Meta's ad system has undergone its most significant architectural shift since the introduction of machine learning–based delivery. If you run Meta ads in any capacity, Andromeda 2.0 is already deciding which of your ads get seen and which ones never make it past the gate. This guide breaks down what changed, what the data says, and exactly what to do about it.
Meta Andromeda 2.0 is the 2026 evolution of Meta's AI-powered ad retrieval engine, fundamentally changing how ads are selected before they ever reach the auction. Here's what matters most:
Meta Andromeda 2.0 is Meta's second-generation AI retrieval engine for Facebook, Instagram, and Audience Network ads. It became fully live by October 2026 and acts as the first filter in ad delivery across all of Meta's advertising platform surfaces.
In plain language, here's what happens every time someone opens their feed: Andromeda 2.0 scans tens of millions of eligible ads and shortlists roughly 1,000–2,000 candidate ads for the downstream ranking stages and auction. This process determines which ads even get a chance to compete. Andromeda evaluates over one billion ads per second, making retrieval decisions at a speed that was impossible just two years ago.
The 2.0 version builds on the original 2024 Andromeda launch with several major upgrades. It introduces a retrieval Mixture-of-Experts (MoE) model, where specialized user towers handle different behavioral patterns rather than relying on a single user tower. In other words, Andromeda 2.0 enhances user tailoring by utilizing multiple specialized AI models instead of a single general model. Meta also co-designed the system with NVIDIA Grace Hopper (GH200) hardware for maximum throughput.
Andromeda identifies ads based on user engagement and contextual signals for better targeting. It's always on in the background-there is no switch in Ads Manager to disable it or opt out. Every single Meta ads campaign runs through this retrieval engine whether you know it or not.
Meta ads delivery now happens in three main stages: Retrieval (Andromeda 2.0), Ranking (using architectures like Meta Lattice and GEM), and Auction/Pricing. Understanding this pipeline is essential because your campaign performance depends on clearing each stage.
Here's how the stages connect:
Retrieval is now far more computationally intensive than before, yet the system delivers results in under 300 milliseconds to avoid feed delays. Meta's advertising infrastructure allows for deeper architecture and feature interactions within this tight latency window. The system can now consider a larger candidate pool than before, meaning more ads are evaluated per impression.
What's critical to understand: improvements or failures at the retrieval layer have downstream effects on CPM, CTR, CPA, and conversion rates. But advertisers only see surface metrics in Ads Manager. You never see a "retrieval score." The system decides which ads advance, and you see the results as better or worse delivery.
Index matching inside retrieval uses targeting rules, creative semantic signals, and user behavior patterns to decide which ads belong in the shortlist for a specific person. This is where the magic-and the frustration-happens.
Andromeda 2.0 represents a major architectural revamp released progressively through 2025, with production benchmarks published in early 2026. This wasn't a minor tuning pass; it was a ground-up rethink of how the retrieval algorithm operates at scale.
Key technical upgrades include:
On the user modeling side, the shift to a retrieval Mixture-of-Experts model means specialist "experts" focus on different user and ad patterns instead of one monolithic network. Andromeda uses a mixture-of-experts model for ad scoring, where each expert captures different facets of contextualized user interests-from shopping behavior to content consumption patterns.
The system also ingests richer signals now: sequences of engagement (view → click → save), user cohorts, semantic understanding of ad creatives (visual elements, on-screen text, audio cues), and session-level context-what a specific person was just doing in the app.
Meta reports a 2.4% increase in Facebook conversions since 2025, attributed partly to Andromeda improvements. In Q3 2025, Meta published that combining retrieval and early ranking improvements produced roughly 14% improvement in ads quality on Facebook surfaces. Andromeda continues to evolve in response to privacy changes affecting ad delivery, relying more on internal behavioral signals as third-party data diminishes.
Andromeda 2.0 doesn't just rank individual ad IDs in isolation. It works on conceptual clusters often referred to by practitioners as Entity IDs-and understanding this mechanism is one of the biggest strategic advantages you can have.
Here's how the pieces connect:
Index matching is the process by which the retrieval index stores ads by semantic and behavioral features. When an impression opportunity arises, Andromeda 2.0 uses index matching to rapidly find relevant ad candidates from the index. This is far more sophisticated than a simple keyword lookup-it evaluates visual similarity, copy meaning, offer structure, and predicted engagement patterns.
An entity ID is a conceptual grouping. Andromeda clusters similar ads into a single entity ID. If you have 20 ads showing the same product on the same white background with only minor headline tweaks, the retrieval system treats them as one "ticket" into the auction. They share retrieval fate.
Classic identifiers like campaign ID, ad set ID, and ad ID still exist in Ads Manager. But Andromeda 2.0's internal clustering layer is what governs retrieval probability. Your ad set might contain 12 ads, but if Andromeda clusters them into 2 entity IDs, you effectively have only 2 retrieval tickets.
The strategic implication is clear: winning more distinct Entity IDs via truly different ads is vastly more valuable than inflating the number of cosmetic variations. Each genuinely distinct concept earns its own entity ID and its own chance of being retrieved for each impression opportunity.
Before Andromeda, ad delivery leaned heavily on manual targeting-interest stacks, lookalike audiences, demographic slices. Andromeda 2.0 largely inverts this relationship. Audience targeting is now driven by creative content, not settings.
Keywords in copy, on-screen text, products visible in frame, and overall tone now function as soft targeting signals for the retrieval model. If your video ad shows a woman running in a city park at dawn while narrating about "fitting in a workout before the kids wake up," Andromeda reads those cues and delivers the ad to people whose behavior matches that scenario-even with broad targeting settings.
Broad targeting and Advantage+ setups give Andromeda 2.0 maximum room to interpret creative intent and match ads to the right person. Heavy interest targeting, by contrast, can actually "fight" the retrieval logic by constraining the search space before Andromeda can do its work.
Manual targeting is not dead. Location, language, age, and necessary exclusions still matter. But obsessively narrow interest stacks are less effective under Andromeda's system than they were 18 months ago. Think of manual targeting as setting the boundaries of the playing field, while your ad creatives determine where the ball actually goes.
For example, an ad explicitly addressing "busy new parents in small apartments" will naturally be directed to that cohort by Andromeda, even without selecting "new parents" as an interest in Ads Manager.
Since late 2025, performance reports from advertisers and agencies paint a nuanced picture. Some accounts see higher reach at weaker conversion rates. Others see efficiency gains after restructuring campaigns for the new system.
Many media buyers noted increased CPM and CPC right after the Meta Andromeda update rollout, especially as the system recalibrated in late 2025. One large-scale analysis of $110 million in ad spend found CTR rose about 9% and cost per click dropped roughly 6%-but conversion rates from clicks to orders dropped approximately 23%, cost per purchase rose 22%, and ROAS fell 26%.
Return on ad spend declined by 7% during the Andromeda rollout across a broader dataset of $834 million in spend from 3,000 advertisers. Top performers lost 31% in return on ad spend after Andromeda, which sounds alarming until you understand that "top performers" had the most optimized legacy setups and thus the most disruption from new retrieval logic.
Meta reports a 2.4% increase in Facebook conversions since 2025 at the platform level. But third-party datasets show a wide spread: some ad accounts lose ROAS while others improve by double digits after adapting their creative strategy and campaign structure.
Emerging patterns most advertisers should watch for:
There is a critical distinction between creative volume (many files) and creative diversity (many distinct ideas). Andromeda 2.0 rewards the latter and largely ignores the former.
If you upload 30 ads showing the same product on the same background with minor text tweaks, they likely collapse into one entity ID, giving you just a single effective auction ticket. Creative diversity prevents ads from clustering into a single Entity ID.
Genuinely different ads should represent different buying motivations:
Ads with diverse creatives achieve 44% higher ROAS compared to accounts running variations of the same concept. Top performers run 33% more ads than bottom performers-but the key differentiator is concept variety, not file count.
A practical range: aim for 8–15 distinct concepts per campaign, then rotate and refine based on performance rather than launching hundreds of near-duplicates.
"Fake diversity" vs. "Real diversity":
The "Entity ID framework" is a practical method to maximize distinct retrieval chances under Andromeda 2.0. The goal: ensure each new ad concept earns its own entity ID rather than being absorbed into an existing cluster.
Two main levers control clustering:
Here's an example for a fitness app targeting three segments:
Each combination of persona + visual environment + benefit creates a sufficiently different concept that Andromeda's semantic and visual similarity checks should classify it as its own entity ID.
Quick checklist-is your new ad a new Entity ID?
If you check at least two of those boxes, you likely have a new entity ID.
Many media buyers are used to tight manual targeting, lookalike audiences, and layered interest stacks to control ad delivery. Under Andromeda 2.0, that playbook needs updating.
Broad targeting aligns better with Andromeda 2.0 because it opens more branches in the retrieval index and lets creative meaning guide where ads appear. When you restrict audiences tightly, you're shrinking the pool of individual users that Andromeda evaluates, which can reduce both retrieval eligibility and learning speed.
Narrow targeting is less effective under Andromeda's system than it was before, but manual targeting still matters for:
The recommended hybrid approach: keep essential constraints (country, age range, exclusions) but test at least one broad ad set per account where Andromeda 2.0 has maximum freedom. Compare performance against your manually targeted ad sets over 2–4 weeks to see which approach delivers stronger results for your specific situation.
Fragmented campaign structures with many small ad sets are now typically outperformed by consolidated, broader setups. If you're running 15 overlapping interest-based ad sets with $20/day each, Andromeda doesn't have enough data in any single one to learn effectively.
Best practices for structuring a Meta ads campaign under Andromeda 2.0:
A simple structure that works for most advertisers:
There is no official Meta rule on ad count, but real-world data reveals clear patterns. The sweet spot lives between "too few to give Andromeda options" and "too many to let any single ad learn."
Ads with 5+ creative variants achieve 38% higher performance compared to ad sets running 1–2 creatives. But going above 15 active ads per ad set often splits learning and ad spend so thin that none exit the learning phase.
Recommended ranges by budget:
More Entity IDs provide more retrieval tickets, but the trade-off is real. If each ad only gets a handful of impressions, Andromeda can't learn which creative resonates with which audience segment.
Base your final decision on your own account data rather than blindly following a magic number. The ranges above are starting points-your ad accounts will tell you where the real sweet spot lives.
Andromeda 2.0 "reads" your ad creatives. It analyzes visuals, copy, and even audio cues to decide who each ad is for. This means your creative assets are doing double duty: selling to the customer and signaling to the retrieval algorithm.
No single ad format-video ads, static image, or carousel-is universally favored. But each surfaces different types of behavioral signals. Using multiple formats increases retrieval opportunities by three times compared to running a single format.
Match format to message:
Write hooks and on-screen copy that directly call out specific people or situations. Phrases like "For remote workers who..." or "If your kitchen is under 100 square feet..." make Andromeda's job easier by providing clear semantic signals about who the ad is for.
Different formats with different ads covering different angles give the retrieval engine maximum flexibility to find the right person for each concept.
Andromeda 2.0 doesn't change auction mechanics directly, but it shifts how efficiently your ad spend reaches qualified users. Think of it this way: better retrieval means your budget competes in auctions where your ad actually has a chance of winning, rather than being wasted on impressions where it never had a shot.
Poor creative diversity or misaligned targeting can lead to wasted impressions, higher CPM, and worse ROAS, even when bids and budgets stay the same. Conversely, improved retrieval can lower CPA over time for advertisers who adapt-more relevant impressions mean fewer wasted clicks and more efficient conversion paths.
To track ad spend efficiency under Andromeda 2.0:
Strong measurement and tracking are even more critical now because Andromeda 2.0 learns from post-click behavior to improve future retrieval. High fidelity signals flowing back to the system directly influence which of your ads deliver in the future.
Essentials to get right:
Index matching uses historical performance, user-level event data, and creative semantics to decide which ads belong in the shortlist for each impression. If your conversion data is noisy or delayed, retrieval and ranking both suffer.
Avoid making major creative, budget, and tracking changes all at once. Stagger adjustments to isolate what actually improves or hurts campaign performance.
Here's a step-by-step blueprint for updating your testing process specifically for Andromeda 2.0:
Step 1: Audit current ads by concept Map every active ad to its core buying motivation (price, convenience, social proof, status, etc.). Identify how many genuinely distinct concepts you're actually running versus how many are cosmetic variants of the same concept.
Step 2: Design 5–10 new, clearly different concepts Each should target a different persona, pain point, or benefit. Vary visuals, format, and tone.
Step 3: Launch in a consolidated, broad campaign Use broad audiences. Don't fragment across multiple ad sets by interest.
Step 4: Run for 7–14 days without over-editing Resist the urge to change budgets or swap creatives during this window. Let Andromeda learn.
Step 5: Cull losers and scale winners based on business KPIs Use cost per purchase and ROAS as primary decision metrics-not CTR or CPC alone.
Test concepts (new benefits, use cases, offers, angles) rather than micro-variations like color shifts or trivial headline changes. Each new ad should teach you something about customer motivations under Andromeda 2.0, not just whether blue outperforms green.
Andromeda 2.0 changes who arrives at your site. It often sends colder, more top-of-funnel traffic than legacy interest targeting did, because the retrieval engine is matching creative signals to new audiences who haven't interacted with your brand before.
Landing page conversion rates fell by 17% across advertisers during the Andromeda transition, largely because many sites weren't designed for this colder traffic profile.
To adapt your landing pages:
The gap between what your ad promises and what your landing page delivers is now more costly than ever. A mismatch doesn't just lose the sale-it trains Andromeda to deprioritize your ads.
Andromeda 2.0 impacts small, mid-sized, and large advertisers differently because of varying budgets, creative capacity, and organizational speed.
Mid-sized advertisers often have an agility advantage. They have enough ad spend to test multiple distinct concepts but lean enough processes to turn around new creatives in days rather than weeks.
Large advertisers may struggle with slow approval pipelines, brand guideline rigidity, and fewer radical creative swings-which can hurt retrieval diversity under Andromeda 2.0.
Small advertisers face creative resource constraints but can compensate with strong concept variety across even a handful of ads.
Vertical-specific considerations:
Catalog Ads, Collection Ads, and other dynamic formats naturally generate creative diversity at the product level. Each product becomes its own entity, giving the retrieval engine many different entry points.
Catalog Ads generate 3.6 times more revenue than static ads in aggregate performance data. Top performers run 72% more Catalog Ads than bottom performers. Catalog Ads maintain a 20-30% higher ROAS over static ads across most verticals.
Large catalogs with 5,000+ SKUs produce many Entity IDs automatically, giving Andromeda 2.0 far more retrieval options. Each product image, description, and price point creates a distinct creative instance.
If you have a small catalog, compensate by:
These strategies expand perceived creative diversity even when your product count is modest.
Several popular misconceptions have emerged since the 2.0 rollout. Let's clear them up.
Misconception 1: "Upload hundreds of variations to increase reach." Simply uploading many similar ads doesn't work because Andromeda clusters them into shared Entity IDs. More ads only help if they represent genuinely distinct concepts.
Misconception 2: "Andromeda replaces the need for good offers and landing pages." Retrieval can't fix a weak business model. Andromeda gets your ad in front of people-but ad quality, pricing, and post-click experience still determine whether they convert.
Misconception 3: "Interest targeting is completely dead." Interest targeting still shapes index matching boundaries. It's weaker than before but not irrelevant. Use it as a guardrail, not a primary strategy.
Misconception 4: "Short-term volatility means the system is broken." Andromeda causes natural volatility during learning and recalibration phases. Watch trends over several weeks and across multiple campaigns before making structural changes. Overreacting to daily swings usually makes things worse.
Here's a concrete checklist you can apply in the next 30 days to align with Andromeda 2.0:
Set measurable goals: improve Event Match Quality above 6.0, reach a target cost per purchase, or increase the number of effective Entity IDs in your account. Review progress in 2–4 weeks.
Adaptation-not panic-is what separates brands that thrive after algorithm updates from those that stall. The advertisers winning under Andromeda 2.0 aren't the ones with the biggest budgets. They're the ones willing to rethink creative, structure, and measurement from first principles.
These FAQs address practical concerns not fully covered in the main sections. Answers are based on public documentation, Meta published benchmarks, and real advertiser experience through mid-2026.
Andromeda 2.0 does not remove manual targeting options, but it reduces their relative influence compared to creative content and behavioral signals. Manual targeting now primarily sets hard boundaries-countries, age ranges, exclusions-while the retrieval engine optimizes within those bounds using creative semantics and user interests.
Keep essential constraints in place while testing at least one broad campaign so Andromeda 2.0 can fully leverage creative-first matching. Most advertisers find that the broad campaign quickly reveals which creative concepts naturally find their audience without manual help.
There is no official Meta rule, but most advertisers perform well with roughly 4–6 active, distinct concepts per ad set at modest budgets. Larger budgets can support 8–12 concepts effectively.
Uploading 50 cosmetic variations of one idea often collapses into a single Entity ID, offering no real retrieval benefit and extending the learning phase. Focus on concept diversity-different benefits, personas, and stories-rather than chasing arbitrary high creative counts.
Andromeda 2.0 itself does not inherently raise prices. But it can expose inefficiencies: weak creatives or over-segmented structures may see higher CPM and CPA because the retrieval engine deprioritizes ads that don't match well with any user segment.
Advertisers who adapt with broad targeting, strong creative diversity, and solid tracking often report stable or even lower blended acquisition costs over time. Monitor cost per result and ROAS before and after structural changes rather than focusing on CPM alone.
Video is not mandatory, but it is highly useful because it provides richer signals-motion, audio, storytelling-for the retrieval model. Static images, carousels, and Catalog Ads can still perform extremely well when they convey clear benefits and distinct concepts.
Run a mix of different formats and let performance data-not assumptions-decide your long-term format split. The same video repurposed across every ad set doesn't help nearly as much as diverse formats carrying different messages.
The practical learning window remains roughly 7–14 days or around 50 optimization events per ad set, consistent with pre-Andromeda guidance. Andromeda scores over one billion ads per second, but learning still requires enough conversion data to calibrate.
Making frequent edits-budget, bids, or major creative swaps-restarts learning and makes it hard to interpret results. Set clear thresholds (maximum acceptable CPA over a minimum number of conversions) and evaluate only after campaigns reach those benchmarks. Patience during the learning phase is one of the highest-leverage things you can do in the next generation of Meta advertising.
