Perhaps it won't take information from the training data but rather take the content from the pages a regular search would return and convert it to a nice answer to your query.

This seems to be the approach https://www.perplexity.ai/ is taking.