Understanding the contribution of different digital marketing channels to customer conversions is crucial for effective resource allocation and campaign optimization. Traditional attribution models, such as first-touch and last-touch, are widely used but often fail to account for the complexity of multi-channel interactions. This study critically evaluates the limitations of heuristic-based attribution methods and presents a data-driven framework leveraging machine learning techniques to improve attribution accuracy. Existing literature predominantly explores rule-based and probabilistic models, such as Markov Chains and the Shapley Value approach, yet these methods often struggle with scalability and real-time adaptability. Recent advancements in machine learning offer new opportunities to enhance multi-touch attribution through predictive analytics. By analyzing clickstream data obtained from Google Analytics 360 via BigQuery, this study constructs a structured four-stage modeling process. The methodology incorporates heuristic models as baselines, a third-order Markov Chain model for probabilistic evaluation, and an XGBoost classification model for predictive accuracy assessment. The results reveal substantial inconsistencies in heuristic-based models, particularly in their allocation of credit among non-dominant channels. The probabilistic Markov Chain approach provides a more balanced distribution of attribution, yet it lacks the flexibility of machine learning-based models in capturing dynamic consumer behavior. The XGBoost model demonstrates superior predictive performance, achieving an overall accuracy of 98.4% and an AUC of 0.846. Feature importance analysis identifies organic search, direct traffic, and paid search as the most influential factors driving conversions. This research advances attribution modeling by introducing a scalable and reproducible machine learning-based framework. The findings offer valuable insights for marketers seeking to refine.