Atlas confidently failed successfully [0]; Kagi [1] and Google [2] nailed it.
This is a perfect example as to why I don't think LLMs should replace search engines any time soon. Search engines help you find the truth. LLMs tell you the truth, even when it's not.
With chatgpt I don't search, I ask. Chatgpt explains, I ask again, and refine and refine. Ask back for sources, etc.
When in doubt, I copy/paste a statement and I search for it with Google. And then Google LLM kicks in.
If it's consistent with chatgpt, I am still wanting to see the links/sources. If not, I notify chatgpt of the "wrong" information, and move on.
70-80% of search is dead. But of course searching for people or things like that, Google is still needed.
But search (the way we know it) was a paradigm that the old Internet created, because it was obviously easier to search for one or two keywords. Semantic search was always something they tried to implement but failed miserably.
Chatgpt is the new way to get information on the internet, like it or not. Even when you think that "it's only trained on recent data, etc", it's only partially an issue, because in many cases it's trained on good information coming from books. And that can be quite useful, much better than a crappy blog that is in the first Google page.
The new paradigm is to use chatgpt as an assistant / someone you can ask information to, in order to answer a question you have. The old paradigm, on the other hand, requires that you start from zero. You need to know already what to search for, in order to get to the fact you wanted to know in the first place. Now it's there, as long as you know a few words.
Hallucinations and misrepresentations aside, my problem with LLMs as search engines is that they purport to give you "the truth" based on the "truth" it thinks you want to see from your query.
While I think LLMs do a much better job of natural language search than search engines do, they also remove the critical thinking that's employed in going through a list of results, processing the information therein and rendering a result.
LLMs might short-circuit the need to trawl through that "crappy blog," but they can, just as well, exclude key information in that blog that adds important context to what you're looking for.
You're doing the right thing by continuously refining your answer and cross-checking it with other LLMs, but I'm sure you've also met many people who use the Gemini AI overview answer and call it a day (which has been hilariously wrong in some searches I've done with it). Regardless, I still strongly believe that well-formed search queries beat rounds of refining LLMs every time.
We aren't making it past the great filter as a species, like it or not.