There's nothing remotely similar to Google Search retrieval happening underneath ChatGPT (as we today have it, anyway). The language model associates certain words with other words on a scale hard for us to imagine, but it does not tie any of them to any specific retrievable documents.

That's why being very specific and detailed about the information was more likely to guide it toward "right" answers - those words are pushing the associative machine into a corner where some of its associations might be relevant. Or then again, they might not be.