LLMs are doing numbers well to predict the next letter, basically. It's It's based on so the data fed into training and then whatever prompt it gets (here's that's not just the search but also results). The algorithm is working just as designed, but it's just a guess on how the text should continue