Does a URL just sitting in a prompt steer an LLM's output toward its content?

This article investigates whether including a URL in an LLM prompt influences the model's output toward the content at that URL. Experiments show that a URL only steers output if its content was in the model's training data. Many sites that rely on JavaScript are missing from training corpora. Descriptive URLs can influence as normal text, while famous opaque identifiers (like arXiv IDs) may decode if memorized. The research highlights the opacity of training data and the impact of SPA design on model knowledge.