Exactly. “Children’s clothing” is a completely useless search until the search engine joins it with a million signals you failed to type. I can’t even begin to imagine how people believe every page in the internet could possibly be ranked in a useful way without those extra implicit terms.

I find it interesting how you manage to draw such weird conclusions. "Children's clothing" is not a useless search query. It is very generic or wide and THEREFORE useful.

Running that query through a GPT to transform it into "Popular brand among this bucket of users that spends a lot on ads" + "storefront" is completely useless, because now you have "enhanced" the query itself to a narrow subset of indices. All the fashion, sustainability, diy blogs are immediately filtered out in the query itself. The "frontend" now does not even have a chance at drawing from different buckets to form SERPs - every result is storefront now.

This is the issue discussed in OP. There is one implicit question here. Not only how much we need to tweak input queries to get the search engine to actually cast a wide net and allow discoverability, but whether that is possible at all.