Meta Ads Andromeda Algorithm Explained: How It Changes Your Campaign Strategy in 2026

What Andromeda Actually Changed

Before Andromeda, Meta’s ad system used a relatively simple auction model: advertisers defined audiences, Meta matched those audience segments to groups of ads, and the highest bidder with acceptable relevance won the impression. The system evaluated roughly 1,000 candidate ads per user per auction [1].

Andromeda expanded that evaluation to approximately 10,000,000 candidate ads per auction, a 10,000x increase in model capacity. More importantly, the matching logic shifted from audience-to-ad-group to individual-user-to-individual-ad [1] [2].

According to Manel Gomez (Crece sin Limite, $3.3M in Meta ad spend), this shift has four operational components that advertisers need to understand [1]:

1. GEM (Generalized Embedding Model): Processes user behavior into dense vector embeddings
2. Lattice: Unifies ranking across all placements into a single model
3. Andromeda retrieval: Handles the 10,000x scale ad delivery
4. Sequence Learning: Tracks user journeys across multiple ad interactions

Multi-expert consensus (Gomez, Dara Denney, Charley T from Disruptor School) confirms the practical takeaway: under Andromeda, your audience targeting settings matter less and your creative diversity matters more [1] [2] [3].

GEM: Why Behavior Beats Interest Targeting

GEM (Generalized Embedding Model) is the behavior processing layer. It converts user actions across all Meta surfaces (Facebook, Instagram, Messenger, WhatsApp) into dense vector representations of behavioral intent [1].

What GEM processes:

Signal Type	Examples	Weight
Direct engagement	Clicks, form submissions, purchases	Highest
Content consumption	Video watch time, scroll depth, time on post	High
Social signals	Shares, saves, comments, reactions	Medium
Cross-platform behavior	WhatsApp opens, Messenger interactions	Medium
Historical patterns	Purchase history, seasonal behavior, lifetime value	Context layer

Before GEM, Meta relied heavily on self-declared interests and lookalike models built from seed audiences. GEM replaces this with real-time behavioral matching. A user who watched 3 SaaS demo videos, saved 2 B2B marketing posts, and spent 45 seconds reading a case study generates a behavioral embedding that GEM can match to your ad, regardless of whether that user falls into your selected interest categories [1].

The practical consequence for advertisers: Interest-based targeting becomes a speed bump, not an accelerator. According to Charley T (Disruptor School, managing $100K+/day accounts), “broad targeting with creative doing the targeting” is now the standard operating approach. GEM identifies high-intent users faster and more accurately than any interest stack an advertiser can build manually [3].

When interest targeting still helps: New accounts with fewer than 50 pixel conversions benefit from interest guidance during the cold start period. Once the pixel accumulates enough data (typically 50-200 conversions), GEM has sufficient behavioral signal to outperform manual interest selections [3] [4].

Lattice: Unified Ranking Across All Placements

Before Lattice, each Meta placement (Feed, Stories, Reels, Explore, Messenger, Audience Network) ran its own ranking model. An ad that performed well in Feed had no bearing on its Reels performance because the systems were separate [1].

Lattice consolidates all placement ranking into a single model. Performance data from one placement directly informs delivery decisions on every other placement [1] [2].

What this means operationally:

Before Lattice	After Lattice
Separate ranking per placement	Unified ranking across all placements
Manual placement selection could outperform auto	Advantage+ placements outperform manual in most cases
Creative optimized for one format dominated	Creative diversity across formats rewarded
Placement-specific learning curves	Cross-placement learning acceleration

According to Dara Denney, Lattice is why the recommended creative format mix is now 30-40% static images, 30-40% video, 10-15% carousel, and the remainder experimental formats (UGC, text-heavy, AI-generated). Each format performs differently across placements, and Lattice optimizes the format-placement match per individual user [2] [5].

The cross-placement learning effect: If your static image converts well in Feed but your video converts well in Reels, Lattice does not treat these as separate data points. It synthesizes both into a unified understanding of which users respond to which creative types in which contexts. This compounds over time, making accounts with diverse creative formats increasingly efficient [1] [2].

Baker’s recommendation: Use Advantage+ placements as the default. Manual placement selection is only justified in two scenarios: regulated industries requiring specific placement controls, or when you need to isolate placement-specific creative performance for testing purposes [3].

Andromeda Retrieval: The 10,000x Scale Engine

The retrieval layer is Andromeda’s core delivery mechanism. It evaluates approximately 10 million candidate ads per auction (up from roughly 1,000 in the pre-Andromeda system) and selects the optimal ad for each individual user at each moment [1].

The scale difference is not incremental. It is architectural. Pre-Andromeda, Meta could only evaluate a small fraction of eligible ads per auction, which meant audience targeting was necessary to narrow the candidate pool. Post-Andromeda, the system can evaluate essentially all eligible ads simultaneously, making narrow audience definitions redundant [1] [2].

What this means for campaign structure:

• Fewer campaigns, broader targeting. The algorithm no longer needs audience segmentation to function. According to Charley T, accounts spending $20K-$50K/month should operate with 2-5 campaigns, not 20-30 [3]
• Creative diversity is the new targeting. Instead of 5 campaigns targeting 5 audiences with the same ad, run 1-2 campaigns with 16+ distinct creatives. Andromeda matches the right creative to the right user automatically [2] [5]
• Consolidation accelerates learning. More data flowing through fewer campaigns means faster learning phase exits. The 50-conversion-per-week threshold becomes easier to hit when budget is not split across dozens of ad sets [3] [4]

Entity ID: How Andromeda Measures Creative Variation

A critical technical concept under Andromeda is the Entity ID. According to Ryan from Leadbase ($1B+ in Meta spend analyzed), Andromeda does not measure creative variation by Creative ID (the individual ad). It measures by Entity ID, which represents the underlying visual concept [6].

Practical example:

Scenario	Entity IDs	Creative IDs	Andromeda Assessment
10 ads, same video with different text overlays	1	10	Low variation
10 ads, same product photo with different copy	1	10	Low variation
10 ads, 10 completely different visual concepts	10	10	High variation
50 ads, 5 visual concepts with 10 variations each	5	50	Moderate variation

The minimum for Andromeda to optimize effectively is 6 distinct Entity IDs. The target for scaled accounts is 6-10 distinct Entity IDs, each with 3-5 variations, totaling 20-50 active ads [6].

The “16 different beats 50 similar” rule: Running 16 visually distinct creatives outperforms 50 variations of the same concept. Each distinct Entity ID unlocks a different audience pocket within Andromeda’s retrieval system. Variations of the same concept compete for the same audience pocket, creating diminishing returns [2] [6].

Sequence Learning: The Full-Funnel Multiplier

Sequence Learning is Andromeda’s user journey tracking component. It maps the sequence of ad interactions each user has over time and learns which creative sequences drive conversions [1].

How Sequence Learning works:

1. User sees educational video ad (awareness)
2. User later sees testimonial ad (consideration)
3. User later sees product/offer ad (conversion)
4. Sequence Learning records this journey and identifies which sequences convert

The algorithm does not just optimize individual ads. It optimizes the sequence of ads a user encounters. This is why full-funnel creative strategies outperform single-message campaigns under Andromeda [1] [7].

What Sequence Learning rewards:

• Creative diversity across awareness stages. Accounts with educational, social proof, and direct response creatives give Sequence Learning more permutations to test
• Longer attribution windows. Sequence Learning tracks journeys across days and weeks, not just single sessions
• Consistent brand signals. The algorithm can identify when brand consistency across the sequence improves conversion rates

According to Manel Gomez, this is why his portfolio approach (one CBO per awareness level: educational, authority, comparison) outperforms single-CBO structures at scale. Each CBO feeds different creative into Sequence Learning’s journey optimization [1] [7].

Baker’s observation from client accounts: B2B SaaS accounts running full-funnel creative (problem-aware, solution-aware, product-aware, comparison) see 25-40% lower ENCAC than accounts running only bottom-funnel direct response creative. Sequence Learning needs diverse creative across the buying journey to optimize effectively.

Baker’s Andromeda Adaptation Framework

Based on the four Andromeda components, Baker’s Andromeda Adaptation Framework maps five operational changes that every Meta Ads account should implement [1] [2] [3].

Change 1: Shift Budget from Audience Testing to Creative Testing

Pre-Andromeda Approach	Post-Andromeda Approach
70% of effort on audience research and testing	70% of effort on creative production and testing
5 audiences with 2 creatives each	2 broad audiences with 10+ creatives each
Test interests, lookalikes, custom audiences	Test hooks, formats, concepts, angles
Audience is the targeting lever	Creative is the targeting lever

The reason is mechanical: GEM already identifies high-intent users better than manual audience selections. Your job is to give Andromeda diverse creative so it can match the right message to the right user [2] [3].

Change 2: Consolidate Campaigns (Fewer Is Better)

Andromeda’s 10,000x retrieval capacity means narrow audience segmentation is no longer necessary for the algorithm to function. Campaign consolidation directly improves performance [3] [4].

Monthly Budget	Pre-Andromeda Campaigns	Post-Andromeda Campaigns
Under $3K	3-5 campaigns	1 campaign (ABO, broad)
$3-10K	5-10 campaigns	2 campaigns (Test ABO + Scale CBO)
$10-30K	8-15 campaigns	3 campaigns (Test + Scale + Retargeting)
$30K+	10-20+ campaigns	3-5 campaigns (portfolio CBOs by awareness level)

Multi-expert consensus (Charley T, Dara Denney, Manel Gomez) confirms: accounts that consolidated from 15+ campaigns to 3-5 campaigns consistently saw improved performance, lower CPAs, and faster learning phase exits [3] [2] [1].

Change 3: Diversify Creative Formats (Feed the Lattice)

Lattice’s unified ranking rewards accounts that provide creative across multiple formats. The recommended mix [2] [5]:

Format	Share of Total Creative	Purpose
Static images	30-40%	Fast to produce, strong for messaging tests
Video (15-60 seconds)	30-40%	Highest engagement, Reels and Stories delivery
Carousel	10-15%	Product comparisons, step-by-step education
Experimental (UGC, text-heavy, AI)	10-15%	Unlock underserved placements and audiences

According to the Meta Creative Benchmarks Report (500K+ ads analyzed), only 5-8% of ads become winners (defined as earning 10x the account’s average spend). This hit rate is consistent across all budget tiers. Top spenders do not have better ideas. They have a production machine that constantly outputs diverse variations [5].

Account Tier	Weekly Creative Output	Expected Winners
Under $15K/month	3-5 new ads	0-1
$15-50K/month	10-15 new ads	1-2
$50K+/month	15-25+ new ads	1-3

Change 4: Build Full-Funnel Creative Sequences

Sequence Learning optimizes the journey, not just individual ads. Accounts need creative at every awareness stage [1] [7].

Baker’s Andromeda Creative Sequence:

Stage	Creative Type	Goal	Example
Problem-Aware	Educational content, industry data	Generate recognition	”78% of B2B SaaS companies overspend on Meta Ads”
Solution-Aware	Framework explanations, how-to	Build consideration	”The 3-Layer approach to Meta Ads structure”
Product-Aware	Case studies, testimonials, demos	Drive evaluation	”How [client] reduced ENCAC by 74%“
Comparison	Feature comparisons, social proof	Convert	”Baker vs. in-house: the numbers”

Each stage feeds Sequence Learning with different creative, enabling the algorithm to orchestrate the optimal journey for each user [1].

Change 5: Use Cost Caps to Control Marginal Spend

With Andromeda’s expanded retrieval, the algorithm can spend aggressively on high-confidence impressions and wastefully on low-confidence ones. Cost caps prevent overspend on the marginal tail [8].

According to Andrew Faris (AJF Growth), most brands overspend because blended ROAS hides declining marginal returns. Cost caps cut the unprofitable tail of spend without requiring daily management [8].

Implementation:

• Start cost cap at 1.25x-1.5x your target CPA to avoid choking delivery
• Example: $50 CPA target = $62.50-$75 initial cost cap
• Reduce the cap gradually as the algorithm learns
• Evaluate weekly or monthly, not daily [8]

The Andromeda Diagnostic Checklist

Use this checklist to assess whether your account is adapted for Andromeda [1] [2] [3]:

Diagnostic	Andromeda-Ready	Needs Restructuring
Active Entity IDs	6+ distinct visual concepts	Fewer than 6 distinct concepts
Campaign count	2-5 campaigns	10+ campaigns
Audience targeting	Broad or Advantage+	Interest stacking with 5+ layers
Creative format mix	3+ formats (static, video, carousel)	Single format dominance
Funnel coverage	Creative at 3+ awareness stages	Only bottom-funnel direct response
Placement strategy	Advantage+ placements	Manual placement restrictions
Learning phase status	Under 20% of budget in Learning Limited	Over 50% in Learning Limited
Creative refresh cadence	3-5+ new ads per week	Fewer than 3 new ads per month

If more than 3 items fall in the “Needs Restructuring” column, your account structure is pre-Andromeda and likely underperforming relative to its budget.

Budget-Tier Implementation Guide

Baker’s Andromeda Adaptation Framework adjusts implementation by budget tier, because creative production capacity scales with spend [3] [4].

Under $3K/Month

• 1 ABO campaign, 3-5 ad sets, broad targeting
• Minimum 6 distinct creatives (prioritize static images for fast production)
• Add 1-2 new creatives weekly
• Do not split prospecting and retargeting
• Focus: find 1 winning creative concept before scaling

$3K-$10K/Month

• Testing ABO (20% of budget) + Scaling CBO (80%)
• 10+ distinct creatives across 2-3 formats
• Add 3-5 new creatives weekly
• Graduate winners from testing to scaling at $900+ spend with 10+ conversions at target CPA
• Advantage+ placements on scaling CBO

$10K-$30K/Month

• Testing ABO (15%) + TOFU CBO (60%) + Retargeting CBO (25%)
• 15+ distinct creatives across 3+ formats
• Full-funnel creative sequence (problem, solution, product stages)
• Weekly creative refresh mandatory
• Begin building Sequence Learning with diverse awareness-stage creative

$30K+/Month

• Portfolio approach: multiple CBOs by awareness level
• 20-50 active ads with 6-10 distinct Entity IDs
• 15-25 new creatives per week
• Cost caps on scaling CBOs to control marginal spend
• ENCAC as north star metric, evaluated weekly [9]

Common Andromeda Mistakes

Mistake 1: Treating Broad Targeting as “No Targeting”

Broad targeting under Andromeda is not the absence of targeting. It is delegating targeting to GEM, which has 10,000x more signal capacity than your manual selections. The creative itself becomes the targeting mechanism. A video about “SaaS metrics for CFOs” naturally attracts CFOs at SaaS companies, regardless of whether you selected those interests manually [2] [3].

Mistake 2: Running 50 Variations of the Same Concept

Andromeda groups creatives by Entity ID. Fifty variations of the same product photo with different copy count as one Entity ID. The algorithm needs conceptual diversity, not copy diversity. Sixteen distinct visual concepts outperform fifty variations of the same concept [6].

Mistake 3: Manual Placement Restriction

Lattice’s unified ranking almost always outperforms manual placement selection. Restricting to Feed-only or Stories-only prevents cross-placement learning and reduces the audience pool Andromeda can access. Use Advantage+ placements unless regulatory requirements mandate otherwise [1] [2].

Mistake 4: Ignoring Sequence Learning

Running only bottom-funnel direct response ads gives Sequence Learning nothing to work with. The algorithm cannot optimize a journey if the only creative available is “buy now.” Educational, social proof, and comparison creative at different awareness stages enables Sequence Learning to orchestrate higher-converting journeys [1] [7].

---

Sources

1. Manel Gomez, Crece sin Limite. “Andromeda Architecture: GEM, Lattice, Retrieval, and Sequence Learning Components.” $3.3M Ad Spend Analysis and Portfolio Framework, 2026.
2. Dara Denney. “Creative Diversity Under Andromeda: Format Mix, Persona Grouping, and Entity ID Variation.” Post-Andromeda Creative Strategy, 2026.
3. Charley T, Disruptor School (Meta MBA graduate). “Broad Targeting, Campaign Consolidation, and the Two-Campaign Model Under Andromeda.” $100K/Day Account Management Framework, 2026.
4. Meta Platform Documentation, Multi-Expert Analysis. “Learning Phase Requirements: 50 Weekly Conversions, Budget Thresholds, and Reset Triggers.” Meta Ads Optimization Framework, 2026.
5. Meta Creative Benchmarks Report. “500K+ Ads Analyzed: 5-8% Winner Rate, Creative Volume vs Quality, and Format Mix.” Industry Benchmark Analysis, 2026.
6. Ryan, Leadbase. “Entity ID Concept and Creative Variation Requirements Under Andromeda.” $1B+ Meta Spend Analysis, 2026.
7. Thomas Owen. “TPS Framework (Test, Prove, Scale) and Full-Funnel Creative Sequencing.” Andromeda-Era Campaign Operations, 2026.
8. Andrew Faris, AJF Growth. “Cost Cap Myths, Manual Bidding as Scale Tool, and Marginal Spend Management.” Post-Andromeda Bidding Strategy, 2026.
9. John Moran, Tier 11. “ENCAC as North Star Metric and First-Click CAPI Attribution.” Beauty Account Case Study, 2026.