Est. reading time: 5 minutes
Andromeda arrived without a launch event for advertisers, which is fitting, because it isn’t a feature you turn on. It’s the retrieval engine Meta rolled into its ads infrastructure in late 2024, quietly became core plumbing through 2025, and now sits underneath essentially everything your campaigns do, deciding which ads are even eligible to enter each auction before bidding ever happens. Most accounts have been running on it for over a year without changing anything, and that’s the problem. Andromeda rewrote what the delivery system rewards, and playbooks built for the previous machine are leaving performance on the table in ways the dashboard doesn’t label.
What Meta Andromeda actually is
Strip away the branding and Andromeda is a personalized retrieval system. When a user’s feed loads, Meta has an enormous pool of candidate ads and milliseconds to pick contenders, and Andromeda is the AI layer that scans that pool and retrieves the ads most likely to produce value for that specific person, in that moment, before the auction ranks and prices them. It’s the difference between an auction everyone enters and an auction you must be selected for, and the selection runs on predicted relevance computed from your signals and your creative.
That’s why it doesn’t appear as a button in Ads Manager. It’s not a campaign type, it’s the engine under the campaign types, most visibly under Advantage+ sales campaigns, where audience expansion, creative selection, placement, and budget allocation already run as one adaptive system. The walls between prospecting and retargeting, and between interest targeting and broad, keep blurring because Andromeda optimizes on probability of value per person rather than on the static segments advertisers define, and the segments were always a proxy for that probability anyway. The proxy just got replaced by the real thing.
Why the staircase funnel stops mattering
The classic architecture, cold, warm, and hot audiences in separate campaigns with tight windows and stacked exclusions, exists to hand-route people through stages. Andromeda routes them itself, reading behavior and predicted value continuously, which means a nominally prospecting campaign will quietly do retargeting-grade work on high-intent users and a broad campaign will nurture where nurturing is what converts. Hard-coding the stages doesn’t add control anymore. It adds constraints the system has to work around, at your expense.
The reframe that works is organizing campaigns by the outcome each one owns rather than the audience each one contains. Customer acquisition, incremental revenue, reactivation, each an objective the engine hunts across the whole user base to serve, with your remaining structural controls, the existing-customer budget cap chief among them, enforcing the business constraints the algorithm would otherwise optimize straight through. For accounts that plateaued under the old architecture, this consolidation is frequently the unlock, part of the pattern we diagnosed in why most Meta ad accounts hit a ceiling. Fewer campaigns with more signal flowing through each one is exactly the shape a retrieval engine learns fastest from.
Your two remaining levers: signal and creative
With audience construction largely surrendered, targeting becomes a question of what your best customers do rather than who you think they are. Conversion events, purchase values, and on-site behavior are the raw material Andromeda uses to find more people like your most valuable users, including people your persona decks never imagined, which means the real targeting lever is signal quality. Pixel plus Conversions API with clean deduplication, a clear primary conversion, values passed correctly, meaningful micro-events, and strong match quality stop being hygiene items and become the input the entire retrieval computation runs on. Feed the engine messy, duplicated, or mis-firing events and you’re asking it to predict value from noise, the input-quality argument we’ve been making since Training Meta’s AI Faster.
Creative is the other lever, and Andromeda raises its stakes in a specific, mechanical way. A retrieval system matching ads to individuals can only match from what exists in your account, so a brand running three variations of one concept is retrievable for exactly one kind of person, while a brand running fifteen genuinely distinct angles gives the engine fifteen chances to be the right ad for someone. The old discipline of building twenty micro-audiences becomes the new discipline of building twenty angles and offers, problem-first, proof-led, demonstration, discovery, and letting the system match angle to person. It’s the shift from audience hacking to signal and creative engineering, still strategic work, just different muscles, and we’ve watched it play out in accounts like the one in our case study on scaling a paid social channel, where broad setups with strong signal and varied creative did what granular targeting never had.
Rebuilding the playbook, in order
Structure first. Consolidate fragmented ad sets into broader campaigns organized by business objective, replace hyper-granular interest stacks with broad or Advantage+ audiences, and accept that the account should look almost suspiciously simple on the surface, because the complexity moved into the modeling underneath, where it works better than it ever did in your ad set names.
Measurement second, because judging this system on last-click, day-one ROAS misjudges it structurally. A retrieval engine optimizing toward predicted value needs its scoreboard at the level where value actually lands, blended MER against backend revenue, incrementality tests for the causal question, and cohort views of LTV by acquisition source, with internal reporting patient enough to let learning-phase dips resolve before verdicts get issued. Short windows measure the old machine. This one pays out on longer ones.
Then iterate rather than converting in one leap. One consolidated campaign launched alongside the legacy structure, events cleaned, automation extended as it earns trust, creative experimentation run as a standing weekly process instead of a quarterly project. Andromeda didn’t remove your control, it relocated it, from slicing audiences to shaping signals, offers, and creative the engine can amplify, and the advertisers pulling ahead on Meta right now aren’t the ones who found new knobs. They’re the ones who noticed which knobs still connect to anything.
