It's also conceivable that companies running LLMs might consider various tweaks to popular questions as they appear in media so as to inflate the model's answer

It's also conceivable that companies running LLMs might consider various tweaks to popular questions as they appear in media so as to inflate the model's answers. No way to learn for sure, but competing companies where PR spells short term disasters, wouldn't put it past them to intervene when necessary. I am not familiar with how one augments and already trained model or if if there is even an architecture that's capable of that, short of input proxies in front of the model. I am curious to learn more if anyone knows.