Summary
Performance advertising on the open internet is fundamentally harder than inside walled gardens. Without logged-in identity, deterministic attribution, or control over the auction, bidders must make decisions from fragmented, late-arriving signals under tight latency. Moloco's answer is CARA (Compound Ad Recommendation Architecture), a system of six domains—Campaign Automation, Supply, Ad Recommendations, Bidding, Creative, and Signals—built to compound improvements across each other.
The architecture relies on a shared ML infrastructure for training, serving, real-time features, and experimentation, enabling continuous learning from over 6 billion daily events and model refreshes as often as every two hours. A key engineering practice is using large offline models to train compact online models via knowledge distillation, balancing accuracy with latency constraints. Live experimentation is essential; Moloco runs over 100 concurrent A/B tests to validate changes before deployment.
In 2025, this process produced 65 validated model improvements that reduced cost per install by 14%, cost per action by 17%, and improved return on ad spend by 27% (per Moloco internal data). For ad ops decision-makers, the takeaways are clear: compound AI systems that integrate learning across all decision points yield outsized returns; one must treat the open internet's complexity as a feature, not a bug, and invest in experimentation infrastructure to keep finding new 'hills' to climb. Generative AI is a complement, not a substitute, helping generate creatives and enrich signals, while the core value lies in precise value prediction and real-time decisioning.
Ultimately, CARA demonstrates that a deep understanding of the advertising problem combined with a compound AI architecture can deliver measurable, compounding gains in performance advertising beyond the walled gardens.
Analyst Note
The significance of Moloco's CARA goes beyond the headline numbers. That 14% CPI reduction and 27% ROAS improvement came from 65 discrete, experimentally validated model updates — a pace of compounding iteration that reflects a deliberate infrastructure bet. What's unusual is the architectural focus on the open internet's anomalies: fragmented identity, first-price closed auctions, and non-stationary supply.
By treating these as first-class constraints, Moloco is challenging the assumption that walled gardens hold an insurmountable data advantage. For UA managers and monetization strategists, the key implication is that optimization is now a continuous, compounding process. The reported gains came from constant retraining and experimentation, not a one-time fix.
That environment is already familiar — performance teams live with volatility — but the scale and speed of model refreshes are worth factoring into expectations. The article also underscores the growing importance of infrastructure: knowledge distillation, real-time feature stores, and training consistency are as critical as model architecture. And the mention of agents and generative creative hints at the next phase of automation.
For buyers, this signals that the independent ad tech stack is becoming more sophisticated, and that the trade-off between reach and precision may be narrowing.
You Might Also Like
AppsFlyerCross-platform measurement: the complete guide for 2026
Cross-platform measurement resolves the common problem of fragmented, device-level reporting that inflates ROAS and misallocates budgets. By unifying customer identity across web, mobile, CTV, and other surfaces, marketers gain a single view of LTV and attribution. AppsFlyer provides this via CUID stitching and Product Line grouping, enabling real-time, deduplicated insights without manual BI work. Key benefits include accurate cross-platform ROAS, elimination of duplicate attribution, and reliable data for AI-driven optimization.
AppsFlyerMarketing attribution: what it is and how to measure it
Marketing attribution is critical for connecting spend to revenue, but platform self-reporting and last-click bias distort budget decisions. Single-touch models (first/last-click) are simple but miss the full journey; multi-touch models (position-based, data-driven) are more accurate but require robust data. Mobile attribution is particularly challenging due to ATT, SKAdNetwork, and cross-platform gaps, necessitating a mobile measurement partner (MMP) for independent, deduplicated measurement. Clean attribution data is essential for AI-driven optimization—bad signals lead to bad decisions. Starting with position-based attribution and incrementality testing provides a practical foundation.
AppsFlyerCustomer Lifetime Value (LTV): What It Is and How to Measure It
Customer lifetime value (LTV) is a critical long-term metric for app success, but most marketers measure it per-device, understating true value by 2-5x. Cross-platform LTV stitches together web, app, CTV, and more, attributing all revenue back to the original acquisition campaign. Key drivers include retention (5% increase boosts profits up to 95%), purchase frequency, average order value, and acquisition quality. To improve LTV, focus on retention, cross-platform adoption, and optimizing acquisition by predicted LTV rather than CPI.
AppsFlyerI Built, Shipped, and Marketed a Mobile Game in 14 Days Using Only AI. Here’s What Actually Happened.
One person built, shipped, and marketed a mobile game in 14 days using AI tools, achieving 5,563 installs at $0.39 eCPI on $2,200 spend. MCPs (Model Context Protocol) were critical for agentic workflows. The AI agent CLAW managed ad campaigns via AppsFlyer MCP and BigQuery. Data Locker streamed raw data for analysis. Key takeaway: vendors must offer MCPs for fast, agentic data access; measurement stack (Data Locker, ROI 360, Creative Optimization) is essential for solo teams; human+AI beats AI alone.
AppsFlyer6 reasons why mobile apps set the gold standard for omni-channel measurement
Mobile app measurement has solved the single-channel problems that plague other digital channels—independent attribution (neutral third-party verification), privacy regulation (survived iOS 14.5 with new methods), signal governance (provenance, chain of custody), fraud detection (15% fraudulent installs, 275% fake installs in some channels), and cross-platform fragmentation. These capabilities, built under duress, now form the foundation for omnichannel measurement. Ad ops must apply mobile-grade rigor per channel first, then connect via CUID, unified attribution logic, and real-time data governance to build a trusted cross-platform framework.
AdjustA guide to user testing for mobile apps | Adjust
User testing reveals the gap between designer intent and user experience, uncovering silent churn causes like unclear onboarding or passive ad chains. Analytics show what happens; user testing explains why. Small tests (5-8 participants) can identify friction points, and improving retention by 10% can significantly boost revenue without changing monetization. For ad ops, this means better user engagement reduces wasted ad spend and increases lifetime value.
TikTokIntroducing Symphony Agent | TikTok For Business Blog
TikTok's Symphony Agent is an AI-powered creative engine that helps advertisers produce trend-driven ads at scale. It powers Symphony Creative Studio for video generation from prompts, Content Suite for AI search of relevant creator videos, and TikTok One for streamlined creator matching and outreach. Key benefits include leveraging platform signals to generate authentic content, reducing manual effort, and enabling fast A/B testing. A limited offer provides ad credits for new SMB advertisers spending $100-$1500.
AppsFlyerWhere European Finance Apps Win or Lose Users
European finance app installs hit 960M in 2025 but grew only 0.4%. BNPL apps grew 40% while crypto fell 35%, signaling a shift to utility. Neobanks win acquisition; traditional banks win retention (1.5-2x Day 30 rates). Web-to-app drives 41.8% of conversions but most brands can't measure the handoff. Nearly 1 in 2 investment app installs in Western Europe is fraudulent, distorting CPI and ROAS. Winning brands prioritize engagement, fraud detection, and cross-platform measurement.
More from Moloco
Moloco launches Agency Partner Program with more than a dozen founding partners
Moloco's new Agency Partner Program helps agencies drive profitable performance via Moloco Ads, with education, dedicate...
Discovery is shifting fast: New insights on AI, apps, and the fight for attention
AI is reshaping discovery, with web traffic declining sharply in information-heavy industries (news -26%, education -22%...
How sports apps are preparing for the next wave of fan engagement
Sports publishers face unique monetization challenges due to event-driven traffic spikes. AI-native SDKs like Moloco opt...
Is your channel mix resilient in an AI-disrupted world?
AI is reshaping consumer behavior, with 80% of Google searches ending without a click and half of consumers using AI for...
Gaming Solved App Monetization From Day One. Why Is the Rest of the App Economy Still Playing Catch-Up?
Gaming publishers treat monetization as infrastructure, using hybrid models with ads and in-app purchases to drive reven...
The Singular 2026 ROI Index showcases Moloco as a top partner outside walled gardens
Moloco ranks 5th in the 2026 Singular ROI Index with 29 leaderboard appearances, highlighting consistent ROI across glob...