Post

I feel like this changed as soon as they got web search integrated ( which varied by platform )

That's my finding with CoPilot. It does a good job of summarizing the results, and I can go to them, if needed. The pithy paragraph it adds at the end is best ignored.

Spot on! (but… ->). Compared to most Google searches and unfortunately to Google competitors results it’s shockingly better. But I thinks that’s ~60% (guessing) caused by the intense enshittification google search has been through during the latest ~8 yrs.

Obsiously. "LLM is bad at doing..." is a bit oversold here.

I’m primarily relying on Anthropic for code-related API use at this point.

It searches when it needs to - occasionally I will say "run a search" but mostly, I have found that to be unnecessary

I loved o3 for web research. Recently I've been using Perplexity since they have been giving out free subscriptions.

I have to concur. For a while (~>1-2 months ago) Perplexity was superior, but either it degraded in it answers or ChatGPT with search has become much better. Maybe a bit of both. Either way: While I used Perplexity nearly exclusively for a while, I now only use it occasionally to check on it.

AI mode replaced perplexity for me. i would prefer AI mode to be automatically selected based on the question

Have you tried AI Mode with Gemini 2.5 Pro on? It's a bit slower, but I feel like it provides even better results : www.google.com/search?udm=5...

Seems extremely fast, though not necessarily better than Perplexity

Fair question. I’ve seen a lot of posts showing nonsense answers from the overview, and asking how to turn it off. Maybe most people aren’t questioning it at all.

It’s not Google’s reputation either — AI overviews are the most visible application of LLMs for millions. They make everyone look bad.

AI image generators were terrible for a long time - and then suddenly, they weren't. Seems like a general pattern - so makes sense that search is working better.

Google's "AI overviews" remain one of the worst implementations of AI-assisted search Confusingly, Google's new "AI mode" is genuinely really good! www.google.com/ai

You mean there's been progress since then? What do you mean I have to update my understanding and try things more often than every few years to keep track of progress?

Great post, a one to one match with my experience. I found one teeny tiny typo though.

Did you try the same searches with GPT-5 Auto? It would be interesting how often it would automatically select Thinking to come to the same results. Also interesting: Would Perplexity "Best" select GPT-5 or GPT-5 Thinking or another model for these searches?

I suspect the effect will be quite dramatic when it becomes possible to have it search through paywalled academic literature

Ah, yes, that makes a lot of sense!

I like the way the name contains an Oxford comma, but GPT-5 tactfully calls it a “serial” comma.