Entity ID Diversity: the concept agencies ignore

Producing 20 variants of the same concept isn't diversity: it's inflated volume. Real diversity requires different narrative angles, different psychological triggers, different formats. Otherwise you're saturating the same semantic space and the system treats your variants as a single creative with multiple instances.

Note on sources, added 29 August 2026. "Entity ID" isn't a term Meta uses in public. It doesn't appear in the Meta Engineering post on Andromeda of 2 December 2024, it doesn't appear in the Lattice paper, and it doesn't appear in any Meta documentation I've managed to find. It circulates in agency blogs, this one included in its first version, where it was presented as a documented mechanism. It isn't. Below you'll find the one sentence Meta has written on the subject first, and then the model I use.

What Meta says, and what I call Entity ID

Andromeda is the retrieval system Meta uses to cut down the inventory of candidate ads before the auction. On the point that concerns us here Meta has written a single sentence, in the Engineering post of 2 December 2024: the system returns "a more diverse set of ad candidates", and "increased ad diversity can improve people's experience with ads and drive better advertiser outcomes".

That's the whole of the public material. Meta says greater diversity among candidate ads produces better outcomes. It doesn't say how it measures that, with what representation, or what happens to two ads that resemble each other.

What follows is my model for working with it, and I call it Entity ID because I need a name for it. I assume every creative has a representation of its content behind it — text, visual, narrative structure — and that two creatives with different copy but the same angle, the same visual style and the same psychological trigger end up close together in that representation. From there I infer they get treated as redundant candidates for the same impression slot.

Meta has written one sentence about ad diversity. Everything else, "Entity ID" included, is a model practitioners build for themselves so they have something to work with.

The practical result: in an auction where you have 5 creatives with similar Entity IDs, you're not entering with 5 independent candidates. You're entering with a cluster, and the system picks the best representative of that cluster — ignoring the others. You pay for the upload, you pay for the management, you get no extra distribution.

Volume vs semantic diversity: the concrete examples

The line between a variant and a distinct creative isn't intuitive. These are real examples of what counts as inflated volume and what counts as semantic diversity.

Inflated volume: same hook "Discover product X", you only change the background photo (color 1, color 2, color 3), same copy in the body, same CTA "Buy now." Andromeda reads three Entity IDs that sit very close together. The system distributes one, two at most. The others die without meaningful impressions.

Real semantic diversity: creative A with a hook based on saving ("It costs you less than a coffee a day"), creative B with a hook based on identity ("Not for everyone — only for people who already get it"), creative C with a hook based on a specific problem ("If X happens to you often, this is why"). Three distinct psychological triggers, three narrative angles, three distant Entity IDs. Andromeda treats them as independent candidates and distributes them to different audiences.

On the accounts I run, groups of creatives built the second way tend to pay a lower CPM. I write that as a professional observation and not as a measurement: I have no publishable data series, with windows and verticals declared, that proves it. Anyone who hands you a percentage on this is making it up, unless they show you the sample.

How to audit your creative library for real diversity

The audit takes three steps.

First: categorize every live creative by primary psychological trigger. The base categories are: saving/economic value, identity/status, problem/solution, social proof/FOMO, curiosity/entertainment. If 70% of your creatives fall into a single category, you have a semantic diversity problem no matter how many variants you have.

Second: map the narrative angles. An angle is the perspective you tell the product from: product benefit, customer transformation, comparison with an alternative, brand story, specific use case. Every angle should have at least 2-3 live creatives. If you have a single angle with 20 variants, you're in the volume inflation case.

Third: check the formats. Video, static image, carousel and collection ads are distinct formats that generate different signals for Andromeda. A video creative with a verbal hook and a static image with the same text message aren't semantically identical — the format contributes to the Entity ID representation.

The 4 dimensions of real creative diversity

Real creative diversity is built on four independent axes.

Narrative angle: how you position the product in the customer's mind. "Product A does X" is an angle. "Your life without product A" is a different angle. "How people like you use product A" is a third angle. Every angle activates different cognitive schemas.

Psychological trigger: the emotional mechanism that drives the click. The six main ones on Meta: scarcity/urgency, identity/belonging, solution to the pain, aspiration/status, curiosity/pattern interrupt, social proof. Don't combine two triggers in the same creative — it weakens both.

Format and structure: the technical form of the creative. Ratio, length (for video), presence of text overlay, presence of voiceover, use of UGC vs branded vs product-only. Different formats reach different audiences even with the same content because consumption behavior changes.

Hook structure: the first 3 seconds for video, the first line for static. A hook based on a direct question ("Have you ever had this problem?"), one based on a provocative statement ("Most brands are lying to you about this"), one based on an immediate visual demo are structurally different hooks — not variants of the same hook.

The performance symptoms when you saturate the semantic space

When you have too many creatives with close Entity IDs, the signals show up in a precise order.

First: CPM rising with no apparent explanation. It isn't seasonality, it isn't external competition. It's the system distributing impressions to the same profiles over and over because your creatives offer no variety of implicit targeting.

Second: CTR flattening out across all creatives almost at the same time. There's no creative that scales while the others decay — they all decay together. This is the signature of semantic saturation, not of individual creative fatigue.

Third: frequency climbing on specific segments while total reach stalls. You're hitting the same people over and over with messages they perceive as identical — because semantically they are.

If you see these three signals in sequence over 2-3 weeks, the problem isn't the budget. It isn't the targeting. It isn't the offer. It's that Andromeda doesn't have enough real diversity to work with. The same principle applies to the Lattice system that handles cross-placement learning.

30-day creative diversification plan

Week 1 — Audit and mapping. Categorize every live creative according to the 4 dimensions above. Identify the clusters of similar Entity IDs. Pause the redundant creatives inside a cluster and keep only the representative with the lowest CPM.

Week 2 — Producing the missing angles. If you only have 1-2 live angles, produce creatives for the 2 angles furthest from your current cluster. Don't optimize — explore. The brief has to specify the target psychological trigger and the hook structure, not just the visual tone.

Week 3 — Testing in isolation. Launch the new angles in a separate ad set with a fixed budget (not Advantage+) for 7 days. The goal isn't immediate ROAS — it's understanding which angles find traction and which Entity IDs the system actually distributes.

Week 4 — Integration into the main system. The 2-3 angles that show a CTR above benchmark go into rotation in the main system. Build 2-3 format variants for each one. From this point on you have a creative library with real Entity ID diversity — and the system has enough material to do real matching.

Changelog

29 August 2026: flagged that "Entity ID" appears in no public Meta documentation. Added the one documented sentence about ad diversity, from the Meta Engineering post on Andromeda of 2 December 2024. Kept that apart from the working model I use, which the earlier version presented as a documented mechanism. Removed the claim that the CPM difference is measurable, replaced with the professional observation and the limit that comes with it.

---

If you want to run this audit on your account — identify the redundancy clusters, build a structured brief for real diversity, integrate the system into Advantage+ — book a 1-to-1 strategy session. A 90-minute session is enough to map the problem and walk out with a concrete plan.