Performance Marketing Visual Deep Dive
How Meta Andromeda Works:The Ads Retrieval Engine
A visual, stage-by-stage guide to Meta's next-generation ads retrieval engine — from impression triggers to deep embedding neighborhoods, isolating documented mechanics from practitioner interpretations.
Deeptanshu Sharma•Updated June 18, 2026•12 min read
How to read this guide: We strictly label claims with either ✅ Documented (Meta Engineering/NVIDIA confirmed) or 🔍 Interpretation (practitioner modeling / inference).
✅ Confirmed🔍 Inferred
Retrieval, Not Ranking
Andromeda is the bouncer of the system. It trims down candidate ads from tens of millions to a few thousand before heavier neural models predict individual conversion values.
Deep Multimodal Ingestion
Using the Generative Ads Model (GEM), the platform interprets ad copy, visual structures, color, and audio to build unified coordinates, reducing reliance on keyword targeting.
Creative as Targeting Signal
Because retrieval matches people directly with creative vector spaces, campaign setups must simplify. Broad targeting with distinct creatives scales best with Andromeda.
SYSTEM TOPOLOGY
Andromeda's Placement in the Delivery Stack
Before your ad can win an auction, it must survive Retrieval. Here is how Andromeda acts as the initial high-throughput bouncer, gatekeeping candidate ads before heavy ranking scoring is applied.
✅ DOCUMENTED MECHANICAL GATEWAYPhase 1 · Retrieval
The Retrieval Filter (Andromeda)
Andromeda processes the massive global pool of eligible ads (tens of millions of candidates) within Meta's few-hundred-millisecond delivery window. It evaluates the raw vector matching distance between the live user sequence embedding and the multimodal creative coordinates.
Candidate Capacity:Tens of millions
Process Latency Target:Within Meta's few-hundred-millisecond delivery window
Silicon Architecture:MTIA & NVIDIA Grace Hopper
Marketer Impact: 🔍 Growth practitioners infer that retrieval operates with format neutrality, meaning clear creative assets guide Andromeda to locate correct vector coordinate neighborhoods, bypassing historical keyword limits.
✅ DOCUMENTED EVALUATION GATEWAYPhase 2 · High-Fidelity Scoring
Deep Ranking & Auction
Only ads shortlisted by Andromeda enter the ranking models. These heavier models evaluate CTR, CVR, and advertiser bid values to compute precise engagement estimates before running the final auction.
Shortlist Candidate Capacity:A few thousand candidates
Prediction Parameters:P(Click), P(Conversion), Ad Quality
Final Decision Node:Bidding Auction Equation
Marketer Impact: While auction bids and historic CTR/CVR scores decide the final impressions, ranking is powerless if your creative vector doesn't pass Andromeda's initial Retrieval gate.
Interactive Walkthrough: The Andromeda Pipeline
Toggle through the 7 chronological steps below to see how a live impression trigger moves through multi-dimensional vector matching down to auction delivery.
✅ DOCUMENTED
Stage 1: The Impression Trigger
A user opens the feed, triggering a ~300ms countdown.
Trigger Activated
300ms Countdown...
The second a person launches Facebook or Instagram, an 'impression opportunity' is born. To deliver the perfect ad without lag, Meta's infrastructure has a fraction of a second to evaluate, filter, and score the ideal ad candidate.
Source: Meta Engineering Blog: 'Meta's Andromeda Advantage: Automation and Next-Gen Personalized Ads Retrieval Engine' (Dec 2024)Read Meta Blog
Interactive Algorithmic Sandbox
Andromeda Algorithmic Exploder
Break the black box apart. Tap the button below to visually "explode" the unified recommendation engine into its separate mathematical components and adjust parameters live.
Unified Engine
Unified Andromeda Engine Model
All systems are currently synced and encapsulated inside a single low-latency delivery wrapper.
Macro Pipeline Equation
MatchScore = Sim(U_seq, C_multimodal) × Pruning(Tree) × P_Rank
🔍Illustrative — not Meta's published equation
FLUXSY PRO-GRADE ALGORITHMIC TOOLS
Andromeda Creative Signal Estimator
Model your prospective ad variables below to calculate your illustrative indexability score (simulated) and see how Andromeda's retrieval pipeline might map your creative vectors.
Configure Ad Properties
Ad Copy Semantic Clarity75%
How explicit is your problem-solution hook? High clarity reduces embedding ambiguity.
Visual Asset Category Specificity60%
Does your visual directly depict your product category? Helps GEM's object detection parse accurately.
Creative Concepts / Variations3
Testing distinctly mapped conceptual angles prevents local vector optimization failures.
Targeting Scope StrategyBroad targeting lets Andromeda scan adjacent vector pockets for incremental value.
Estimated Andromeda Indexability
🔍 Illustrative estimate · Not an official Meta metric
68%
Retrieval BreadthBroad but Fragmented
Algorithmic Drift RiskHigh
Algorithmic Action PlanModerate Signal: Andromeda has a basic grasp of the content. Increase contrast, simplify visual graphics, and let Advantage+ utilize broader targeting to gather adjacent vector signals.
INTERACTIVE EMBEDDING SIMULATOR
Andromeda Semantic Vector Sandbox
Type your own ad copy below or load pre-built campaign templates. Andromeda's Generative Ads Model (GEM) reads this text, extracts core entities, parses the semantic layout, and converts it into high-dimensional vector coordinates. Watch your ad's dot glide dynamically across clusters in real time!
Draft Ad Headline & HookLocal B2C Services
Characters: 107/240Deterministic Hashing Active
Load High-Contrast Templates
GEM Extracted Entity Mappings🏷️ Emergency🏷️ Local Service🏷️ Home RepairConfidence: 99.8%
YOUR AD VECTORX:233, Y:74
Dynamic GEM Text FocusText Hook Clarities40%
Visual Category Object20%
Price/Value Anchor40%
Active Vector Coordinates8-Dimensional Float Representation-0.46
+0.60
-0.35
+0.71
-0.23
+0.83
-0.11
+0.94
💡 Why the dot moves: As you type, the keyword mappings shift. When you transition from SaaS keywords (e.g., automated cloud deployment) to E-commerce coffee bean templates, GEM's internal coordinates transform, and Andromeda retrieves candidate shortlists from the correct D2C neighborhood, optimizing bid auctions natively.
Demystifying the Black Box
Meta Andromeda is the unheralded engine driving modern campaign distribution. Historically, marketers optimized delivery by manually segmenting audiences — setting geographic boundaries, exclusions, interest layering, and demographic parameters. However, with the release of Meta's MTIA silicon and the integration of NVIDIA Grace Hopper processors, Meta has entirely automated the audience selection layer at the retrieval level.
Before your ad ever reaches an auction screen, Andromeda evaluates whether it deserves to compete. Understanding this stage-by-stage pipeline separates tactical operators from average agencies.
CRITICAL ETHICAL RIGOR AUDIT
What Meta Has NOT Documented
Many growth marketing guides present inferred practitioner ideas as official Meta engineering facts. To safeguard your strategy, keep these critical demarcations in mind:
- 🔍 1.The 'Creative Cluster ID / Entity ID' Concept: Meta documents mathematical proximity inside embedding layers, but has never confirmed a static, named system that labels and isolates creatives into locked boxes. These are helpful mental models, not rigid databases.
- 🔍 2.Colour-Based Audiences: While multimodal parsing captures visual features including colour and composition to compile vector embeddings, claims stating 'Andromeda routes ads entirely by color profiles' are over-specific practitioner speculation.
- 🔍 3.Placement-Based Filtering: Placement acts as an auction delivery variable, not a documented retrieval axis. Andromeda is format-neutral during initial matches.
Operational Blueprint for Advertisers
With Andromeda acting as the gatekeeper, your approach to creative creation must shift from aesthetic vanity to mathematical precision. Here is how top performance agencies navigate the retrieval protocol:
1
Isolate Hooks Semantically: Avoid cryptic, metaphorical copy. Lead immediately with direct product outcomes or exact customer pain points to help GEM's text models parse the vector accurately.
2
Unify Design Assets: Frame core objects within the central 80% viewport. If the algorithm cannot identify objects or text inside visual frames, indexability decays instantly, triggering higher CPMs.
3
Broaden Optimization Boundaries: Trust Advantage+ and broad targeting. Andromeda relies on finding adjacent intent coordinates; hyper-tight demographic segmentation locks the retrieval algorithm out, causing performance fatigue.
The Complete End-to-End Journey: From Creative File to Quality Feedback
The 7-stage pipeline above is the core loop. This section zooms all the way out and all the way in at once — every stage, broken into its individual steps, in the order they actually happen, from the moment you export a creative to the moment its performance teaches the system what to do next.
We've split it into three acts:
- Act I — Pre-Flight (before anyone sees your ad): creative build → ingestion → embedding → indexing.
- Act II — Live (a single impression, in ~300ms): trigger → retrieval → ranking → auction → render.
- Act III — Post-Impression (after the ad shows): measurement → attribution → learning loop.
A running note on honesty: Act I and Act II are heavily ✅ documented. Act III is where documentation thins out — Meta confirms the system learns from outcomes, but not a per-ad, step-by-step of how your specific creative gets "quality analysed." So Act III carries more 🔍 labels, each with the reason it's inferred.
ACT I — PRE-FLIGHT: from your export folder to the ad index
Step 1.1 — You build and publish the creative
🔍 InterpretationYou export the asset (image/video), write the copy, headline, and CTA, and publish through Ads Manager or the API. Nothing algorithmic yet — but every choice here becomes signal downstream, which is the whole reason creative now matters at the retrieval level. Reason for label: This is on your side, not Meta's system.
Step 1.2 — The ad enters review & becomes an eligible candidate
✅ DocumentedBefore anything else, the ad passes policy review and becomes one of the "tens of millions of eligible ad candidates" in the pool Andromeda draws from. (Meta describes the candidate pool explicitly; the review gate is standard platform behavior.)
Step 1.3 — Multimodal ingestion: the creative is read as content
✅ DocumentedMeta's Generative Ads Model (GEM) — the "central brain" feeding retrieval and ranking — learns from content across modalities: text, images, audio, and video. Your copy, your visuals, your video frames, and any audio are parsed as content, not just as attached metadata. This is the documented basis for "the algorithm understands your creative."
Step 1.4 — The creative is encoded into an embedding
✅ DocumentedThe parsed multimodal content is transformed into a dense vector (embedding) — a set of coordinates placing your creative in a high-dimensional space where semantically similar things sit close together. Meta's recommendation models are explicitly founded on learning these high-dimensional representations from features.
Step 1.5 — Visual and stylistic features fold into that vector
🔍 InterpretationBecause the image/video is embedded multimodally, visual characteristics — composition, the objects depicted, overall style, and colour as part of the image — plausibly influence where the creative lands. This is a reasonable implication of multimodal embedding, not a documented "colour" or "style" axis. Fair to assume the vector reflects your visuals; not fair to claim "it clusters by colour."
Step 1.6 — The creative takes its place in the hierarchical index
✅ DocumentedAndromeda organizes the ad universe into a jointly-trained hierarchical index — trained with the retrieval model so the index reflects the model's own sense of relevance. Your newly-embedded creative is slotted into this structure so it can be found fast at match time. Similar creatives occupy nearby branches.
Step 1.7 — "Neighbourhoods" form around similar creatives
🔍 InterpretationCreatives with similar embeddings naturally sit near one another — the documented behavior of embeddings. Practitioners call these groupings "clusters" or "entity IDs." Useful mental model; not a named, static Meta database of locked creative boxes. The durable truth: similar creative → similar vector → similar future audience.
ACT II — LIVE: one impression, roughly 300 milliseconds
Step 2.1 — The impression trigger fires
✅ DocumentedA person opens Facebook or Instagram. An "impression opportunity" is created, and the system has a few hundred milliseconds to fill it. This latency budget is the entire reason a fast retrieval pre-filter (Andromeda) exists instead of ranking everything at once.
Step 2.2 — The user is represented as a live embedding
✅ DocumentedThe person is encoded into their own embedding — built via sequence learning from the order and relationships of their recent interactions, capturing current intent rather than static demographics. (Meta documents sequence learning as a paradigm shift specifically for this.)
Step 2.3 — Retrieval matches user-vector to ad-vectors
✅ DocumentedAndromeda searches the hierarchical index for ad embeddings sitting near the user's embedding — semantically "nearby" candidates for this person, this moment. Crucially, this is why broad targeting can still reach precise audiences: the match is being made on vector proximity (which your creative drives), not just on your targeting settings.
Step 2.4 — The pool is pruned: tens of millions → a few thousand
✅ DocumentedUsing the hierarchical index and the deep retrieval network, Andromeda narrows the candidate set to a shortlist of a few thousand — without scoring every ad individually. This pruning is what the ~10,000× jump in model complexity (on Grace Hopper + MTIA hardware) made possible.
Step 2.5 — The shortlist is handed to ranking
✅ DocumentedRetrieval's job ends here. The few-thousand-ad shortlist passes to the heavier ranking stage. If your ad didn't make this shortlist, it is now out — it never reaches the auction, regardless of bid.
Step 2.6 — Ranking predicts value per ad
✅ DocumentedLarger, more sophisticated ranking models predict how likely this person is to take the action you're optimizing for (click, install, purchase) and combine that with your bid to estimate value.⚠️ Note: Meta has not published the ranking model's parameter count or its exact equation. Any formula shown (e.g., value ≈ P(action) × bid) is an illustrative representation of value-based ranking, not Meta's confirmed equation.
Step 2.7 — The value-based auction picks the winner
✅ DocumentedThe familiar auction ranks the shortlist on estimated action rates, ad quality, and bid, and selects the winning ad. Andromeda chose who competed; the auction chooses who won.
Step 2.8 — The ad renders in the feed
✅ DocumentedThe winning creative is served into the placement and the person sees it. This is the impression — the finish line of Act II and the start of the data that drives Act III.
ACT III — POST-IMPRESSION: measurement and the learning loop
Honesty header for this act: Meta confirms the system learns from outcomes and retrains on engagement and conversion signals. It does not publish a per-ad, step-by-step "quality analysis" procedure. So the existence of this loop is ✅; most of the specific internal steps are 🔍 — reasonable inferences from how the documented models are trained.
Step 3.1 — The impression and its context are logged
✅ DocumentedThe system records that the ad was shown, to whom (as signals, within privacy limits), in what context. This event becomes part of the data the models learn from.
Step 3.2 — The user's response becomes a signal
✅ DocumentedWhether the person scrolls past, clicks, watches, or converts, that reaction is captured as an interaction. Meta's sequence-learning approach is explicitly built on aggregating a person's actions over time — so this response feeds both future user representations and model training.
Step 3.3 — Conversions flow back via your measurement setup
✅ DocumentedDownstream events (add-to-cart, purchase, install) return through the Conversions API / Pixel / SDK you've configured. The quality and completeness of your measurement setup directly determines how good a signal the system receives — one of the few parts of Act III you actually control.
Step 3.4 — Signals are attributed and aggregated
🔍 InterpretationResponses and conversions are attributed to the ad within the configured windows and aggregated into the datasets that train the models. The existence of attribution is documented; the exact internal path from "a conversion happened" to "the model updated" is not published.
Step 3.5 — The ad's real performance updates its standing
🔍 InterpretationAs real engagement and conversion data accumulate, the system's estimate of how this creative performs — for which kinds of people — sharpens. Strong real-world response tends to mean the creative gets retrieved and ranked more favorably for similar users; weak response, less so. This is the mechanism practitioners experience as "the algorithm figured out my creative," and it's a sound inference — but Meta doesn't publish a named per-ad quality-score update.
Step 3.6 — Creative fatigue and saturation emerge
🔍 InterpretationOver time, as the same creative is shown repeatedly to a narrowing pool, response typically declines (rising frequency, falling CTR). The system shifts delivery accordingly. This is real and universally observed, but it's an emergent outcome of the loop, not a documented "fatigue module."
Step 3.7 — Learnings feed the next retrieval
✅ DocumentedThe aggregated signals become training data that improves future retrieval and ranking — Meta describes GEM and sequence learning as continuously learning from interactions to sharpen future predictions. The loop closes: today's impressions and outcomes shape tomorrow's shortlists.
The whole loop, in one line
Creative → embedding → indexed → (impression trigger) → user embedding → retrieval → shortlist → ranking → auction → impression → response captured → attributed → performance sharpens the estimate → feeds the next retrieval. The front half is well-documented; the feedback half is documented in principle and inferred in detail. Your leverage is at both ends you control: the creative you feed in (Act I) and the measurement you wire up (Act III).
ARCHITECTURE DIAGRAMS
Recreated Systems Architecture & Flows
These original diagrams recreate key components of Meta's documentation in the Fluxsy brand design system, explicitly mapping verified, documented flows against inferred practitioner systems.
FIGURE AAndromeda Ads Retrieval Architecture
Illustrates how candidate ads (tens of millions) pass through Andromeda's candidate matching model (User & Ad embeddings) utilizing a jointly-trained hierarchical index tree to slice down candidates before high-fidelity scoring is called.
Caption: Figure A charts the high-throughput filtering mechanism. High-fidelity ranking algorithms only execute on the retrieved candidate shortlist, cementing retrieval as the critical gatekeeper.Source: Meta Engineering (Dec 2024)
FIGURE BGEM Foundation Model and Vertical Model Mapping
Represents Meta's central multimodal ads model (GEM) serving as the unified brain layer, processing copywriting, visual images, audio files, and video structures down to task-specific downstream networks.
Caption: Figure B charts Meta's foundational multi-task architecture. GEM serves as the central brain. Read our dedicated companion guide: Source: Meta Engineering (Nov 2025)
FIGURE CSequence Learning & User Interest Attention Stack
Illustrates how chronological user interaction sequences are processed by transformer-based attention modules to calculate real-time dynamic intent coordinates.
Caption: Figure C charts sequence learning mechanics. Instead of relying on static user traits, the network parses user activity chronologically to calculate accurate instant user intent coordinates.Source: Meta Engineering (Nov 2024)
Primary Sources & Citations
- [1] Meta Engineering Blog. "Meta's Andromeda Advantage: Automation and Next-Gen Personalized Ads Retrieval Engine." December 2024. Available at: engineering.fb.com/2024/12/02/production-engineering/meta-andromeda-advantage-automation-next-gen-personalized-ads-retrieval-engine/
- [2] Meta Engineering Blog. "Sequence Learning for Personalized Ads Recommendations." November 2024. Available at: engineering.fb.com/2024/11/19/data-infrastructure/sequence-learning-personalized-ads-recommendations/
- [3] Meta Engineering Blog. "Meta's Generative Ads Model GEM: The central brain accelerating ads recommendation AI innovation." November 2025. Available at: engineering.fb.com/2025/11/10/ml-applications/metas-generative-ads-model-gem-the-central-brain-accelerating-ads-recommendation-ai-innovation/
- [4] NVIDIA Accelerator Solutions. "Scaling Ad Recommendation Architectures via Grace Hopper Superchips." secondary documentation, 2025.
Frequently Asked Questions
Request a Retrieval Audit
Fluxsy specializes in structuring campaigns and creatives designed to align directly with Meta's vector retrieval architectures. Request a free diagnostic assessment to evaluate your Andromeda signal health.
Compounding Growth Insights
Growth & Retrieval Intelligence
We document strict performance blueprints, privacy-centric analytics, and pricing structures directly linked to conversions.Highly recommended read for this page: Signal & Inertia