If your Meta Ads have behaved differently in the last 18 months, there's a specific reason. Meta has rebuilt its ad delivery system from the ground up. The new engine is called Andromeda. By October 2025 it was fully rolled out globally, which means every advertiser on Meta is operating under it now, whether they've adapted or not.
Most Andromeda guides you'll find are written for eCommerce or general advertisers. That's a problem if you run mobile apps, because the way Andromeda interacts with your account is genuinely different. SKAdNetwork conversion delays, install-event learning patterns, and app inventory dynamics all change how Andromeda evaluates and delivers your ads. Apply the eCommerce playbook to a mobile app account and you'll get suboptimal results.
This guide explains what Andromeda actually is, how it differs from what came before, what's specifically different about how it treats mobile apps, and what to do about it. We draw on Meta's own published research, current industry analysis from Jon Loomer, Search Engine Land, and others, and how we operate under Andromeda for mobile app clients at The Social Outline.
If you spend meaningfully on Meta for a mobile app, this is the system you're now operating in. Worth understanding it properly.
Key takeaways
- Andromeda is Meta's new ad retrieval engine, rolled out December 2024 and completed globally by October 2025, representing a 10,000x increase in model complexity at the retrieval stage.
- Ad delivery now happens in two stages: Andromeda retrieves around 1,000 candidate ads from billions in milliseconds, then the traditional auction ranks them.
- Andromeda evaluates each ad on three factors: predicted action rate, bid value, and user experience impact.
- Mobile apps face unique Andromeda dynamics because SKAdNetwork postback delays slow conversion signal, install events fatigue audience pools faster than purchase events, and app inventory competes with non-app content in feed.
- Andromeda's semantic analysis of creative content means visually distinct ads with similar psychological angles register as similar in the algorithm's eyes, accelerating fatigue.
- For mobile apps in 2026, optimising for Andromeda means broad audiences, distinct creative concepts (not visual variations), clean conversion data via the Conversions API, and consolidated account structure.
What Andromeda actually is
Andromeda is Meta's new ad retrieval system. It's the first stage of a now two-stage ad delivery process that replaced Meta's previous rule-based approach to selecting which ads compete in each auction.
Meta announced Andromeda in December 2024. The global rollout completed by October 2025. Every campaign on Meta now operates under Andromeda, regardless of campaign type or account size.
When a Meta user opens Instagram, Facebook, or any Meta-owned surface, the platform has potentially millions of eligible ads to serve them. Running the full ranking algorithm against millions of options for every single impression is computationally prohibitive. So Meta split delivery into two stages.
Stage 1: Retrieval (Andromeda). Andromeda's job is to scan billions of ads and narrow them to roughly 1,000 candidates that are eligible for the auction. This happens in milliseconds using a hierarchical tree structure based on user intent, ad characteristics, and semantic similarity. Andromeda doesn't decide the winner. It decides who gets to compete.
Stage 2: Ranking (the auction). The sophisticated bidding system you're already familiar with then evaluates those 1,000 candidates and selects the final ad. This stage hasn't changed as much as Stage 1, but its inputs have.
Andromeda represents roughly a 10,000x increase in model complexity at the retrieval stage. Meta's engineering investment was in building deep neural networks and computer vision capabilities that read the actual content of your creatives (not just the metadata you tag) and match them to individual users in real time.
Andromeda has a sister system called GEM (Generative Ads Recommendation Model) that handles downstream prediction work. The two operate together: Andromeda decides eligibility, GEM helps predict user intent and conversion probability.
How Andromeda differs from the old system
The old Meta ad delivery model was rule-based. You defined the audience, uploaded creative, set the budget. The system delivered ads within the boundaries you specified. Your advantage as an advertiser came from defining those boundaries precisely.
Andromeda flips the logic.
Now Andromeda decides who's eligible to see your ads. It does this by analysing the user's recent behaviour, content interactions, micro-preferences, and contextual signals at a depth the old system couldn't match. Then it cross-references this with your creative content (which Andromeda can read semantically) and your conversion goal.
The strategic implication is that narrow targeting now hurts performance. When you constrain Andromeda to a small audience pool through layered interest targeting, you reduce the algorithm's ability to find high-converting users at scale.
The new playbook, broadly accepted across the industry by mid-2026:
- Run broad audiences (or use Advantage+ which removes targeting restrictions entirely).
- Feed Andromeda strong conversion signal through the Pixel and the Conversions API.
- Compete on creative quality and creative diversity rather than audience precision.
- Consolidate campaigns and ad sets rather than fragmenting them.
The three factors Andromeda evaluates
For every potential ad impression, Andromeda balances three factors in real time:
Predicted action rate. How likely is this specific user to take the conversion action this ad is optimising for? An ad with strong predicted conversion rate can win delivery even with a lower bid.
Bid value. What is the advertiser willing to pay? Higher bids signal commercial value but do not guarantee delivery. A lower bid with a higher predicted conversion rate can outcompete a higher bid with weaker engagement signals.
User experience. Will showing this ad improve or degrade the user's experience on the platform? Ads that get hidden, generate negative feedback, or drive users away from the platform get penalised. Andromeda weights this factor heavily.
For mobile apps specifically, the user experience factor is sometimes underrated. App ads that lead to App Store conversion friction get penalised over time. The algorithm can't directly measure what happens in your app, but it can measure how users behave toward your ads on Meta surfaces and infer.
What Andromeda means specifically for mobile apps
This is the section the eCommerce-focused Andromeda guides don't cover. Mobile apps face Andromeda dynamics that differ from web-based businesses in five significant ways.
SKAdNetwork conversion data feeds Andromeda slower. Web conversions report back to Meta in real time via the Pixel and Conversions API. SKAN postbacks for iOS install events arrive with built-in delays of 24 to 72 hours, and the data is aggregated rather than user-level. Andromeda learns slower on iOS app accounts than it does on equivalent web accounts.
Install events fatigue audience pools faster than purchase events. Lower-commitment actions saturate audience pools faster. Andromeda's retrieval logic responds by tightening the candidate pool as audiences saturate, which compresses effective reach. See our creative fatigue guide for the full diagnostic.
App ad inventory competes with non-app consumer content. Mobile app install ads compete in the same auction as eCommerce, lead-gen, brand awareness, and every other ad type. There's no separate auction lane for app ads. Apps with weaker creative get systematically deprioritised at the retrieval stage.
iOS apps experience different Andromeda dynamics than Android apps. iOS conversion data flows through SKAN and is delayed and aggregated. Android conversion data flows through Google Play Install Referrer and is closer to real-time. Most app marketers benefit from running them as separate campaigns.
App Store conversion is a second layer Andromeda can't see directly. Your creative drives the click. The App Store page drives the install. Apps where the App Store page underperforms the ad creative end up training Andromeda on a degraded conversion signal.
The Creative Similarity factor
This is the deepest connection point between Andromeda's mechanics and what most mobile app advertisers experience as "creative fatigue."
Andromeda's retrieval engine uses semantic analysis to evaluate ad content. The model doesn't just look at the targeting parameters or the metadata you tag. It reads the actual creative. It evaluates the imagery, the messaging, the emotional tone, the framing, the call to action. From these signals it builds an understanding of what each ad is "about" at a conceptual level.
It then uses that understanding to evaluate similarity. Two ads that share visual elements, messaging structure, emotional charge, or audience framing get clustered as similar in the retrieval logic, even if they look visually distinct to a human viewer.
This is why visual refresh stopped fixing creative fatigue.
Three different fitness app ads with three different creators, three different colour palettes, and three different opening hooks can all hit the same psychological notes: "transform your body in 30 days." Same valence, same self-concept hook, same intensity. To Andromeda's semantic analysis, these are three variants of the same ad.
The operational answer is creative diversity at the conceptual level, not the visual level. Ten ads built around ten different ideas will beat fifty variations of one idea.
How to optimise for Andromeda as a mobile app advertiser
1. Feed Andromeda clean conversion signal. If your conversion data is incomplete or noisy, Andromeda optimises toward the wrong things. Specifically:
- The Pixel installed and firing correctly across all relevant events
- The Conversions API set up for server-side event delivery (critical for SKAN-affected accounts)
- SKAdNetwork configured with the right conversion value schema for your monetisation model
- Mobile Measurement Partner (MMP) integration for app event passback (AppsFlyer, Adjust, Branch)
2. Run broad audiences or Advantage+ campaigns. Andromeda is built to do the targeting work. Broad audiences with strong creative consistently outperform narrow audiences with the same creative on Andromeda accounts.
3. Compete on creative diversity, not variation. Ten distinct concepts across different psychological zones will beat fifty variations within one zone.
4. Consolidate account structure. The old playbook of fragmenting into dozens of ad sets with different audiences hurts you under Andromeda. Fewer campaigns, fewer ad sets, more creative within each.
5. Respect the learning phase. Andromeda's learning phase under SKAN-affected iOS conditions can run 7 to 10 days, not the 3 to 4 days the system flags.
6. Match creative format to channel reality. Most app budget on Meta runs on video and static formats. Playable ads convert better but only run at scale on AppLovin and ironSource.
Common mistakes mobile apps make with Andromeda
Over-fragmenting campaigns. Splitting into iOS vs Android plus by country plus by audience type plus by creative theme produces dozens of underfunded ad sets. Most app accounts should run 3 to 5 campaigns, not 20.
Killing the learning phase too early. App accounts with SKAN delays look worse in days 1 to 5 than they actually are. Hold for at least 7 days before making any campaign-level decisions.
Adding interest layers to Advantage+ campaigns. Advantage+ is designed to remove targeting restrictions. Adding interest layers on top defeats the purpose.
Refreshing creative within psychological zones. Andromeda's semantic analysis sees through this. Refreshes need to be conceptually distinct, not just visually different.
Optimising for the wrong SKAN conversion value. Many app accounts set their SKAN conversion value schema based on the install event alone, when their actual monetisation depends on a Day-2 or Day-7 in-app event.
Frequently asked questions
Will Andromeda work better with Advantage+ campaigns for mobile apps?
Yes. Advantage+ App Campaigns are designed to give Andromeda maximum flexibility in audience retrieval and creative testing. Meta is investing most heavily in Advantage+ as the primary structure going forward, and Andromeda's performance is strongest on these campaign types. For most mobile app advertisers in 2026, Advantage+ should be the default campaign structure.
How does Andromeda handle SKAdNetwork data?
Andromeda processes SKAN postbacks the same way it processes other conversion signals, but with the inherent delays and aggregation limits SKAN imposes. The practical implication is slower learning phases and more variance in early campaign performance for iOS app accounts. Cross-reference dashboard performance with postback-corrected views over 7 to 14 days before making campaign decisions.
Should I run separate iOS and Android Andromeda campaigns?
In most cases, yes. iOS and Android conversion data flow through different attribution systems with different latencies and granularity. Running them in separate campaigns lets Andromeda learn each platform's conversion logic independently rather than averaging across them. The exception is small-budget campaigns where consolidation produces better learning than fragmentation.
How long is the learning phase under Andromeda for mobile apps?
7 to 10 days for SKAN-affected iOS campaigns. 3 to 5 days for Android campaigns with cleaner attribution. Meta's exit-learning-phase notification fires earlier than this, but performance is often still volatile until full conversion signal has arrived. Wait for at least one full SKAN postback window plus a 3-day stabilisation period before making major changes.
Does Andromeda work with lookalike audiences?
Yes, but with caveats. Lookalikes act as a seed signal Andromeda uses to refine retrieval, not as a hard targeting constraint. Lookalike-1% and lookalike-3% audiences typically outperform broader lookalikes on Andromeda accounts because they give the algorithm cleaner intent signal. But Advantage+ campaigns without lookalikes often match or beat lookalike-targeted campaigns for apps with healthy conversion data flow.
How does Andromeda affect creative refresh strategy for mobile apps?
Significantly. Visual refresh alone no longer resets fatigue because Andromeda's semantic analysis sees psychological similarity through visual variation. Effective refresh strategy under Andromeda is concept-based, not visual-based. Rotate the underlying psychological angle (valence, identity hook, language intensity) rather than just the imagery.
Want this run for you?
Andromeda has made the operating model harder. Strong conversion data, broad audiences, distinct creative concepts, and clean account structure are now table stakes. The agencies and in-house teams that have adapted are pulling ahead. The ones still running old-system playbooks are watching their CPIs climb.
At The Social Outline we run performance creative for mobile apps spending £25k or more per month on Meta. If you want a system built for the algorithm you actually run on, apply to work with us.