Meta Andromeda Explained: What Performance Marketers Need to Know in 2026

Meta Description: Meta Andromeda changed Facebook ads forever. Learn what this AI algorithm means for your 2026 performance marketing strategy, backed by official Meta data and practical optimization tactics.

URL Slug: meta-andromeda-explained-performance-marketers-2026

Primary Keyword: meta andromeda
Secondary Keywords: meta andromeda update, meta ai advertising, facebook ads ai, andromeda meta ads

Target Audience: Performance marketers, paid media specialists, CMOs, and marketing directors at B2B challenger brands

Featured Image Alt Text: "Meta Andromeda AI advertising system visualization showing creative-first ad delivery architecture for Facebook and Instagram advertisers in 2026"

Word Count: 6,200 words

On December 2, 2024, Meta quietly announced something that would fundamentally change how Facebook and Instagram ads work. Most advertisers missed it. The ones who paid attention gained a massive competitive advantage.

The announcement? Meta Andromeda: a complete replacement of the ad delivery infrastructure that's powered Facebook advertising since its inception.

Here's the reality: If you're still running Meta ads the way you did in 2024, you're likely wasting budget. The system doesn't work the same way anymore. Audience targeting matters less. Creative matters more. And the gap between advertisers who understand this shift and those who don't is widening every week.

Search interest tells the story. In early 2024, "Meta Andromeda" got zero searches. By October 2025, it peaked at 2,400 monthly searches. That's a clear sign that advertisers finally realized something fundamental had changed.

This guide breaks down what Meta Andromeda actually is, how it changes your Meta advertising strategy, and exactly what you need to do differently in 2026. No fluff. Just practical frameworks backed by official Meta data and real-world performance patterns.

What is Meta Andromeda? (And Why It Matters)

The Technical Explanation

Meta Andromeda is an AI-driven ads retrieval system. It's the first stage in Meta's multi-stage ad recommendation process. According to Meta's Engineering team (December 2024), Andromeda processes tens of millions of active ads and narrows them down to thousands of candidates for each user, each time they open Facebook or Instagram.

The system runs on NVIDIA's Grace Hopper Superchip (GH200) and represents a 10,000x increase in model complexity compared to the previous ad delivery system. That's not marketing speak. That's the actual technical specification from Meta's engineering announcement.

The breakthrough? Sublinear inference cost. In simple terms, Andromeda can handle exponentially more creative variations without exponentially more computing power. This is why Meta can now process 15 million+ AI-generated ads per month (from over 1 million advertisers) without the system collapsing.

The performance impact is measurable. Meta for Business reports that Andromeda improved ads quality by 8% and recall by 6% across the platform. That might sound modest, but across a $196 billion ad revenue base (Meta's 2025 total), small percentage improvements represent billions in additional advertiser ROI.

The Simple Explanation

Think of Andromeda as a personal concierge who knows exactly what each of Meta's 3+ billion users wants to see. It makes decisions based not on who they are, but on the creative content itself.

Before Andromeda (2014-2024):
You (the advertiser) defined your target audience using demographics, interests, and behaviors. Meta's system started with that audience, filtered it down, and ranked users most likely to convert. Your targeting choices determined who saw your ads.

After Andromeda (2025-present):
Your creative content determines who sees your ads. Andromeda scans the entire user base instantly, matches creative signals (visual elements, copy themes, messaging angles, format) to user intent embeddings, and retrieves the best matches. Your creative choices determine who sees your ads.

This is the fundamental mindset shift: You're no longer buying access to audiences. You're buying algorithmic distribution based on creative signals.

Timeline: When Meta Andromeda Rolled Out (And What Changed)

The rollout wasn't a single flip-the-switch moment. Meta deployed Andromeda in phases:

Here's what advertisers noticed during the transition:

Broad targeting started outperforming narrow interest stacks. Campaigns with "Advantage+ Audience" or broad targeting (18-65+, all interests) were delivering better ROAS than precisely segmented campaigns. This went against a decade of Facebook ads best practices.

Simplified account structures won. Advertisers running 1-3 consolidated campaigns outperformed those with complex funnel structures and dozens of ad sets. Less became more.

Creative fatigue accelerated. Ads that used to perform for weeks started burning out in days. The system was learning faster and needed more creative input to maintain performance.

Advantage+ campaigns improved dramatically. According to official Meta data, advertisers using Advantage+ Shopping campaigns with creative AI features saw 22% ROAS increases compared to manual campaigns.

The common thread? The system was no longer rewarding manual optimization expertise. It was rewarding high-quality creative input and letting the AI do the rest.

How Meta Andromeda Works: The Three-System Stack

Andromeda doesn't work alone. It's part of a three-system AI stack that Meta deployed throughout 2024-2025. Understanding how these systems interact is key to optimizing for them.

System 1: Andromeda (The Retriever)

Role: Personalized ads retrieval engine

Andromeda sits at the top of the funnel. Every time a user opens their feed, Andromeda evaluates millions of active ads based on:

• Historical user engagement patterns

• Ad copy and messaging themes

• Creative elements (visuals, video content, format)

• Semantic meaning of the ad (not just keywords, but actual concepts)

It predicts which ads this specific user will engage with and narrows millions of possibilities down to thousands of candidates. According to Meta's Engineering team, Andromeda processes three orders of magnitude more ads than the next stage in the ranking process.

The key innovation is hierarchical indexing. Instead of evaluating every ad individually, Andromeda groups ads by semantic similarity and evaluates clusters. This enables the 10,000x model capacity increase without proportional computing cost increases.

Why this matters for advertisers: Andromeda supports the explosive growth in creative variations from Advantage+ tools. Meta reports that over 1 million advertisers created 15 million+ ads in a single month using AI generation tools. Without Andromeda's sublinear scaling, this volume would crash the system.

System 2: GEM (The Brain)

Role: Generative Ads Recommendation Model

GEM (announced March 2025) is the "brain" behind the operation. It's a large language model-based architecture that identifies patterns across organic content interactions and paid ad sequences.

GEM synthesizes:

• Engagement signals (likes, shares, saves, time spent)

• Behavioral patterns (what users do after seeing ads)

• Conversion data (what drives actual business outcomes)

• Cross-surface learning (patterns that work across Facebook, Instagram, Reels, Stories)

According to Meta for Business (March 27, 2025), GEM is 4x more efficient at driving ad performance gains compared to Meta's original models. It improved ad conversions on Meta Reels by 5% within the first quarter of deployment.

Think of GEM as a super brain that reads an entire library in seconds and understands relationships between all elements. It then feeds those insights to Andromeda to improve retrieval predictions.

GEM is trained on thousands of GPUs, and Meta doubled GPU allocation for GEM training in Q4 2025. This is a massive infrastructure investment that signals how central GEM is to Meta's ad strategy.

System 3: Meta Lattice (The Library)

Role: Unified ad ranking architecture

Meta Lattice (also announced March 2025) replaced Meta's previous collection of specialized ranking models with one giant system. Instead of separate models for purchase campaigns, lead campaigns, app installs, and different placements, Lattice learns from all of them simultaneously.

The key innovation? Generalized learning across campaign objectives and surfaces. If Lattice learns that a specific creative pattern drives purchases on Instagram Stories, it applies that learning to Facebook Reels purchase campaigns automatically.

The performance impact:

• +12% ad quality increase (Meta for Business)

• +6% ad conversion increase (Meta for Business)

Meta describes Lattice as understanding purchase journeys across all surfaces simultaneously, rather than treating each placement as an isolated event.

How the three systems work together:

1. GEM identifies what types of content and creative patterns drive results

2. GEM feeds those insights to Andromeda

3. Andromeda uses those insights to retrieve the best ad candidates for each user

4. Lattice ranks those candidates and determines which ad actually gets shown

5. The cycle repeats, with each system learning from the results

This is why "feed the algorithm more data" isn't just advice. It's how the system is designed to work.

The Fundamental Shift: Audience-First vs. Creative-First

Understanding the old vs. new approach is critical because most advertisers are still mentally operating in the old paradigm.

The Old Way (Pre-Andromeda)

How it worked:

• Advertisers defined audience targeting using demographics, interests, behaviors, and lookalike audiences

• Meta's system started with that defined audience, filtered it down based on budget and competition, then ranked remaining users by likelihood to convert

• Account structures were hyper-segmented: separate campaigns for different audience segments, interests, funnel stages

• Optimization meant manual testing, budget reallocation, audience refinement

• Creative was a secondary consideration: something to test within your defined audiences

Performance model: Targeting precision = Success

The best advertisers were audience experts who could identify niche interests, build sophisticated lookalike stacks, and structure accounts to isolate audience performance.

The New Way (Post-Andromeda)

How it works now:

• Creative content determines ad retrieval. The visual elements, messaging themes, language, format, and semantic meaning of your ad signal who should see it

• Andromeda scans the entire user base and retrieves matches based on creative signals, not advertiser-defined boundaries

• Account structures are simplified: fewer campaigns with broader targeting and more creative variations

• Optimization means creating diverse creative angles and letting AI learn which patterns work

• Creative is the primary lever. It's both your targeting and your conversion driver

Performance model: Creative diversity + AI learning = Success

According to Meta's official statement: "Andromeda does not optimize for novelty. It optimizes for semantic meaning."

This is crucial. Creating 20 versions of the same ad with different headlines doesn't help. Creating 10 ads with fundamentally different value propositions, messaging angles, and visual approaches does.

What This Means for Your Campaigns

Targeting is less deterministic. Interest-based targeting and custom lookalikes are underperforming broad targeting in most cases. Industry analysis from Search Engine Land (January 2026) confirmed this pattern: "Broad targeting began outperforming interest stacks" across hundreds of advertiser accounts analyzed.

Creative is the primary signal. The visuals you choose, the hooks you use, the themes you explore, and the language you write determine who sees your ads more than any targeting setting.

Volume enables AI. More creative inputs give the algorithm more signals to learn from. One great creative in 10 variations performs worse than 10 different creative concepts.

Simplification wins. Consolidated campaigns with larger data pools enable faster AI learning than fragmented structures with isolated datasets.

Official Performance Data: What Meta's Numbers Show

Let's look at the documented performance improvements from Meta's official announcements. These aren't case studies or testimonials. These are actual platform-wide data:

Scale Stats:

• 1 million+ advertisers used Meta's GenAI tools in a single month

• 15 million+ ads created using AI generation

• Model complexity: 10,000x increase

• Ad revenue 2025: $196 billion (from $201B total revenue)

These aren't cherry-picked success stories. These are platform-wide averages measured across Meta's entire advertiser base.

The numbers tell a clear story: The AI systems work, and the performance gains are real. But they require adapting how you run campaigns.

How to Optimize Your Meta Ads for Andromeda in 2026

Here's the practical implementation framework. This isn't theory. It's what's working now based on official Meta guidance and confirmed performance patterns.

1. Creative Strategy Becomes Your Core Lever

What to do:

Volume over micro-variations. Create 10-15 truly different creative angles, not 20 versions of the same ad with different headlines. Andromeda optimizes for semantic meaning. Minor variations read as identical to the algorithm.

Semantic diversity. Test different:

• Hooks and opening statements

• Value propositions (solve different problems)

• Messaging frameworks (emotional vs. logical, aspirational vs. practical)

• Social proof types (testimonials vs. data vs. case studies)

Format mix. Combine:

• Static images (high-quality product shots, infographics, quotes)

• Videos (demos, testimonials, explainers, behind-the-scenes)

• Carousels (step-by-step guides, product features, before/after)

• UGC-style content (customer videos, user testimonials)

Clear hooks in first 3 seconds. Don't build to a hook. Start with it. 80% of Reels plays happen with sound off, so use captions and text overlays.

Practical Framework:

• Minimum: 10 creative variations per campaign

• Optimal: 15-20 variations with distinct angles

• Test dimensions: Different personas, pain points, benefits, formats, tones

The goal: Feed the AI a buffet of creative signals so it can learn what semantic patterns drive results for your business.

2. Simplify Your Account Structure

What to do:

Consolidate campaigns. Move from 10+ campaigns to 1-3 core campaigns. Typical 2026 structure for most advertisers:

1. Testing/Challenger campaign (new creative, new audiences, experiments)

2. Scaling campaign (proven winners with increased budget)

3. Evergreen campaign (always-on core messaging)

Broader targeting. Replace interest-based targeting with:

• Advantage+ Audience (for Sales objective)

• Broad targeting (18-65+, all interests)

• Lookalikes only as starting suggestions, not constraints

Consolidated budgets. Larger budget pools enable faster AI learning. Campaign Budget Optimization (CBO) performs better than ad set budgets in most cases because it gives the AI more flexibility to allocate spend.

Why it works: Andromeda and GEM learn faster with more data volume. When you split budget across 15 ad sets, each one has limited data to learn from. Consolidated campaigns create larger learning datasets.

Specific tactics by objective:

• E-commerce: Use Advantage+ Shopping Campaigns as default

• Mobile apps: Use Advantage+ App Campaigns

• Lead generation: Broad targeting + Advantage+ placements + high-quality lead forms

Meta's official data shows 22% ROAS lifts when advertisers turn on Advantage+ features. That's not a marginal improvement. It's the difference between profitable and unprofitable campaigns.

3. Embrace Learning Stability

What to do:

No-touch windows. Commit to 7 days or 50-75 conversions before making any campaign changes. Yes, this feels uncomfortable if you're used to daily optimization. But the data is clear: frequent edits reset the learning phase and delay performance.

Stop micro-optimizations. Every time you edit targeting, change budgets significantly (more than 20%), or pause/restart campaigns, the algorithm has to relearn. This kills performance velocity.

Evaluate on trends, not daily fluctuations. Look at 3-7 day rolling windows. Day-to-day volatility is normal system behavior. It doesn't mean something's wrong.

Decision Framework:

• Week 1: No changes unless the campaign is truly broken (spending with zero results, technical issues, wrong audience entirely)

• Week 2: Evaluate trends. If clear underperformance vs. benchmarks, adjust creative or budget

• Week 3+: Optimize based on accumulated learning. Add new creative, increase budgets on winners, pause clear losers

GEM identifies sequence patterns over time. If you keep interrupting the sequences, the system can't learn what actually works.

4. Budget as a Signal

What to do:

Align budget to conversion event. Higher-intent events (purchases, qualified leads) require more budget than awareness events (page views, content views) because the AI needs more volume to find converters.

Minimum thresholds. Meta's guidance: Aim for 50+ conversions per week per campaign. Below that, the algorithm struggles to find patterns.

Consistent investment. Steady daily budgets enable consistent learning. Stop-start budget patterns (big spend Monday, zero spend Wednesday) confuse the algorithm.

Budget Reality Check:

Your budget needs to support the conversion volume the AI needs to learn. If your cost per lead is $25 and you're spending $50/day, you're getting 2 leads per day or 14 per week. That's below the optimal learning threshold (50/week).

Specific guidance by objective:

• Lead generation: $50-100/day minimum (adjust based on your CPL)

• E-commerce (purchase optimization): $100-200/day minimum

• App installs: $50-100/day for install campaigns

• Awareness/Traffic: $20-50/day minimum

If you can't hit these minimums, consider optimizing for an upper-funnel event with higher volume (landing page views instead of purchases, for example) until you have budget to support lower-funnel optimization.

For more on performance marketing budget optimization, see our guide on how AI-powered marketing intelligence helps challenger brands compete.

5. Rethink Your Role as a Marketer

Here's the uncomfortable truth: The skills that made you successful in 2020 aren't the skills that will make you successful in 2026.

The mindset shift:

From: Manual optimizer, audience expert, budget controller, performance analyst
To: Creative strategist, brand guardian, framework builder, AI collaborator

Your new responsibilities:

Define clear brand positioning and messaging. The AI can't create your brand voice. You have to give it strong inputs that reflect your actual positioning. If your creative is generic, the AI will distribute it to generic audiences and you'll get generic results.

Create strong creative inputs and frameworks. Build a creative system: documented brand guidelines, approved messaging angles, visual style standards, hook templates. Then create diverse executions within that system.

Collaborate with design teams for scalable creative production. You need to produce 10-15 new creative variations per week. That requires either in-house design resources, contractors, or AI generation tools (Meta offers these natively). Build the workflow now.

Set guardrails for brand integrity. Advantage+ will automate many decisions. Your job is defining what's off-limits: brand safety parameters, content categories to avoid, messaging that doesn't align with positioning.

Provide strategic judgment on campaign goals. The AI optimizes for whatever conversion event you give it. You need to decide if you're optimizing for the right business outcome.

What humans do better than AI:

• Novel creative ideas and unexpected concepts

• Brand judgment and positioning decisions

• Strategic campaign planning and goal-setting

• Cross-channel coordination and storytelling

• Understanding customer context the data doesn't capture

What AI does better than humans:

• Pattern recognition across billions of user signals

• Real-time bidding and budget allocation

• Audience matching at scale (3+ billion users)

• Rapid testing across thousands of variables simultaneously

• Continuous optimization 24/7 without fatigue

The goal isn't to fight the AI. Focus on what you do better and let the AI handle what it does better. Learn more about restructuring your marketing team for AI collaboration.

The Challenger Brand Advantage: Why Smaller Teams Can Win

Here's the counterintuitive truth: Andromeda favors agile teams over bloated structures.

Why challenger brands have an edge:

Speed to execution. Smaller teams make decisions faster, test creative faster, and iterate faster. In a system that learns from volume and speed, this is a massive advantage.

A challenger brand can go from creative concept to live campaign in 2 days. An enterprise with approval layers, legal reviews, and brand committees takes 2 weeks. Over a quarter, that's 45 creative tests vs. 6. The volume advantage compounds.

Creative agility. Less bureaucracy means more creative risk-taking. The brands winning with Andromeda are testing bold creative concepts, not playing it safe. Challenger brands do this naturally.

Simplified structures. Challenger brands are already running lean. You probably have 3-5 campaigns, not 50. That means less legacy structure to unwind and simpler transitions to Andromeda-optimized setups.

AI as the great equalizer. Andromeda's intelligence is accessible to all advertisers regardless of budget. A challenger brand with a $50K/month Meta budget using Advantage+ and strong creative can outperform an enterprise with a $500K/month budget running legacy manual optimization.

The enterprise disadvantage:

• Complex approval processes slow creative production

• Legacy account structures are difficult to simplify (political resistance to change)

• Risk-averse cultures resist testing bold creative

• Siloed teams struggle with the consolidated approach Andromeda rewards

• "We've always done it this way" thinking prevents adaptation

How DOJO AI helps:

This is where an AI Marketing Operating System makes the difference. DOJO AI consolidates your marketing data into a unified intelligence layer and provides:

• Performance monitoring across Meta and other channels in one place

• Creative performance analysis at scale (which themes, hooks, formats drive results)

• Campaign recommendations based on AI-learned patterns

• Cross-channel insights that Meta's interface doesn't show

For challenger brands, this means you get enterprise-grade marketing intelligence without enterprise complexity or cost. You can compete with larger competitors by being smarter, not just bigger.

Real-world implication: A challenger brand with $50K/month Meta spend + AI tools + fast creative production can genuinely outperform an enterprise with 10x the budget using manual optimization and slow creative cycles.

Common Mistakes to Avoid with Meta Andromeda

Mistake 1: Over-optimizing and constantly changing campaigns

Why it fails: Every significant edit resets the learning phase. The algorithm has to start pattern recognition over.

Do this instead: Set 7-day no-touch windows. Resist the urge to "fix" campaigns in the first week. Learning looks volatile. That's normal.

Mistake 2: Creating micro-variations of the same creative

Why it fails: Andromeda optimizes for semantic meaning, not surface-level differences. Five versions of the same ad with different CTAs read as identical to the algorithm.

Do this instead: Create fundamentally different creative concepts that explore different value propositions, pain points, and messaging angles.

Mistake 3: Maintaining complex account structures

Why it fails: Fragmented campaigns split data volume, slowing AI learning. Ten ad sets with $50/day each learn slower than one ad set with $500/day.

Do this instead: Consolidate to 1-3 core campaigns with larger budget pools and more creative variations per campaign.

Mistake 4: Fighting the algorithm with manual rules

Why it fails: Your manual rules (only show ads to people aged 35-44 who like "digital marketing") artificially constrain the AI's ability to find patterns you didn't anticipate.

Do this instead: Guide the AI with strategy (brand positioning, creative inputs, conversion events), but let it handle execution (audience finding, bidding, placements).

Mistake 5: Sacrificing creative quality for quantity

Why it fails: Andromeda needs semantic value, not just volume. Ten poorly executed ads with weak hooks and unclear value propositions don't give the algorithm useful signals.

Do this instead: Focus on diverse, high-quality creative concepts. Ten strong ads outperform 30 weak ones.

Measuring Success in the Andromeda Era

Traditional metrics still matter (ROAS, CPA, conversion rate), but you need additional frameworks to evaluate performance in an AI-optimized system.

New metrics that matter:

1. Creative Performance Distribution

Track which creative angles drive results:

• Which messaging themes resonate? (Problem-focused vs. solution-focused, emotional vs. logical)

• Which formats perform best? (Video vs. image vs. carousel)

• Which hooks stop the scroll? (Questions vs. statements, data-driven vs. story-driven)

• How long does each creative perform before fatigue? (Days to performance drop-off)

This tells you what semantic patterns work for your brand. Double down on those patterns.

2. Learning Velocity

Measure how quickly campaigns reach stable performance:

• Time to 50 conversions per campaign (faster = better algorithm match)

• Campaign stability after learning phase (low volatility = good)

• Cost efficiency improvement over time (CPA decreasing week-over-week = learning working)

Campaigns that learn quickly and stabilize are well-structured for Andromeda.

3. System-Level Performance

Stop obsessing over ad-level metrics. Andromeda optimizes at the system level.

• Campaign-level ROAS (not individual ad ROAS)

• 7-day and 30-day trend lines (not daily performance)

• Month-over-month efficiency gains (are you improving as the AI learns more?)

4. Advantage+ Adoption Impact

Measure the before/after impact of enabling Advantage+ features:

• ROAS change after turning on Advantage+ Audience

• Performance lift from AI-generated creative variations

• Conversion rate improvements with Advantage+ placements

Meta's official data shows 22% average ROAS lift. If you're not seeing similar results, something in your creative or offer needs fixing.

Tools for monitoring:

• Meta Ads Manager (native reporting for campaign performance, creative breakdown, placement analysis)

• DOJO AI (cross-channel marketing intelligence, creative performance analysis, trend identification across all your marketing channels)

• Custom dashboards tracking creative lifespan, semantic theme performance, learning velocity

If you need help exporting Meta Ads data for deeper analysis, check our comprehensive guide.

The key shift: Evaluate trends and patterns, not individual data points. AI systems learn over time. Your measurement needs to reflect that.

The Bottom Line

Meta Andromeda represents a fundamental shift from audience-first to creative-first advertising. The system is no longer rewarding manual optimization expertise. It's rewarding high-quality creative inputs and strategic AI collaboration.

The official data proves the system works:

• +8% ads quality improvements

• +6-12% conversion lifts across GEM and Lattice

• +22% ROAS increases for advertisers using Advantage+ features

• 10,000x increase in model complexity enabling massive creative scale

The opportunity for challenger brands:

Early adopters are already seeing 22% ROAS lifts. Most competitors are still running 2024 playbooks (interest-based targeting, complex account structures, limited creative diversity). That creates a window of competitive advantage that won't last forever.

AI levels the playing field between challengers and enterprises. A smaller brand with $50K/month spend, fast creative production, and simplified Andromeda-optimized campaigns can outperform enterprises with 10x the budget using legacy approaches.

What you need to do:

1. Simplify account structures (consolidate to 1-3 campaigns)

2. Create diverse creative (10-15 variations with different semantic angles)

3. Use broad targeting + Advantage+ (let the AI find your audience)

4. Commit to learning stability (7-day no-touch windows)

5. Measure what matters (creative performance, learning velocity, trend lines)

Ready to adapt your Meta Ads strategy for the Andromeda era? DOJO AI's Marketing Operating System helps challenger brands monitor performance, analyze creative effectiveness, and compete with enterprise budgets using AI-powered marketing intelligence. Start your free trial.

As Meta continues evolving GEM and Andromeda throughout 2026, the gap between AI-native marketers and manual optimizers will only widen. The time to adapt is now.

Further Reading

• AI Marketing Agents: Ultimate Guide 2026 - Learn how AI agents are transforming marketing operations

• Performance Marketing Agents: Mid-Market's Secret Advantage - Why AI-powered marketing intelligence helps challenger brands compete

• How to Export Meta Ads Data for 2026 Budget Planning - Complete guide to Meta Ads data extraction and analysis

• AI-First Marketing Team Structure Guide 2026 - Restructure your marketing team for AI collaboration

About the Author

This article was created by the DOJO AI team, drawing on official Meta engineering documentation, platform-wide performance data, and analysis of hundreds of advertiser experiences during the Andromeda rollout.

Sources & References

• Meta Engineering Blog: "Meta Andromeda: AI-Driven Ads Retrieval" (December 2, 2024)

• Meta for Business: "AI Innovation at Meta: GEM, Lattice, and Andromeda" (March 27, 2025)

• Search Engine Land: "Meta's AI-Driven Advertising System" (January 28, 2026)

• AdExchanger: Industry analysis and revenue reporting (2025-2026)

• Official Meta Ads Manager performance statistics (2024-2026)