Generative AI is increasingly used to produce marketing content, but off-the-shelf models are often misaligned with business goals. Standard fine-tuning approaches rely on experiments that identify what content correlates with higher engagement. Optimizing only on these outcomes, however, risks overfitting, reward hacking, and poor generalization—often leading models to drift toward clickbait. Learn about a new framework that enables large language models to learn why content works. The model automatically generates and validates hypotheses about underlying mechanisms—such as emotional tone or framing—and combines this knowledge with observed performance when generating new content. Using data from over 23,000 A/B-tested news headlines, we show that this approach outperforms standard methods in improving engagement while remaining more robust and less prone to clickbait, because it is grounded in generalizable principles rather than surface correlations.