How large AI models filter information and construct their own information loopholes has always been a mystery, but researchers have been trying to uncover some of its secrets. I came across an interesting study attempting to understand the "taste list" of media outlets cited by AI models. This study primarily used over 10,000 responses generated in the 18 days leading up to the US midterm elections for a large-scale audit, roughly outlining how the "information funnel" is constructed when AI models answer voters' political questions. This study utilized the specific window of opportunity before the midterm elections, but it offers valuable insights for those working on "GEO" (Geographical Occurrence Analysis) and considering the "bias" of AI models (after all, this is where bias is most easily detected, as media political leanings are very obvious). Several interesting findings (note that these findings are all based on answering voters' politically relevant questions): - Each model also has its own "dietary habits," meaning it has its preferred primary news source: GPT prefers the Associated Press, Claude chooses the traditional Yahoo and its content aggregation network, Gemini favors PBS, and Grok surprisingly chooses The New York Times (Musk doesn't seem to like the NYT very much);