This study presents EcoPoetico, a multimodal web app designed to enhance digital literary mediation by recommending poems based on semantic relatedness. The app addresses key challenges in digital reading platforms, such as the ambiguity of literary mediation in online contexts and the cognitive challenge of processing poetry, which is often problematic in curriculum design and not always favored by teachers. Using a corpus of \(\sim \) 5,000 Spanish-language poems, EcoPoetico employs the BERT-based language model Beto and cosine similarity to offer personalized poetry suggestions. To evaluate its performance, the study conducted both usability and recommendation tests. The usability test showed that both students and teachers found EcoPoetico engaging and easy to use, with potential for teaching purposes. However, teachers emphasized the need for Explainable AI to enhance the platform’s educational value by providing insights into how recommendations are generated. The recommendation test confirmed the system’s effectiveness, as cosine similarity resulted in recommendations that aligned more closely with expert criteria for adequacy, significantly outperforming random suggestions. A cosine similarity analysis reveals that the similarity score is not the primary factor in determining the adequacy of a recommendation, but rather its variance among the recommended poems. Higher variance allows for better discrimination and leads to more effective recommendations. EcoPoetico’s open-source code and modular architecture allow for easy adaptation to other languages or corpora.