Most brands running Meta in 2026 know that Andromeda shifted the targeting lever from audiences to creative. What fewer brands understand is the specific mechanism that punishes them for acting like the old playbook still works.
Inside Meta's ad retrieval system, every creative you run gets assigned to an Entity ID. The system uses computer vision to evaluate the visual composition, color distribution, and structural layout of your assets. When two creatives share enough visual DNA, the algorithm groups them into the same Entity ID.
When that happens, those ads don't compete for different users. They compete against each other for the same slot. You're paying to run 50 variations of the same ad — and Meta is serving them as one.
How Entity IDs work
Andromeda operates in two stages: retrieval and ranking. In the retrieval stage, the system pulls candidate ads from your active set. In the ranking stage, it decides which of those candidates to serve to which user.
Entity IDs are a retrieval-stage construct. If the system determines that creative A and creative B are structurally similar enough to be the same "concept," they share a single Entity ID. That Entity ID gets one retrieval slot — the same slot a single ad would get.
This is why flooding the account with 50 variations has stopped working. You're not giving the algorithm 50 shots. You're giving it 1, dressed 50 ways. The algorithm has gotten good enough at seeing through the costume.
The 60% threshold
The specific trigger is a Creative Similarity Score above 60%.
Performance data from Q2 2026 across hundreds of accounts shows that once two creative assets score above this threshold, they get collapsed into a single Entity ID in retrieval. The combined reach of your "variant A, B, C, D" ends up roughly equivalent to running just one of them.
Scores below 40% are what top-performing brands are targeting. At that level, each asset registers as a genuinely distinct candidate — which means each one gets its own shot at finding new users.
The gap between a score of 38% and a score of 65% doesn't look dramatic in Ads Manager. In delivery terms it's the difference between four independent ads finding four distinct audience segments and four ads fighting over one.
What's actually in most accounts right now
The most common pattern: a brand has 100+ active creatives, 80 of which are minor variations on two or three winning concepts. Different CTA button. Different background color. Different hook line but the same visual structure. Andromeda's computer vision treats all of them as the same creative.
The practical result is inflated impression frequency on a narrow audience, rising CPMs, and a plateau on scale — with no clear explanation in Ads Manager. You're running 100 ads and getting the delivery equivalent of five.
The brands outperforming benchmarks in Q2 2026 are doing something different. They're running fewer total ads, but each asset comes from a structurally distinct concept: different visual treatment, different talent, different scene context, different product angle, different emotional register. Research across leading DTC accounts shows the top quartile is running 33% more unique assets and 40% fewer total creatives than the prior year.
Not more ads. Different ads.
What "genuinely different" actually means
The instinct when you hear "creative diversity" is to change what's easy to change. That's where most of the wasted work happens.
Swapping the CTA from "Shop Now" to "Get Yours" doesn't move the similarity score. Changing the background from white to cream doesn't either. Even using a different product SKU on an identical visual template can still produce a 70%+ similarity score if the compositional structure stays the same.
What actually moves it:
- Scene composition and shot type — lifestyle vs. product, wide vs. close-up
- Dominant color palette, not one accent color
- Human presence vs. no human presence
- Motion pattern and camera movement style
- Visual hierarchy — where the eye lands first
These are the signals Andromeda's vision model weights. If those don't change, the text changes are invisible to the retrieval engine.
What to do this week
Meta's Creative Hub shows Similarity Scores at the creative level. Pull your top 20 active creatives and group any that score above 60% with each other. Each cluster is a single Entity ID in practice.
Then make a binary decision: is this concept strong enough to earn a genuinely new visual treatment, or should the whole cluster be retired and replaced with something structurally different?
The accounts getting the most out of Andromeda right now are the ones feeding the retrieval system enough variety that it has something to work with. The algorithm is good at finding the right person for a message. It's not good at manufacturing a different message when all you've given it is the same one wearing different colors.
If you want to see how your creative distribution looks against broader account benchmarks, the free audit at Gromerce surfaces the patterns Meta's default dashboard won't show you.
Sources: Meta Andromeda documentation, admetrics.io Q2 2026 performance analysis, DTC account research, July 2026
