It’s more like:
Traditional search pipes first page of results to the bot. The bot reads the pages from the results and tries to identify an answer or the best result from the set. Both the bot summary and the adjusted ranking for the results are returned. This gives a chance at a better experience for the user because they don’t have to read all the pages themselves to try find the answer they were looking for. However there is a huge margin for error since the bot is underpowered due to Google balancing the amount they pay for each search with the amount they earn for each search. So there end up being misinterpretations, hallucinations, biased content etc.
If they used a top end model like Claude Sonnet 3.7 and piped it enough contextual information, the AI summaries would be quite accurate and useful. They just can’t afford to do that and they want to use their own Gemini bs.
It’s more like: Traditional search pipes first page of results to the bot. The bot reads the pages from the results and tries to identify an answer or the best result from the set. Both the bot summary and the adjusted ranking for the results are returned. This gives a chance at a better experience for the user because they don’t have to read all the pages themselves to try find the answer they were looking for. However there is a huge margin for error since the bot is underpowered due to Google balancing the amount they pay for each search with the amount they earn for each search. So there end up being misinterpretations, hallucinations, biased content etc.
If they used a top end model like Claude Sonnet 3.7 and piped it enough contextual information, the AI summaries would be quite accurate and useful. They just can’t afford to do that and they want to use their own Gemini bs.