MARCO Decomposes Click Intent for Calibrated Ad Conversion Prediction

Shiwen Shen, Xiru Huang, Liang Luo, Jianbo Sun, He Lyu, Zihang Fu, Ivonne Xu, Zhizhuo Li, Zhengyu Zhang, Pei-Ju Sung, Yunmiao Wang, Zixuan Wang, Zhengli Zhao, Qiang Jin, Mike Jermann, Mingda Li, Yang Xiao, Bhavana Challa, Brooke Bian, Yang Li, Ashish Chamoli, Bibek Bhusal, Danning Di, Yuan Jin, Meet Raval, Zhiwen Chen, Boyao Sun, Shuguang Wang, Yunlong He, Yantao Yao, Sagar Chordia, Wenlin Chen, Santanu Kolay, Qin Huang, Ellie Wen· August 12, 2026 View original

Key takeaways

  • Treating all ad clicks equally leads to biased conversion rate predictions.
  • MARCO decomposes clicks by user intent, training specific CVR models for each intent.
  • This approach significantly improves per-intent calibration and overall conversion rates.
  • Deployment showed a +2.80% lift in conversions per click and +0.98% in topline metrics.

Who benefits

AdTechE-commerceRetailMarketingMedia

Summary

This research introduces MARCO, a framework that improves ad conversion prediction by decomposing clicks based on user intent, recognizing that not all clicks are equal. By training per-intent conversion rate models and composing their estimates, MARCO corrects calibration biases and significantly lifts conversions per click and overall topline metrics.

Current industrial ad ranking systems often treat all clicks uniformly when predicting conversion rates, despite clear evidence that different types of user interactions with an ad signal varying levels of intent. This uniform treatment leads to biased predictions, where high-intent clicks are underestimated and low-intent clicks are overestimated, even if aggregate calibration appears accurate. The MARCO (Multi-intent Ads Ranking Composition Optimization) framework addresses this by disaggregating clicks based on their underlying intent. It leverages logged click types as a free behavioral label to train distinct conversion rate (CVR) models for each intent group. At serving time, MARCO combines these per-intent CVR estimates using a predicted distribution over intents. The theoretical underpinnings of MARCO demonstrate that this decomposition never increases population risk and provides a clear path to improved estimation and calibration. Online deployment of MARCO, even with binary intent granularity, has shown remarkable results: per-intent calibration improved to nearly 100%, conversions per click increased by 2.80%, and cumulative topline metrics saw a 0.98% improvement.

Why it matters

Marketing and sales professionals can achieve significantly more accurate ad conversion predictions, leading to optimized ad spend, higher ROI, and improved user experience by better matching ads to true intent.

How to implement this in your domain

  1. 1Analyze existing click data to identify distinct user interaction patterns that could signal different intents.
  2. 2Implement a system to log and categorize various click types as behavioral labels for intent.
  3. 3Develop and train separate CVR models for each identified click intent, as proposed by MARCO.
  4. 4Integrate the MARCO framework into ad serving systems to dynamically compose per-intent CVR estimates.
  5. 5Conduct A/B tests to validate the impact of intent-based CVR prediction on key advertising metrics.

Original post by Shiwen Shen, Xiru Huang, Liang Luo, Jianbo Sun, He Lyu, Zihang Fu, Ivonne Xu, Zhizhuo Li, Zhengyu Zhang, Pei-Ju Sung, Yunmiao Wang, Zixuan Wang, Zhengli Zhao, Qiang Jin, Mike Jermann, Mingda Li, Yang Xiao, Bhavana Challa, Brooke Bian, Yang Li, Ashish Chamoli, Bibek Bhusal, Danning Di, Yuan Jin, Meet Raval, Zhiwen Chen, Boyao Sun, Shuguang Wang, Yunlong He, Yantao Yao, Sagar Chordia, Wenlin Chen, Santanu Kolay, Qin Huang, Ellie Wen

"arXiv:2608.10562v1 Announce Type: new Abstract: Not all clicks are equal. Industrial ads ranking decouples conversion probability into click-through rate (CTR) and post-click conversion rate (CVR), yet treats every click as the same event. In reality, users provide a free, self-g…"

View on X

Originally posted by Shiwen Shen, Xiru Huang, Liang Luo, Jianbo Sun, He Lyu, Zihang Fu, Ivonne Xu, Zhizhuo Li, Zhengyu Zhang, Pei-Ju Sung, Yunmiao Wang, Zixuan Wang, Zhengli Zhao, Qiang Jin, Mike Jermann, Mingda Li, Yang Xiao, Bhavana Challa, Brooke Bian, Yang Li, Ashish Chamoli, Bibek Bhusal, Danning Di, Yuan Jin, Meet Raval, Zhiwen Chen, Boyao Sun, Shuguang Wang, Yunlong He, Yantao Yao, Sagar Chordia, Wenlin Chen, Santanu Kolay, Qin Huang, Ellie Wen on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses

More in AI in Marketing

AI in MarketingAI Engineering & DevToolsAI in Sales

Causal Optimization Boosts LinkedIn Feed Marketing by 7.2%

This paper introduces a decision-centric framework for large-scale targeting and recommendation systems that optimizes for incremental impact rather than just predictive scores, using a causal neural network, a Bayesian neural-bandit layer, and a dual-based linear programming layer. An online A/B test on LinkedIn Feed marketing traffic showed a 7.20% lift in long-term value.

Changshuai Wei, John Bencina, Phuc Nguyen, Andre Assuncao Silva T Ribeiro, Benjamin ZelditchAug 12, 2026
AI Engineering & DevToolsAI in MarketingAI Research

Ex-Omni-2D Creates Expressive Omni-Modal Dialogue with Visual Avatars

Ex-Omni-2D is an omni-modal dialogue framework that generates coordinated responses including text, personalized speech, and reference-conditioned video, giving AI avatars a native visual presence. It uses a Visual Thought Plan and a shared acoustic-temporal interface to learn from heterogeneous data and enables efficient incremental generation.

Haoyu Zhang, Zhipeng Li, Xiaoying Tang, Tianshu Yu, Yiwen GuoAug 12, 2026
AI Engineering & DevToolsAI ResearchAI in Marketing

New AI Handles Incomplete Multimodal Sentiment Analysis

Researchers propose MIDAS, a unified framework for multimodal sentiment analysis that effectively handles incomplete or corrupted inputs by disentangling shared and exclusive latent factors. It uses an uncertainty-aware fusion mechanism to robustly integrate features, outperforming existing methods across various incomplete data scenarios.

Yuhua Wen, Yingying Zhou, Qifei Li, Yingming Gao, Zhengqi Wen, Jianhua Tao, Ya LiAug 12, 2026