Meta Creative Similarity Score: What It Is and How to Lower It

Advertisers have started hearing about a "similarity score" Meta uses to decide which of their ads actually get delivered — and most have never seen it, don't know where it lives in Ads Manager, or what number is actually bad. It's real, it's genuinely easy to miss, and it explains a specific, common symptom: frequency climbing, unique reach flatlining, and ROAS sliding, even with a healthy-looking number of ads live in the account.

What the score actually measures

The Creative Similarity Score estimates how visually and structurally alike your active ads are to each other within an ad set. It functions as a stand-in for something Meta doesn't expose directly — Entity ID overlap, the internal grouping Andromeda uses to decide whether two ads are different candidates or the same one wearing different text. The higher the score, the more your "different" ads are actually being read as one ad.

Where to actually find it in Ads Manager

It's not a column that shows up on a standard reporting table, which is exactly why most advertisers never stumble onto it. It sits under Analyze & Report → Ads Reporting → Account Insights, alongside a related Creative Fatigue Score and a Top Creative Themes breakdown. Because it's new and not part of the default view, it's worth deliberately checking rather than expecting it to surface on its own.

One honest caveat: because this metric is still relatively new, it's best treated as a directional signal rather than a hard target to optimize against obsessively. Meta may reshape how it's calculated or surfaced over time, and there's no fixed formula it's discloses. Use it as a check, not a religion.

How the bundling actually drives your numbers down

When several ads score high on similarity, Andromeda collapses them into a single Entity ID — meaning they compete against each other for delivery instead of each opening up new reach. If you launch 30 ads that share the same template, background, and structure with different headlines over the top, the system may functionally treat that as one ad with 30 costumes, not 30 real opportunities. That's the mechanism behind a familiar, frustrating pattern: frequency creeping up on the same shrinking pool of people while overall reach stays flat, because the algorithm never actually had 30 distinct candidates to work with.

Reporting suggests scores above 60% are where retrieval suppression kicks in meaningfully, and practitioners generally aim to keep the score under 40% across active assets.

The fix: from visual variations to real concept diversity

Different headlines on the same hero shot won't move this number. The fix has to be structural — a genuinely different persona, angle, or offer, not a reworded version of the same ad. This is exactly the logic behind the Persona × Angle × Offer concept grid — each axis you change is a real, structural difference the algorithm can actually register as a new Entity ID, not just a new caption.

• Swap the persona — a different visual style, setting, or talent, not just different text over the same footage

• Swap the angle — a different emotional entry point or pain point, which usually also changes the visual, not just the copy

• Swap the offer — a different framing of the deal itself, which often naturally produces a different visual treatment too

A small-budget nuance worth knowing

Accounts spending under roughly $100/day face a steeper version of this problem — a smaller daily budget gives Andromeda less signal to work with, so it needs each concept to be more clearly distinct to categorize it correctly. For a small account, the fix isn't necessarily more ads. It's fewer, more deliberately opposite ones — two or three genuinely polar concepts (a plainly rational, stats-driven static against an emotional, story-led video, for example) rather than ten ads clustered around one visual template.

The number itself isn't the point. It's a diagnostic for a question worth asking before launching any new batch: if Meta looked at these ads with a computer, not a person's eyes, would it actually see something different?

It's also worth remembering why this matters at all — nobody, including experts, can reliably predict which specific ad in a batch will win. The similarity score isn't about picking a winner. It's about making sure the batch actually gives the algorithm real, distinct options to find one from — produced at