In systems that have a quality component (e.g.

In systems that have a quality component (e.g. search engines, or LLMs), the query stream coevolves with the underlying quality of the service.

Users as a population clue into what it can do and give it queries it will do well at.

There's always some bizarro random queries that don't work, but the creators of the service will typically study those carefully and use that to decide which quality improvements to prioritize (a variant of "pave the cowpaths").

Over time, those random queries help extend the service in ways that other users come to expect to work, too.

When there's a huge first mover like ChatGPT, everyone else will be evaluated on how well they perform on queries optimized for ChatGPT primarily.

If services like Gemini don't do well on those, then users might say "eww not as good as ChatGPT" and not come back.

So services like Gemini have to be at least good enough on ChatGPT-style queries, and then actively great at some other differentiated things that some users might think to try.