Andromeda, GEM, and what to change in your ads

First, separate the systems

Andromeda and GEM are names for parts of Meta’s advertising technology. They are not campaign types you need to recreate in your account. Knowing roughly what they do can help you understand a platform update without treating every announcement as an instruction to start over.

Meta describes Andromeda as a retrieval system: it narrows a very large pool of possible ads before later ranking stages make more detailed predictions. The original engineering explanation was published in December 2024. It is not a new 2026 switch that an advertiser has to turn on. ^[1]

Meta’s October 2026 GEM update describes a central foundation model that learns across advertising surfaces and transfers that learning to downstream systems. The announcement discusses platform-level engineering and vendor-reported improvements. It does not establish a guaranteed result for an individual account. ^[2]

That is enough technical context for most commercial decisions. You do not need to reverse-engineer Meta’s models to write an accurate product brief or identify a broken checkout.

Improve the material you give the system

A reasonable practical response is to make your inputs more useful. Start with a correct product catalog, a clear commercial objective, an accurate conversion signal, and creative that communicates different reasons to buy. This is our recommendation, not a claim that we know the platform’s private ranking formula.

Consider two sets of ads. The first contains twelve versions of the same photograph with slightly different text placement. The second includes a product demonstration, a size comparison, an explanation of maintenance, and an honest answer to a common objection. The second set gives a customer more distinct information to respond to.

That does not mean every additional concept deserves budget. A different idea can still be weak, misleading, or irrelevant. Creative variety is useful when it adds a real buying reason. Count the questions you answer, not just the files you upload.

Keep the commercial constraints visible

Automation does not know every practical limit of your business. You may have locations you cannot serve, a product that is running out, a margin that cannot support the proposed acquisition cost, or a sales team that cannot respond to more leads this week.

Write those constraints down and check where they are represented in your setup. Some belong in campaign controls; others belong in the catalog, creative, landing page, or operations plan. The fact that a platform can find more activity does not mean all of that activity is useful.

For a service business, a lower cost per form submission can even create more work without adding sales. Look at contactability, service fit, booking, and completed jobs. Use the later outcomes to judge the value of the initial signal rather than asking the first event to represent the entire business.

Resist the account-wide reset

When performance changes, list plausible causes before making a sweeping structural change. The offer may have changed. The page may be slower. A popular product may be unavailable. Tracking may be recording different events. The customer mix or competitive environment may also have shifted.

Changing campaign structure, creative, budget, and measurement at the same time makes it harder to understand the result. If a change is urgent, record what you changed and why. If it is not urgent, preserve a useful comparison and work through the most likely explanation first.

An engineering announcement can give you a reason to revisit a practice. It is not evidence that a particular reset will rescue your account. Ask the person recommending a change what outcome it should improve, under what conditions, and what would count as evidence that it did not help.

Translate the update into a brief

Here is a practical brief for the next creative round: identify three different customer questions, choose the evidence that answers each one, and produce one clear execution of each. Keep the product and offer accurate. Make the opening understandable with sound off. Check the landing page before launch.

Then decide how you will learn from delivery. Ordinary campaign results can guide operations, while stronger causal claims need an appropriate experiment. Record the difference. A promising ad does not become a universal creative law because it performed well once.

The useful lesson from better recommendation systems is fairly ordinary: give them something worth recommending and measure the result in the terms that matter to your business. That leaves plenty of work for the team, but it is work the team can actually influence.