +8% ads quality on selected segments, the GEM and ARM conversion lifts). They describe system-level improvements Meta measured across its inventory — not a promise about your account.

## FAQs

**What is Meta Andromeda?** Andromeda is Meta's personalized ads retrieval engine, deployed on Facebook and Instagram and announced in December 2024. Retrieval is the first step in Meta's multi-stage ads recommendation system: it selects a few thousand relevant candidate ads from tens of millions before the ranking stages score them. Meta reports a 6% recall improvement and an 8% ads-quality improvement on selected segments.

**Do advertisers need to do anything to enable Andromeda?** No. It is delivery infrastructure, not a campaign setting. What you control is the input: creative portfolio breadth, conversion signal quality, and account structure.

**Is there a "creative similarity score" in Meta Ads Manager?** No. Meta has not shipped a similarity score, a diversity score, or any Andromeda-branded metric, and has published no numeric similarity thresholds. Treat any hard similarity percentage you read as the author's extrapolation.

**Does Andromeda mean I should run more ads?** More distinct ads, not more near-duplicates. Meta's stated position is that increased ad diversity improves both people's experience and advertiser outcomes; its Performance 5 best practices include creative diversification alongside account simplification.

**How does Andromeda change creative fatigue management?** The mechanics are unchanged — Meta's "Creative fatigue" status still fires only after cost per result doubles against your own history. What changes is the replacement discipline: a refresh needs a genuinely new concept for the system to route, not a re-crop, and rotation should run as a portfolio cadence rather than a panic response.

**How many creative concepts should I run per ad set?** Meta has not published a number, and this post will not invent one. Cover your funnel's distinct buyer motivations with genuinely different concepts, keep enough headroom to retire a fatigued concept without starving delivery, and size the portfolio to what your team can measure.

## Related Resources

- [Creative Feature Models: Beyond the Winning Ad](/blog/topics/creative-analytics/creative-feature-models-beyond-the-winning-ad) - The measurement-layer companion essay
- [Creative Fatigue in Meta Ads](/blog/creative-fatigue-in-meta-ads-detection-and-management-strategies.md) - Detection thresholds, worked math, and refresh cadence
- [Creative Diversification](/blog/creative-diversification-how-to-create-high-converting-targeted-ad-creative.md) - The production side of concept diversity
- [Creative Testing Budget Calculator](/resources/tools/calculators/creative-testing-calculator.md) - Size a creative contrast before running it
- [Creative Quality Grader](/resources/tools/analyzers/creative-quality-grader.md) - Structured pass over hooks, formats, and proof
- [Creative Fatigue Glossary](/resources/glossary/creative/creative-fatigue.md) - Learn about creative fatigue concepts
- [Meta Advantage+ Glossary](/resources/glossary/platform/meta-advantage-plus.md) - The campaign-automation layer above the delivery stack

## Sources

- Engineering at Meta, "Meta Andromeda: Supercharging Advantage+ automation with the next-gen personalized ads retrieval engine," December 2, 2024. https://engineering.fb.com/2024/12/02/production-engineering/meta-andromeda-advantage-automation-next-gen-personalized-ads-retrieval-engine/
- Engineering at Meta, "Meta's Generative Ads Model (GEM): The Central Brain Accelerating Ads Recommendation AI Innovation," November 10, 2025. https://engineering.fb.com/2025/11/10/ml-applications/metas-generative-ads-model-gem-the-central-brain-accelerating-ads-recommendation-ai-innovation/
- Engineering at Meta, "Meta Adaptive Ranking Model: Bending the Inference Scaling Curve to Serve LLM-Scale Models for Ads," March 31, 2026. https://engineering.fb.com/2026/03/31/ml-applications/meta-adaptive-ranking-model-bending-the-inference-scaling-curve-to-serve-llm-scale-models-for-ads/
- Reuters, "Meta aims to fully automate advertising with AI by 2026, WSJ reports," June 2, 2025. https://www.reuters.com/business/meta-aims-fully-automate-advertising-with-ai-by-2026-wsj-reports-2025-06-02/
- Meta Business Help Center, "Creative fatigue recommendations in Meta Ads Manager." https://www.facebook.com/business/help/1346816142327858
- Meta Blueprint, "Maximize campaign results with Performance 5." https://www.facebookblueprint.com/student/path/253157-performance-5