Meta Andromeda: What It Is and How to Advertise for It (2026)

Meta Andromeda is Meta's machine-learning system for the retrieval stage of ad recommendation. Retrieval is the first step in a multi-stage process: it narrows tens of millions of eligible ads to a few thousand relevant candidates, then later ranking models estimate value and determine which ads can be shown. (Engineering at Meta, December 2, 2024)

That definition matters because much of the advice about “the Andromeda algorithm” confuses retrieval, ranking, the auction, and Advantage+ campaign products. Andromeda does not write your strategy or guarantee a winner. It helps Meta choose a smaller, more personalized candidate set at enormous scale.

For advertisers, the useful response is not to chase a secret setting. It is to give the delivery system clear conversion goals and genuinely varied creative inputs, then watch whether spend and results concentrate too heavily on a small part of the portfolio. This practical guidance is interpretation based on Meta's system descriptions and its own creative-diversification guidance, not a claim that Meta publishes one mandatory account structure. (Meta for Business, April 22, 2025; Meta Blueprint, May 15, 2025)

What is Meta Andromeda?

Andromeda is a personalized ads retrieval engine introduced by Meta in the second half of 2024. Meta built it through joint work on machine-learning models, feature representation, training, indexing, inference, and hardware including NVIDIA Grace Hopper and Meta's MTIA accelerators. Its job is to retrieve promising ads fast enough for later stages to rank them. (Engineering at Meta, December 2, 2024; Meta Q2 2025 prepared remarks, July 30, 2025)

Before Andromeda, Meta says retrieval relied on isolated model stages and many rule-based heuristics, which limited personalization and made end-to-end optimization harder. Andromeda replaced much of that complexity with a custom deep neural network and a jointly trained hierarchical index. The index focuses computation on relevant branches rather than comparing every candidate in the same way. (Engineering at Meta, December 2, 2024)

Andromeda is part of a larger ads stack. Meta's Q2 2025 remarks distinguish it from GEM, a later-stage ranking system that works after retrieval. Meta Lattice is another model architecture used across objectives, surfaces, and ranking stages. (Meta Q2 2025 prepared remarks, July 30, 2025; Meta AI, May 11, 2023)

What Meta says Andromeda changed

Much greater retrieval-model capacity

Meta says Andromeda's customized network enabled a 10,000-fold increase in model capacity with sublinear inference cost. It uses interaction features to model more complex relationships among people, products, services, and ads. This is a capacity statement about the retrieval model, not a promise that an advertiser's performance will improve by the same factor. (Engineering at Meta, December 2, 2024)

Better retrieval quality in Meta's reported tests

In its launch article, Meta reported a 6% recall improvement in retrieval and an 8% ads-quality improvement on selected segments. In Q2 2025, Meta said Andromeda enhancements expanded coverage to Facebook Reels and drove nearly 4% higher conversions on Facebook mobile Feed and Reels. In Q3 2025, Meta reported a 14% ads-quality increase on Facebook surfaces after combining models across retrieval and early-stage ranking. These are Meta-reported results for specific deployments, not universal advertiser benchmarks. (Engineering at Meta, December 2, 2024; Meta Q2 2025 prepared remarks, July 30, 2025; Meta Q3 2025 prepared remarks, October 29, 2025)

An architecture built for more eligible creative

Meta says Advantage+ automation and generative creative tools increase the volume of eligible ads. Andromeda's hierarchical index is designed to handle that growth while improving retrieval precision and recall. Meta for Business connects Andromeda with greater creative volume and diversity for personalization. Neither source says that uploading duplicates is beneficial. (Engineering at Meta, December 2, 2024; Meta for Business, April 22, 2025)

Continued engineering investment through 2026

Andromeda is not a one-time switch that Meta stopped developing. In April 2026, Meta reported that its KernelEvolve system improved Andromeda inference throughput on NVIDIA GPUs by more than 60%. That is infrastructure throughput, not a 60% advertiser-performance claim. (Engineering at Meta, April 2, 2026)

What Andromeda means for advertisers in practice

The following is advertiser interpretation, grounded in Meta's technical posts and Blueprint guidance.

Creative diversity gives retrieval more useful choices

Meta for Business defines creative diversification as a range of ads with different themes, messages, and visuals for different audience segments. Meta Blueprint adds signals such as format, messaging, persona, and hook type, and says insufficient diversity may limit reach and efficiency. The practical conclusion is to build ads around distinct ideas, not just export the same idea with a new background color. (Meta for Business, April 22, 2025; Meta Blueprint, May 15, 2025)

A diverse test might vary the customer problem, promise, proof, persona, narrative, and format. For example, one concept can demonstrate the product, another can answer an objection, and another can tell a customer story. Each can then be adapted to a static, short video, or creator-led format. The numbers in that example are a planning device, not a Meta requirement.

Use our Facebook ads creative best-practices guide for execution details and the creative testing framework for a repeatable test plan.

Audience micro-targeting is a smaller lever in automated setups

Meta says Advantage+ can automate audience creation, budget allocation, placement, and creative generation. Its Q3 2025 remarks say Advantage+ sales, app, and lead flows can automatically choose criteria such as who sees an ad and where it appears. An advertiser can therefore test broader delivery where the objective, geography, exclusions, economics, and legal constraints allow it. (Engineering at Meta, December 2, 2024; Meta Q3 2025 prepared remarks, October 29, 2025)

This does not mean every audience control is obsolete. Retargeting logic, customer exclusions, geographic limits, regulated-category rules, and business constraints still matter. The interpretation is to avoid creating tiny segments without a clear reason.

Let the system explore, then judge business outcomes

If each concept is isolated in a separate low-budget ad set, the system has fewer opportunities to compare eligible creative within the same delivery context. A consolidated test can give retrieval and ranking more options, provided the ads share the same objective and economic target. This is an operating interpretation, not an official instruction to use one campaign or one ad set.

Do not expect equal spend across ads. Delivery systems choose among candidates. Evaluate the portfolio on conversions, cost per result, and incremental learning, while also inspecting whether a promising concept received enough delivery to produce a useful read.

What Meta Andromeda does not mean

Myth 1: Andromeda is the whole ads algorithm

It is not. Meta defines it as retrieval, the first stage that supplies candidates to later ranking models. GEM and other ranking systems operate after retrieval. (Engineering at Meta, December 2, 2024; Meta Q2 2025 prepared remarks, July 30, 2025)

Myth 2: Targeting no longer matters

Andromeda improves personalized retrieval, and Advantage+ automates more delivery choices. Neither statement removes objectives, locations, exclusions, policies, or the conversion signal an advertiser supplies.

Myth 3: More files always produce better results

Meta connects Andromeda with a growing volume of eligible creative, while its business guidance defines diversity through different themes, messages, and visuals. Twenty near-identical crops may add files without adding meaningful choice. (Engineering at Meta, December 2, 2024; Meta for Business, April 22, 2025)

Myth 4: There is one required “Andromeda structure”

Meta's public technical material explains the system, not a universal campaign-count formula. Structure should follow distinct objectives, markets, budgets, constraints, and the questions your test needs to answer.

Myth 5: Andromeda fixes weak offers or measurement

Retrieval can select among eligible ads. It cannot make an unclear offer persuasive or turn the wrong conversion event into a useful business outcome. If browser tracking is part of the problem, use the Meta Pixel Helper guide before drawing conclusions from campaign results.

A practical account and creative-testing structure

This framework is Adship's interpretation of the sources, not a Meta-mandated template.

1. Define one business question. Decide whether the test is about a new concept, format, audience assumption, or offer.
2. Group compatible economics. Keep ads together when they share an objective, market, optimization event, and acceptable cost. Split when those constraints genuinely differ.
3. Create distinct concepts first. Write a one-sentence job for each ad, such as demonstration, objection handling, comparison, proof, or use case.
4. Adapt concepts into formats. Turn selected ideas into static, short video, creator-led, carousel, or placement-specific executions where relevant.
5. Name ads for analysis. Encode concept and format so reporting can separate real diversity from cosmetic variations.
6. Launch with controlled variables. Avoid changing audience, offer, creative, and optimization event at the same moment if you need a clean learning answer.
7. Read outcomes and distribution. Review conversions and cost, plus frequency, spend concentration, and fatigue.
8. Refresh the portfolio. Add a genuinely new angle when a concept weakens rather than multiplying tiny edits.

Meta's business post connects creative diversification to broader automated testing, while Blueprint connects testing structure with learning and delivery efficiency. (Meta for Business, April 22, 2025; Meta Blueprint, May 15, 2025)

Signals to watch after launch

Frequency

Frequency shows how often the reached audience saw an ad on average. A rising value is context, not a universal stop rule. Read it beside reach, spend, CTR, conversions, and cost per result. Our Facebook ad frequency guide explains the metric in more detail.

Creative concentration

Creative concentration asks whether one or two ads absorb most of a campaign's spend. High concentration is not automatically bad because a winner may deserve more delivery. It becomes a testing concern when the rest of the portfolio never gets a useful opportunity or when the dominant ad begins to weaken.

Fatigue

Fatigue is a pattern, not one number. In Adship, the fatigue workflow reads recent frequency and CTR direction to surface ads that may need a refresh. Use the signal as a prompt to inspect the concept and audience rather than as proof that Andromeda “punished” an ad. See the creative fatigue guide for the broader diagnostic process.

How Adship fits an Andromeda-ready workflow

Adship's Canvas creates copy, image, video, and actor-based variations inside one creative workspace. The launch flow can assign creatives and copy across Meta targets. The Facebook ads agent can inspect account signals and stage recommendations for approval.

After launch, Adship's read-only creative-concentration analysis reports top-ad and top-two spend share, purchase share, the number of meaningful creatives, and a concentration classification. Its fatigue detection uses recent frequency and CTR direction. These features do not claim to expose Andromeda's internal decisions. They help an advertiser manage the creative inputs and observable outcomes around the system.

That distinction is useful. You cannot inspect why every candidate entered or left Meta's retrieval set. You can build more distinct concepts, launch them coherently, and monitor whether delivery is narrowing or a creative is tiring.

Meta Andromeda FAQs

When did Meta introduce Andromeda?

Meta says it began introducing the architecture in the second half of 2024 and published its technical announcement on December 2, 2024. (Engineering at Meta, December 2, 2024; Meta Q2 2025 prepared remarks, July 30, 2025)

Is Andromeda a ranking model?

It primarily powers retrieval. Later systems rank the few thousand candidates it selects from tens of millions of potential ads. (Meta Q2 2025 prepared remarks, July 30, 2025)

Is Andromeda the same as Advantage+?

No. Andromeda is back-end retrieval architecture. Advantage+ is Meta's suite of advertiser-facing automation products, which increases automation and the pool of eligible ads. (Engineering at Meta, December 2, 2024)

Does Andromeda eliminate audience targeting?

No. It increases retrieval personalization, while Advantage+ can automate more audience and placement decisions. Advertisers still set objectives and applicable constraints.

How many creatives does Andromeda need?

Meta's cited sources do not prescribe one universal number. Prioritize distinct themes, messages, visuals, personas, hooks, and formats over an arbitrary file count. (Meta for Business, April 22, 2025; Meta Blueprint, May 15, 2025)

Are small creative variations enough?

They can test execution details, but they are not the same as concept diversity. Meta Blueprint frames creative signals across format, messaging, persona, and hook type. (Meta Blueprint, May 15, 2025)

Should I consolidate every campaign?

No. Consolidation is useful only when objectives, markets, budgets, and economics are compatible. Meta has not published one mandatory account structure for Andromeda.

How do I know whether Andromeda selected my ad?

Meta does not provide an advertiser-facing log of each retrieval decision in the cited material. Use delivery, spend, conversions, concentration, and fatigue as observable signals, not as a reconstruction of the internal candidate set.

Sources

• Meta Andromeda: Supercharging Advantage+ automation with the next-gen personalized ads retrieval engine, Engineering at Meta, December 2, 2024
• New AI advancements drive Meta's ads system performance and efficiency, Meta AI, May 11, 2023
• The Creative Advantage: Unlocking the Power of Diversification with Meta Andromeda, Meta for Business, April 22, 2025
• Ask the Experts: Creative Diversification, Andromeda & The Auction, Meta Blueprint, May 15, 2025
• Meta Q2 2025 prepared remarks, July 30, 2025
• Meta Q3 2025 prepared remarks, October 29, 2025
• KernelEvolve: How Meta's Ranking Engineer Agent Optimizes AI Infrastructure, Engineering at Meta, April 2, 2026