Why would they, there is nothing for them to gain. Their content gets trained upon without being able to even consent and they gain nothing, not clicks, ad revenue etc.
So in answering a question, it also still uses the wider dataset. The references aren’t necessarily the source of the whole answer.