Beyond the Surface: Probing the Ideological Depth of Large Language Models