Meta's Andromeda rewards different ideas, not more versions of one ad
Meta says its Andromeda retrieval system works best when you give it real creative variety. Here is what Meta has actually said, and how to build and read a test.
Advertisers have been talking about Andromeda for months, and much of the advice online goes well beyond anything Meta has said. Start with the primary sources. Meta's engineering team described Andromeda in December 2024 as its machine learning system for personalized ads retrieval, the first stage of the ads recommendation pipeline. A March 2025 Meta business post said that as businesses add creatives to support a diversification strategy, Andromeda powers more complex models to choose the right creative for each person. That is the documented part. What follows is the interpretation, which we will label as such.
What retrieval does
According to Meta's engineering post, retrieval narrows tens of millions of ad candidates to a few thousand that move on to later ranking stages. Meta reported a 6 percent recall improvement and an 8 percent ads quality improvement on selected segments, and said Andromeda uses a hierarchical index to handle the growing volume of creative. It also said the architecture is expected to move toward a loss function that should produce a more diverse set of ad candidates, which it described as a future direction, not a current result.
The reason this matters is simple. Before an ad can win an auction, it has to be selected as a candidate for a person. If the system treats your ads as near-identical, you may be giving it one real option, however many files you upload.
Our reading, not Meta's
Meta has not, in anything we could find, published a threshold for how different two ads must be, or how many ads you should run. Some commentators cite specific similarity cutoffs, but we found no Meta source for them, so we would ignore them. What we can say is that the direction of Meta's own language, "diversification" rather than volume, argues for variety in substance:
- Different problems or benefits. One ad leads with saving time, another with cost, another with ease of getting started.
- Different people. Ads that speak to a different role, situation or motivation.
- Different formats and structures. A short demonstration, a testimonial-style walkthrough, a static comparison, a creator-style talk to camera.
- Different opening seconds. A change of hook is a bigger change than a change of caption.
Cosmetic variants, such as the same footage with a new headline or a swapped color, are iteration. They are useful for polishing a concept that already works. They are not the same as testing a new concept.
A way to organize the work
Think in concepts, not ads. Suppose you plan to run 12 ads. Three concepts with four variations each is a different test from 12 separate ideas. Tag every ad with its concept, format and audience angle in the name, so that you can aggregate results by concept afterward. The ad-level numbers are noisy, and the concept-level numbers are the ones that tell you what to make next.
- Brief by concept. Write the customer problem, the proof and the call to action for each, before anyone opens an editing tool.
- Start with fewer, more distinct concepts. Three to five clearly different ideas is a manageable start for a modest budget. That number is our practical suggestion, not Meta guidance.
- Keep structure simple. Fragmenting budget across many ad sets starves each of data. Where your objective allows, let one campaign carry the concepts.
- Judge on a cost-per-result and volume basis after the system has had time. Early spend is concentrated on a few ads by design, which tells you what the system favors, not necessarily what is best.
- Refresh by learning. When a concept leads, make new versions of the idea. When none do, replace the weakest concepts with new ideas instead of new edits.
The trade-offs
More distinct concepts means more production work and more review. Small teams should resist the urge to make everything at once. Pick the concepts with the clearest customer reasoning. Also be aware of your own measurement: if your attribution is partial, concept-level conclusions will be too, so confirm important winners with a controlled test where you can.
Ignore claims that a single technique guarantees better delivery. Meta's own description of the gain concerns selected segments, and the engineering post gave no instructions to advertisers.
The practical takeaway: the next time you brief creative, ask for three ideas that a stranger would not confuse with each other, and name them so you can read the results by idea.