Mirror or diverge? Auditing how LLMs shape news agendas relative to youtube and traditional media outlets

As Large Language Models (LLMs) increasingly serve as primary information intermediaries, they assume a critical role in algorithmic agenda-setting. This study conducts a systematic audit comparing the news agendas constructed by ChatGPT and Gemini against traditional media (GDELT) and platform-based media (YouTube). Our results indicate that LLMs do not merely "mirror" upstream media. Instead, they systematically de-prioritize contentious domestic politics compared to traditional news media in favor of "factual utility" topics while maintaining a more even distribution than YouTube. Compared to platform agendas, both LLMs remain semantically closer to legacy journalism. This study contributes to intermedia agenda-setting (IAS) by identifying a "decoupling" of issue selection from issue semantic structure and by repositioning the IAS theory as multidimensional independence in an AI-mediated environment. It also introduces a new measurement framework for understanding how generative AI could reshape the public’s perception of reality in the age of "Communication and Change."