Meta Andromeda
What Is Meta Andromeda and Why It Matters for Ecommerce Ads
Meta Andromeda is Meta machine learning system for retrieval in ad recommendation: the stage that decides which ads are even eligible to be shown to you, before any ranking happens. Meta announced it in December 2024 and reports a 10,000x increase in model capacity over the previous generation. For advertisers it moves the leverage decisively: manual audience construction matters less, and the range and quality of your creative matters more. This guide covers what it is, how the retrieval engine works, and what to change in your account.
The short version
- Andromeda is ads infrastructure, not hardware. It is the retrieval stage of Meta ad delivery, announced 2 December 2024.
- Retrieval narrows tens of millions of candidate ads to a few thousand. Lose there and your ad is never ranked or served.
- Meta reports +6% recall, +8% ads quality on selected segments, and a 10,000x increase in model capacity.
- There is nothing to configure. The leverage moved to conversion signal quality and genuine creative range.
What Is Meta Andromeda?
Meta Andromeda is Meta proprietary machine learning system for retrieval in ad recommendation. Meta announced it on 2 December 2024 as its next-generation personalized ads retrieval engine, built alongside the Advantage+ automation suite. Its job is selection: narrowing tens of millions of eligible ads down to a few thousand candidates, which the ranking models then order.
Two clarifications, because both come up constantly. It is not a setting in Ads Manager and there is nothing to enable. And it has no connection to Meta augmented reality glasses: that programme is Orion, with a reported consumer successor codenamed Artemis. Andromeda is ads infrastructure, running underneath campaigns you are already spending on.
How Meta Andromeda's Retrieval Engine Works
Ad delivery has always run in stages, and retrieval is the first one. It exists because ranking every eligible ad for every impression is computationally impossible, so a cheap first pass produces a shortlist and the expensive models work only on that. Historically retrieval was the crude stage: light models and hand-built rules, optimised for speed because it had to run against everything. Andromeda changes what is affordable at that stage, and Meta reports a 10,000x increase in model capacity, a +6% recall improvement, and a +8% ads quality improvement on selected segments.
-
1
Retrieval, not ranking
Ad delivery runs in two stages. Retrieval picks a shortlist from everything eligible, then ranking orders that shortlist. Andromeda is the retrieval half. Meta describes it narrowing tens of millions of candidate ads down to a few thousand before the ranking models see anything, which means an ad that loses at retrieval is never ranked, never auctioned, and never served.
-
2
A deep neural network in place of rules
The previous generation of retrieval leaned on lighter models and hand-built rules because the stage had to be fast above all else. Andromeda replaces that with a deep neural network carrying far more compute complexity and massive parallelism, so the shortlist is now selected with something close to ranking-grade judgement rather than a cheap first pass.
-
3
Hierarchical indexing for creative volume
Advantage+ creative means one advertiser now generates many variants of the same ad, so the candidate pool grows exponentially while the number of distinct advertisers barely moves. Hierarchical indexing handles that by grouping candidates into a tree and evaluating only the most relevant nodes, which cuts inference steps instead of scanning everything.
-
4
Model elasticity and the hardware underneath
Andromeda adjusts model complexity in real time through a segment-aware design, spending heavier computation where personalization pays and less where it does not. Meta reports that elasticity combined with the hierarchical structure delivers a 10x efficiency gain, and the networks are custom-designed for the NVIDIA Grace Hopper Superchip.
Entity IDs Versus Creative Volume
The distinction driving the whole architecture is that two things grew at very different rates. The number of distinct advertisers, products and pages, the entity IDs, grows slowly and predictably. The number of creative assets attached to them has exploded, because generative tools mean one product now carries dozens of variants instead of three. Meta reports more than one million advertisers using its generative AI tools to create over 15 million ads in a single month.
A retrieval system indexed on entities cannot tell those variants apart, so it either treats them as one candidate and wastes the variation, or evaluates each one and collapses under the cost. Hierarchical indexing is the answer to exactly that: group the candidates, evaluate only the promising branches, and keep per-creative resolution without paying per-creative compute. Meta also reports over 100x improvement in feature extraction latency and throughput, and more than 3x in end-to-end inference queries per second, with a further 1,000x increase in model complexity projected as its own MTIA silicon takes over more of the work.
Why the Shift Matters: Creative Strategy Over Targeting
For a decade the skilled part of Meta advertising was audience construction: interest stacks, lookalike percentages, exclusion logic. That skill has been quietly deprecated. A system matching individual people against millions of creatives in real time is not improved by a manual segment describing women aged 25 to 34 interested in yoga. It already knows more than that segment encodes.
What the system cannot generate is the argument. It optimizes distribution of the creative you supply; it does not invent a better reason to buy. That makes creative range the binding constraint, and it is why Nielsen's creative-effectiveness research credits creative, not targeting or media spend, with up to 89% of a digital ad's in-market success and Meta's own research reaches the same conclusion about high-quality creative increasing ad ROI. The cost of getting it wrong keeps rising too, with Facebook cost per lead up almost 21% year over year in 2025, per WordStream's benchmarks.
This is a real loss of control and worth naming as one. You can no longer manually steer who sees what. What you gain is that a genuinely better ad now travels further on its own, because the system finds its audience without you describing it. Sarah Levinger, Founder of Tether Insights, points at where that better ad comes from: “I wish more brands would use the data they have and start analyzing for emotional insights instead of just transactions.” For the wider account view, the Facebook ads strategy guide covers structure and budget alongside this.
Adapting Your Ad Creative for Andromeda
The single hero ad is the format that suffers most. One brilliant execution gives retrieval exactly one thing to match against everybody, and the system has no room to find the person for whom a different angle would have worked. Several genuinely different executions give it range. The word doing the work there is different: five crops of the same photograph are one candidate wearing five outfits, and they add nothing at the feature level.
Different means a different argument. One ad answers the price objection, another leads on the durability proof, a third opens on the problem before the product appears, a fourth is a customer saying it in their own words. Those are four distinct candidates because they are four distinct claims, and that is the diversity the retrieval stage can actually use. Creative strategy covers how to build that set deliberately rather than by accident.
On formats, short-form vertical video carries prospecting because it reads as native to the feed, static and carousel work hardest in retargeting where the buyer already knows you, and creator-style footage covers the gap where a brand voice would be distrusted. Advantage+ will adapt any of these across placements automatically, which is genuinely better than manual placement setup. Hand it the mechanical work and keep the argument with a person. Meta ads templates covers the format specs.
The practical constraint is that four distinct arguments require four pieces of customer evidence, and that is where most teams stall. Producing more variants is easy now; knowing which four objections are worth answering is the part that still needs research. Mirella Crespi, CEO of Creative Milkshake, names the step teams skip: “There is a method to the madness of actually looking through creatives, taking those learnings from your competitor's research, and then turning it into something that you can actually execute. This is what we call the creative analysis stage of research.”
Retrieval rewards genuinely different arguments, and the hard part is knowing which four are worth making. Selzee reads your reviews, comments and campaign performance and returns the ranked objections and hooks behind them. Book a session and we will build a set on your products.
Request a DemoBest Practices and Optimization Checklist
-
1
Fix tracking before you touch creative
Andromeda personalizes against conversion signal, so a broken signal produces a confidently wrong shortlist. Confirm the Conversions API is live alongside the pixel, that deduplication is working, and that your highest-value event is the one being optimized for. Every recommendation below is worthless on top of a bad signal.
-
2
Give retrieval a real spread of candidates
A single hero ad gives the system one thing to shortlist. Ship several genuinely different executions per campaign so retrieval has range to match against different people. This is the one change with a direct mechanical link to how the system works rather than a general best practice.
-
3
Test concepts, not cosmetics
Changing a button colour or trimming two words produces variants that are near-identical at the feature level, so the system treats them as the same candidate and you learn nothing. Test a different argument: a new objection, a new proof, a new buyer. Distinct inputs are the only ones that produce distinct results.
-
4
Run a refresh cadence rather than a rescue
A steady split works better than an emergency rebuild when performance drops. Roughly half your budget on proven winners, about a third on variations of what is working, and the remainder on genuinely new concepts keeps a live pipeline without betting the account on untested work.
-
5
Let Advantage+ automate placement, not judgement
Advantage+ handles placement, format adaptation and audience expansion better than manual setup now. What it does not do is decide what your ad should argue. Hand it the mechanical decisions and keep the message with a person who has read your customers.
-
6
Give changes time to settle
A campaign that is restructured every few days never leaves the learning phase, and the system never accumulates enough signal to personalize well. Set a review window, let it run, then judge. Impatience costs more performance than most creative mistakes.
-
7
Keep a written record of what won and why
The system optimizes delivery. It does not tell you which objection your winning ad answered. Log the angle behind every winner so the next brief starts from evidence instead of a blank page, which is the difference between an angle bank that compounds and one that resets each quarter.
On Refresh Cadence
Treat the roughly 50/30/20 split above as a planning default rather than a rule from Meta, which publishes no such guidance. Its value is behavioural: it forces new concepts into the schedule every cycle instead of only after a drop. Meta Ads Manager surfaces creative fatigue as a formal recommendation, and by the time it does, the replacement work should already be in flight. Facebook ad fatigue covers reading the signals and telling a tired execution apart from a spent angle.
Meta Andromeda vs the Previous Retrieval Generation
The previous generation is described rather than named, because Meta has not published a product name for what came before. The comparison that matters is not raw speed, it is where the advertiser has leverage: the last column is the one to read twice.
| Dimension | Previous retrieval generation | Andromeda |
|---|---|---|
| How the shortlist is chosen | Lighter models plus hand-built rules, optimised for speed over judgement | A deep neural network with far greater compute complexity and parallelism |
| Scale of the candidate pool | Sized for a world where each advertiser ran a handful of distinct ads | Hierarchical indexing built for exponential creative growth from Advantage+ |
| Compute allocation | Broadly uniform cost per request | Model elasticity varies complexity by segment in real time |
| Reported quality gains | Baseline | +6% recall and +8% ads quality on selected segments, per Meta |
| Where the advertiser has leverage | Audience construction, placements, bid strategy | Creative range and conversion signal quality |
Performance figures are Meta own, published 2 December 2024, and describe selected segments rather than every account. The strategic consequence holds regardless of the exact numbers: efficiency gains at retrieval accrue to whoever supplies the best and most varied creative, which is why diversifying arguments beats optimising a single winner.
"Every platform change lands the same way. Teams rush to learn the new mechanics and skip the only input the machine cannot generate, which is a real reason somebody should buy. The accounts that came through this fine were already reading their customers."
Meta Andromeda: FAQ
01 What is Meta Andromeda? +
Meta Andromeda is Meta machine learning system for retrieval in ad recommendation, announced on 2 December 2024. It selects which ads are eligible to be shown, narrowing tens of millions of candidates down to a few thousand before the ranking models order them. It is infrastructure inside Meta ad delivery, not a feature you switch on in Ads Manager.
02 Is Meta Andromeda related to Meta AR glasses? +
No. Andromeda is ads infrastructure. Meta AR glasses prototype is Orion, its reported consumer successor is codenamed Artemis, and the intermediate display glasses are reported as Hypernova. The names get mixed up often, but Andromeda has nothing to do with hardware.
03 What is the difference between retrieval and ranking? +
Retrieval picks the shortlist from everything eligible. Ranking orders that shortlist and decides what actually gets served. Andromeda is the retrieval stage. The practical consequence is that an ad which fails at retrieval never reaches ranking, so it is never auctioned and never seen.
04 Do I need to change my Meta Ads setup because of Andromeda? +
There is no Andromeda setting to configure. What changes is where your effort pays off. Clean conversion signal and a genuine range of creative now matter more than manual audience construction, because those are the two inputs the retrieval system reads.
05 How many creative variations should I run? +
Enough that retrieval has genuinely different candidates to match against different people, which in practice means several distinct concepts rather than a dozen colour variants of one ad. Variants that differ only cosmetically look nearly identical at the feature level and add no range.
06 Does this mean audience targeting no longer matters? +
Broad manual targeting has lost most of its edge, because the system matches at a granularity no manual segment reaches. Targeting still matters for exclusions, for regulated categories, and for genuinely distinct product lines. What has moved is the leverage: creative now carries the decision that targeting used to.
07 How does Selzee fit into this? +
Selzee does not manage your campaigns or interact with Andromeda. It reads your reviews, comments, product data and campaign performance and returns the objections, angles and hooks worth testing, which is how you produce the range of genuinely distinct creative the retrieval system now rewards.
Andromeda decides which ads get a chance. What it cannot do is decide what yours should argue. Selzee reads the reviews, comments, product data and campaign performance you already own and returns the objections, customer phrases and hooks worth testing, with the proof attached. Book a session and we will build a set on your products.
Request a DemoKeep exploring: the Facebook ads strategy guide, Facebook ad fatigue, creative strategy, Meta ads templates, market testing for ads.