Or offer two search results when they suspect one is spam and see which one a user likes and train off of that, just the way they do now with ChatGPT.