Research Paper

Core Innovation

This paper introduces MARCO, a framework that decomposes clicks by user intent using logged click types as behavioral labels. It trains separate CVR models per intent and composes their predictions under a predicted intent distribution, improving calibration and conversion prediction accuracy over standard CVR models.

Why It Matters

Accurate conversion prediction is critical for optimizing ad spend and maximizing ROI in digital advertising. MARCO addresses biases in standard models that treat all clicks equally, enabling better targeting of high-intent users. This leads to improved conversion rates and revenue growth at scale for advertisers and platforms.

Potential Customers & Pain Points

• Digital advertisers – Inefficient ad spend due to poor conversion prediction
• Ad tech platforms – Need for better calibration and user intent modeling
• E-commerce companies – Desire higher conversion rates from ads
• Marketing agencies – Require accurate attribution and performance insights

Business Model

SaaS platform or API licensing to ad tech companies and advertisers, offering improved conversion prediction models and analytics for better ad targeting and spend optimization.

Competitive Landscape

• Google Ads
• Facebook Ads
• The Trade Desk
• Criteo
• AppNexus

Implementation Challenges

• Integration complexity with existing ad tech stacks
• Data privacy and user tracking regulations
• Requirement for large-scale click and conversion data
• Adoption resistance due to model complexity

Validation Strategy

• Conduct offline A/B testing on historical ad campaign data to measure calibration and conversion lift
• Deploy online A/B tests with partner advertisers to validate real-world performance improvements
• Monitor key metrics such as conversions per click and topline revenue impact
• Iterate model based on feedback and scale deployment across multiple ad platforms