Small, repeatable experiments on how AI assistants pick what to recommend and what to cite. Each one shows the data.