What is Meta Andromeda?
Andromeda is Meta's rebuilt ads-retrieval engine: the machine-learning layer that decides, for every auction, which candidate ads even get considered for a given person. Meta announced it in December 2024 with a 6% recall improvement and 8% ads-quality gain on tested segments, and completed the global rollout by October 2025. You cannot toggle it, buy it, or opt out of it. You can only structure your account so it works for you instead of around you.
What does a retrieval engine actually do?
Meta serves ads from a pool of millions of live candidates. Before any ranking or bidding happens, a retrieval step narrows that pool to the few thousand ads worth scoring for this person, this auction. Andromeda rebuilt that step on modern accelerator hardware with much larger models, so the narrowing got smarter: it reads creative content and conversion-event signals rather than leaning on the audience boxes an advertiser ticked. Retrieval quality compounds, because every downstream ranking decision inherits the candidate set.
What changed for advertisers in practice?
Three things, all observable in accounts. Detailed targeting lost its pull: the engine finds buyers from creative and event signals, so interest stacks mostly constrain it. Creative volume gained force: with 10 to 20 live ads the engine has a vocabulary to match people against; with 3 recycled ads it is mute, and fatigue arrives fast (Meta's research shows conversion likelihood dropping about 45% after four exposures). And fragmentation turned from tidy to expensive: consolidated accounts in published panels gained roughly 8 to 10% simply by giving the engine aggregated signal. The full structure we deploy is on the management page.
What is feeding the engine, and why it matters
Andromeda eats your event data. Every Purchase and Lead event, with its hashed customer parameters, is a signal about who converts; retrieval learns your buyer from that stream. Which means a polluted pipeline, duplicate events, match quality at 5, teaches the engine a slightly wrong buyer at industrial speed. That is not a hypothetical: the pet-supplements account in the track record spent months optimizing toward purchases 9.4% of which were counted twice. Clean events first, structure second, creative third; the order is the strategy.
Where is this heading?
Further in the same direction. On its Q2 2026 earnings call Meta credited new GEM ranking models with an 8.3% lift in Facebook ad clicks and a 15.7% conversion uplift in early deployment, and the company has said it wants fully AI-generated, AI-targeted ads available by the end of 2026. Each step moves advertiser control from levers (bids, audiences) to inputs (creative, events, structure). Agencies that sold lever-pulling are converting to input quality, or should be; it is openly what our process is built around.
Why did Meta rebuild retrieval at all?
Scale made the old approach the bottleneck. Meta serves ads to 3.58 billion people a day across its apps, against inventory from an advertiser base last officially counted above 10 million, and it booked $196.18 billion in ad revenue in 2025. At that volume, a 6% improvement in retrieval recall is worth billions, which is why the company moved the workload onto modern accelerator hardware and much larger models rather than tuning the legacy system. Advantage+ products built on this foundation crossed a $75 billion annual revenue run-rate by mid-2026; the engine is not an experiment, it is the business.
What can Andromeda not fix for you?
The inputs it never sees. A weak offer retrieves fine and converts badly. Thin margins make a true 1.88 ROAS unprofitable no matter how well the auction matches you. A polluted event pipeline, as above, gets amplified rather than corrected. And measurement honesty is entirely outside its remit: the engine optimizes toward reported conversions, while whether those conversions were incremental stays your question to answer. Retrieval quality raised the ceiling; it did not move the floor.
What should you do this quarter?
Score your event pipeline before anything else; delivery inherits its errors. Collapse fragmented structures to one to three ad sets per campaign. Build the creative calendar to keep 10 to 20 live ads rotating. Then judge the result on a two-ledger report over at least a learning cycle, roughly 24 days to stability in our accounts, before concluding anything. The engine rewards patience with signal and punishes cleverness with noise.
One habit worth adopting immediately: read your account the way the engine does. Open the creative library and ask what a model could learn about your buyer from those assets alone, with every interest checkbox deleted. If the answer is "we sell to everyone, cheaply," that is what retrieval will find for you. Specific creative for specific buyers is now the only targeting brief that reaches the auction.
Short versions, for forwarding
Did Andromeda change how targeting works?
Effectively, yes. Retrieval now selects candidate ads per person per auction from signals in the creative and your event data, which is why detailed-targeting stacks matter less each quarter. Your creative slate is the targeting input you still control; interest checkboxes mostly narrow what the engine could have found anyway.
Do I need to restructure my account because of Andromeda?
If it is fragmented, yes, and the payoff is measured: panels tracking the consolidation pattern of 1–3 ad sets with 10–20 creatives reported roughly 8–10% gains. Fragmented budgets stall in the learning phase, and the engine aggregates signal better when you stop slicing it into interest-sized pieces.
Is Andromeda the same thing as Advantage+?
No. Andromeda is infrastructure: the retrieval layer deciding which candidate ads enter each auction. Advantage+ is the product family (sales, leads, audience) built on top of it. You experience Andromeda only through the behavior of campaigns; there is no Andromeda toggle in Ads Manager.
Want your account graded against the post-Andromeda structure? That is the first page of the audit.
Book a working session