Recommendation systems play a key role in many aspects of life, from housing recommendations to music suggestions. As a result, recommendation systems have become an increasingly significant part of a company’s digital business plan. Given the economic impact of music recommendations, research has suggested that using LLM-generated song summaries can result in better recommendation quality as opposed to using other textual features when designing music recommendation systems. This paper seeks to expand on this idea by examining what sort of textual features are helpful for music recommendations. In particular, we study the impact of 3 types of textual features derived from a song. The first option is to use public information about the song, such as the song name, to generate an input feature. In the second and third options, we prompt a LLM with the artist and song names, and with the song lyrics, respectively, to generate a song summary, which is then used as an input feature for the recommendation model. The fourth option we explore is to use part of the song lyrics as an input feature. The third and fourth options require parsing the lyrics of a song, which may be copyrighted. Our analysis suggests that while the context of the song, such as the song name, already provides improved recommendation performance, a more effective input feature would be to directly use the truncated song lyrics or at least use a summary of the song generated from a LLM as an input feature in cases of copyright barriers.