We demonstrate how distortion of content arises endogenously when a content-curating or content-creating algorithm is rewarded for maximizing exposure to particular content regardless of user preferences. We show that when insights from the static model extend to the dynamic setting, characterizing trends in distortion due to fundamentally productive and counter-productive effects. Applying the framework to financial advice by Large Language Models (LLMs), we microfound user utility through portfolio choice. When users suffer from confirmation bias, optimal customization balances productivity enhancements with confirming preferences. Even when they do not, better information quality can paradoxically hurt the precision of users by relaxing their participation constraint, enabling more distortion that offsets the direct gains.