--- title: What Andromeda, GEM and Lattice are, and what they change for advertisers slug: meta-andromeda-gem-lattice summary: Plain-language model of Meta's 2025–2026 ad delivery stack (Andromeda retrieval → GEM ranking → Lattice unified models), the Advantage+ sales rollout of mid-2025, and the practical consequence that creative, not targeting or structure, decides delivery. domain: marketing tags: [meta-ads, andromeda, algorithm, advantage-plus, creative-strategy] related: [meta-creative-similarity-entity-ids, meta-targeting-broad-vs-interests, meta-account-structure-2026, meta-persona-angle-creative-dimensions, meta-consolidation-vs-sprint-approach, creative-diversity-levels] sources: [yt:x1b4h42HTG8, yt:ZqM6jPUAaeQ, yt:kn4ByanAe-s, yt:aNCUFwQw524, yt:-BYAC0ALW78, yt:FyWyHJh_6Ng] updated: 2026-10-24 --- # Andromeda, GEM and Lattice — what they are and what they change ## TL;DR - **Andromeda = retrieval**: it narrows millions of ads to a shortlist (~1,000) that *could* be shown to one user. If your ad isn't on the shortlist, bid, budget, structure and audience don't matter.[^1][^2] - **GEM = the "superbrain" ranker** that picks from the shortlist using deep behavioural signals; **Lattice = the unified model architecture** they run on (replacing many separate models).[^3][^4] - Andromeda launched Dec 2024 — it isn't "new" and didn't by itself kill your results. Losing accounts usually lost to competitors with bigger budgets and more, more diverse creative.[^5][^6] - Consequence: interest targeting, prospecting-vs-retargeting splits and elaborate account structures matter far less; **creative is the targeting**.[^7][^8] - Meta told Smart Marketer that strong creative earns up to a **50% cost "subsidy"** (the "creative cost multiplier"); weak creative effectively pays a 50% penalty.[^9] ## Details **The pipeline.** Dara Denney's metaphor: Andromeda is a personal assistant pre-selecting a few white t-shirts from many brands; GEM narrows further using what it knows about you (cotton vs linen); Lattice — the library holding GEM and the other models — makes the final pick.[^3][^4] Ignacio Manchi (Ready Set) uses a library: Lattice is one giant library that replaced many small, siloed libraries (one per vertical), Andromeda the front desk building the shortlist, GEM the head librarian making the final recommendation.[^10][^19] GEM uses **sequential learning**: after you book a ski resort you get ski gear and lift-pass ads, not two weeks of hotel ads.[^11] Manchi stresses Andromeda is a *living* system ("Meta's AI ad delivery and recommendation system"), not a one-off update.[^12] **Why old segmentation stopped paying.** Separate models used to make interest, lookalike and prospecting/retargeting splits useful; with Lattice the right ad can be matched to the right user directly.[^4] **Platform changes (as reported Jan 2026).** In June 2025 Meta began replacing Advantage+ Shopping with **Advantage+ sales campaigns**, finished by end of July 2025; sales, leads and app-install objectives were unified in one campaign type. Meta's three recommendations: broad audiences, Advantage+ (automatic) placements, campaign-level budgets.[^13][^14] **What Meta "reads".** Per Molly Pittman, Meta indexes ad copy, thumbnails and video content — down to the speaker's accent and pace — and even the history of the Facebook Page the ads run from (she questioned whether a marketer-audience Page could launch a pet brand).[^15][^16] Josh Brampton (Motion): the first "viewer" of every ad is Meta's AI, so you are making ads for AI first.[^17] **CTC's reading (Tony Chopp, Jul 2026).** Andromeda is a creative-first retrieval engine using creative attributes, user context and behavioural signals to build each impression's candidate set; Meta's engineers describe a **10,000× increase in model capacity** per impression opportunity. GEM (Generative Evaluation Model) ranks at "LLM scale". Implication: give the system more options, trust its allocation within cost controls, and stop ad-level decisions from backward-looking dashboards (see [[meta-breakdown-effect-judge-in-aggregate]]).[^20][^21] **Personalised frequency planning.** Ben Heath: Meta plans impression frequency per user within a campaign (e.g. four impressions over 48 hours) and post-Andromeda personalises *which* ad each person sees — one reason identical duplicate campaigns perform worse.[^22] **What to do.** Dara: know your audience segments, build diverse formats per segment, then layer messaging angles and awareness levels on proven formats; the bottleneck is creative velocity (how fast good ideas ship), not raw volume.[^6][^18] ## Disagreements / caveats - Who makes the *final* pick: Dara says Lattice; Manchi says GEM. Both are practitioner interpretations of Meta's public explanations — neither works at Meta.[^4][^19] - Dara frames poor results as a creative/competition problem; Molly says the many 2025 updates (not just Andromeda) changed the platform and some accounts still do better with the "old school" approach — see [[meta-consolidation-vs-sprint-approach]].[^14] - The 50% multiplier is a Meta rep's claim relayed by Molly, not a published figure. ## See also - [[meta-creative-similarity-entity-ids]] — how to get onto the shortlist - [[meta-targeting-broad-vs-interests]] - [[meta-account-structure-2026]] - People: [[people/dara-denney]], [[people/ignacio-manchi]], [[people/molly-pittman]], [[people/josh-brampton]] [^1]: yt:ZqM6jPUAaeQ @ 05:03 — Ignacio Manchi [^2]: yt:ZqM6jPUAaeQ @ 09:11 — Ignacio Manchi [^3]: yt:x1b4h42HTG8 @ 04:58 — Dara Denney [^4]: yt:x1b4h42HTG8 @ 05:44 — Dara Denney [^5]: yt:x1b4h42HTG8 @ 00:50 — Dara Denney [^6]: yt:x1b4h42HTG8 @ 07:20 — Dara Denney [^7]: yt:x1b4h42HTG8 @ 00:00 — Dara Denney [^8]: yt:kn4ByanAe-s @ 08:50 — Molly Pittman [^9]: yt:kn4ByanAe-s @ 05:07 — Molly Pittman [^10]: yt:ZqM6jPUAaeQ @ 06:44 — Ignacio Manchi [^11]: yt:ZqM6jPUAaeQ @ 11:39 — Ignacio Manchi [^12]: yt:ZqM6jPUAaeQ @ 01:37 — Ignacio Manchi [^13]: yt:kn4ByanAe-s @ 16:53 — Molly Pittman [^14]: yt:kn4ByanAe-s @ 19:26 — Molly Pittman [^15]: yt:kn4ByanAe-s @ 09:39 — Molly Pittman [^16]: yt:kn4ByanAe-s @ 10:43 — Molly Pittman [^17]: yt:aNCUFwQw524 @ 07:48 — Josh Brampton [^18]: yt:x1b4h42HTG8 @ 08:18 — Dara Denney [^19]: yt:ZqM6jPUAaeQ @ 09:58 — Ignacio Manchi [^20]: yt:-BYAC0ALW78 @ 03:14 — Tony Chopp [^21]: yt:-BYAC0ALW78 @ 05:48 — Tony Chopp [^22]: yt:FyWyHJh_6Ng @ 16:46 — Ben Heath