ads-meta

# Meta Ads Deep Analysis ## Andromeda + GEM + Lattice (2026) Meta's delivery stack was rebuilt across three releases: - **Andromeda** (Oct 2025) — ad-retrieval ranking model with 10,000× more model capacity than the previous funnel ([Meta Engineering, Dec 2024](https://engineering.fb.com/2024/12/02/production-engineering/meta-andromeda-advantage-automation-next-gen-personalized-ads-retrieval-engine/)). Filters the candidate creative set before the auction layer ever sees it. - **GEM** (Generative Embedding Model, late 2025) — replaces the feature pipeline. Creative *content* embeds directly into the targeting space, which is why "creative is the new targeting" is now mechanical truth not slogan. - **Lattice** (rolled out late 2025 / early 2026) — sequence-aware optimizer on top of GEM that uses user-action sequences to rank candidate ads. Net effect: creative diversity is now the #1 performance lever. Ads with Similarity Score >60% (per [Confect's measured threshold](https://confect.io/tactics/meta-andromeda-2026)) get retrieval suppression — the algorithm clusters near-identical creatives and silently limits their delivery. **100 minor variations perform no better than 10 genuinely distinct ones.** Prioritize concept / angle / format diversity over variant volume. ### Creative-as-targeting scoring rubric When auditing a creative library against Andromeda's retrieval logic, score across these 5 axes (each 0-2, total 0-10): | Axis | 0 (Risk) | 1 (OK) | 2 (Strong) | |------|----------|--------|------------| | Concept diversity | Single core message / value prop across all assets | 2 distinct messages | 3+ distinct angles (problem-led, social proof, comparison, …) | | Format diversity | One format (e.g. all static image) | 2 formats | 3+ (image, video, carousel, collection) | | Visual diversity | One palette / one model / one composition | 2 distinct visual treatments | 3+ visually distinct treatments | | Hook diversity (video) | All hooks ≤3s look alike | 2 hook patterns | 3+ hook patterns (UGC POV, question, claim, demo, …) | | Headline diversity | All headlines paraphrase the same line | 2 headline structures | 3+ structures (number-led, question, claim, comparison) | Score 8-10 = LOW Entity-ID clustering risk. Score 4-7 = MEDIUM risk (some suppression likely). Score 0-3 = HIGH risk (significant retrieval ticket loss). ### Entity-ID Clustering Predictor (pre-launch) Before launch, predict which creatives Meta will cluster. Cluster-mates share retrieval tickets — only one wins per impression opportunity. **Predictor heuristics (apply to every pair of creatives in the launch set):** 1. **Visual fingerprint** — same product hero, same model, same backdrop, same lighting → **likely cluster**. Different products or different visual identities → likely *not* a cluster. 2. **Headline fingerprint** — same first 4 tokens → likely cluster (e.g. "Save 30% on" + "Save 30% off" + "Save 30% — limited time"). 3. **Body copy finger