Let me tell you about this new little-known technology that's been gaining traction in the last few years...
Even assuming no hallucinations, an LLM can replace some use cases for search but not all.
It did use Google to provide me with the answers though, sooo...
Imagine you’re in the car or hands free or disabled and you just want the question answered.
Such is the case for people trying to poison--ahem "influence" LLM results for "what is the best restaurant in $my_locale". You're playing a dangerous game.
Sure. But that's not what's happening here, yeah? Both examples are providing sources. Actually I'd say that the LLM is doing a better job at that. You click a button and see 9 URLs to 9 news sites.
I think for the "influence LLM results" fears, it would be ridiculous to argue about which multi-billion dollar American company can be trusted more not to manipulate you.
Does it show relevant snippets for each of the 9 URLs? If not, then it's as bad as Google's AI summary. It's not that rare that the reference disagrees with the LLM.
The value of the non-LLM search engine is that in the search results, you see the relevant snippet, and if you want more information, you know quickly which links to click.
Not saying there's no place for the LLM for many search engine use cases, but a proper search engine replacement it is not.
Sounds like you're saying that the LLM isn't any better than Google's LLM-created snippets. The claim was that Google has no moat against LLMs, and your evidence against this claim is Google's LLM generated content.
So what I'm saying is:
Google Search Results > LLM results
And yes, that includes:
Google Search Results > Google Summary on search results page
Google's UX is terrible - but a search engine (a tool that can put you right into contact with the FAQ or forum where actual experts are discussing your problem) is always going to be powerful.
If you want to know what the relevant pages are, you need a search index.
I can assure you that it is not. If you download any model off of huggingface, it does not also include an index of the internet.
That screenshot shows the model making a tool call to an external search index.
Prompt: Who won the 2026 World Cup?
Answer: Spain won the 2026 World Cup, beating Argentina 1-0 after extra time in the final at MetLife Stadium on July 19. Ferran Torres scored the only goal in the 106th minute, coming on as a substitute in the 62nd minute. It’s Spain’s second World Cup title, having also won in 2010.
Prompt: Nearby BBQ places open now?
Answer: (a geolocation permission request prompt for the browser followed by) Right in [redacted] both [redacted] (4.6 stars, open until [redacted]) and [redacted] ([redacted]) are close and currently open.
A bit further out but highly rated: [redacted]
Seems like LLM does a good job on those questions…
The only issue is that the eyeballs stayed with the LLM, which allows it to hold a position that is potentially very powerful. This is particularly problematic with people who believe that computers are infallible.
But the user doesn't necessarily know anything about that, and they don't necessarily care.
If/when the time is reached when external search engines are no longer present in the LLM loop, it seems likely that regular folks won't even notice this shift. As long as the answer-making machine keeps making answers, they won't have any reason to pay attention to this kind of back-end minutiae at all.
It gets its answers from an archive of whatever used to be the Internet, and the last remaining publication: The Costco Flyer.
> How do we double check the answer-machine's answers?
That's the best part: We don't!
and maybe you really need the index and not the engine?
On the LLM side as well, I have not seen much people using Gemini vs the market share of Claude / Open AI.
The assumption of the long tail still using Gemini because it is bundled might be correct, but I am not even sure if that is something that Google will be happy with.