How Andromeda and GEM work
Meta serves ads from a pool far too large to score one by one for every impression, so the job is split in two. A retrieval layer narrows the enormous set of eligible ads to a shortlist for this person at this moment. A ranking layer then predicts, for each candidate on that shortlist, how likely the person is to respond and what the advertiser’s bid is worth, and the auction resolves from there. Andromeda and GEM are Meta’s names for parts of that machinery, covering retrieval and the recommendation model used in ranking.
What matters to an advertiser is not the internal architecture but its consequence. The system decides, per impression, which ad and which person fit together, using behavioural signals it holds across its apps. Your inputs are the creative you supply, the objective and conversion event you optimise towards, the budget, and the constraints you set. Everything else is inference you cannot see.
Why Andromeda and GEM matter
They explain a shift most advertisers have already felt. As retrieval and ranking improve, hand-built interest targeting adds less than it once did, because the system can find responsive people from broader pools on its own. That is why broad targeting with strong creative now often beats the narrow audience stacks that used to win.
It also raises the value of creative. When the system’s job is to match an ad to a person, the range of ads you supply becomes the range of people it can reach, which is what people mean when they say creative acts as targeting.
Where the talk about Andromeda and GEM goes wrong
Treat vendor descriptions of these systems carefully. They are Meta’s own account of Meta’s technology, published for investors and press as much as for advertisers, and nobody outside the company can inspect the models or verify the claims. That does not make them false; it makes them unconfirmed.
There is no setting for either. Any advice promising to optimise your account for Andromeda, or claiming that a particular posting pattern feeds GEM, is invention. The controls available to you are the ones in Ads Manager, and they did not change because a model was named.
The opposite error is fatalism, concluding that because the system is opaque nothing an advertiser does matters. Poor conversion signals, thin creative and a badly chosen objective all limit what any ranking model can do with your account, and those are entirely yours to fix.
How to act on it
Compete where you actually have control. Send accurate conversion data, including server-side events where the browser loses them, so the ranking has something reliable to predict against. Supply enough genuinely different creative that the system has real choices to make. Set the objective to the outcome you want rather than a proxy that is simply easier to generate.
Then test rather than believe. If you are told broader targeting now works better, run it against your current structure with real budget and read your own results, because your account’s data is the only evidence about your account. That discipline matters more than any explanation of the model, and it applies to every automated layer in Meta advertising, not only the current one.
