The Kagi user base has self-selected to want a search experience that can better the "big tech" search process that optimizes for ad conversions over search results. In a way, Google/Bing/etc. are incentivized to offer just good enough results that you don't give up entirely, but offer as many ads and bad results to get the most impressions/clicks.
I built a prototype on SearXNG that would anonymously track linger time on results for a given query, if a user searched for XYZ, and clicked a link then closed it and returned quickly to look for another result, that link would be downranked. The "terminal result", i.e., the last one accessed by the user as part of that search was upranked, and links visited for a modest amount of time were either kept neutral or slightly upranked.
By doing a lightweight embedding on the query, you can find a nearby neighborhood of searches and corresponding automatic user-submitted (but anonymous) up/down ranks for those results.
This should be something that's very explicitly opt-in as it would be collecting anonymous data from users. The code I wrote simply started a timer when a link was clicked, and if focus returned to the search results tab/window within 30s, the result would be lowered in rank, and if the tab was closed/not returned to, then it would be upranked. One of the challenges is for users who perform a search, open multiple results in tabs, and then work through them without returning to the results.
I see this as a seamless crowd-sourced improvement on the existing collated ranking of domains, but one in which user utility of a result can result in improvements for other users, and have it more closely tied to the semantics of a search query.
