I think the biggest thing Google has working against them is the move to put lightweight models in everything for everyone. Whatever model they use for search is probably somewhere around an ~8B model, and flash 2.0 is decent but still a light weight model.
People now associate Google AI/Gemini with shitty search results and bad answers.
Meanwhile, their SOTA models have been strong, and Gemini 2.5 looks like it might actually have taken the AI throne yesterday.[1]
> People now associate Google AI/Gemini with shitty search results and bad answers
I'm curious how valid this statement is. Anecdotally, I really like those results. For whatever reason, I don't directly interact with Gemini, I just type questions into google and if an AI response comes up I definitely give it a first look before clicking through links. If for some reason no AI response comes up for a question, I copy-paste it into ChatGPT and I usually get the answer I'm looking for.
As I typed that^ I realized something funny. I used to make fun of my parents for typing certain questions into google e.g. "What's the best pizza place in New York?". I would tell them to type "Best pizza place New York", because Google seemed to work better that way. But then that sort of stuff was hijacked by the SEO industry. So the meme of typing "Best Pizza Place New York Reddit" started so we could get actual answers from real people. With the rise of LLMs in search I'm personally back to typing "What's the best pizza place in New York?" right into Google/ChatGPT and it's been pretty successful so far. I'm curious what the next thing will be.
People now associate Google AI/Gemini with shitty search results and bad answers.
Meanwhile, their SOTA models have been strong, and Gemini 2.5 looks like it might actually have taken the AI throne yesterday.[1]
[1]https://simonwillison.net/2025/Mar/25/gemini/