For vast amounts of data and diverse user needs, accurately customizing search results and delivering personalized content to enhance user experience and platform profitability remain critical challenges in the field of personalized services. From the perspective of dynamic optimization, this paper regards a personalized search and recommendation task as a dynamic qualitative index optimization problem and proposes a transfer learning-driven surrogate-assisted differential evolution algorithm with user generated contents. The proposed solution begins by preprocessing and vectorizing multi-source heterogeneous user-generated content, where textual data is vectorized using a pre-trained language model. Next, convolutional neural networks (CNNs) and singular value decomposition (SVD) are combined to extract user preference and item attribute features, leveraging transfer learning to use them as fixed feature extractors, thereby enhancing feature extraction efficiency in dynamic environments. Subsequently, an attention factorization machine is employed to interact with these extracted features, constructing a surrogate model based on user preferences to approximate the objective function of the recommendation task. Finally, the surrogate model is integrated with a differential evolution algorithm. Within the differential evolution framework, specific evolutionary operators, strategies, and a dynamic model management mechanism are introduced to adapt the evolutionary search process based on user preferences. New evolutionary individuals are generated iteratively with detailed user preference information, refining the recommendation list and enhancing personalization accuracy in each iteration. Extensive experiments on several practical, commonly used real-world datasets verify the feasibility and effectiveness of the proposed method.