I have a confession. I pay twenty dollars a month for a glorified link finder.
Not for code generation (though that's nice). Not for writing emails (hell no). No, my main use for LLMs is asking them to go read hundreds of web pages for me so I can ignore the report they write and just grab the links out of it.
This is fine. This is a totally normal thing to do with a language model.
How this whole mess started
Every major LLM these days has a "research" mode. OpenAI calls it Deep Research. Google has one too. Anthropic just calls it Research, as if they own the word. They all work the same way: you type a question, maybe it asks you one clarifying thing, then it disappears for five to fifteen minutes, reads through web pages, and eventually spits out a long report with citations.
I got hooked the first time I used it. Not because the report was good, it definitely was not. What got me was that I was researching something I knew nothing about, and I had no idea what keywords to type into Google. You know that feeling when you're so early in learning a topic that you don't even know what to search for? The LLM solved that. I gave it a vague, rambling question and it figured out the vocabulary for me.
Also, I have a visceral hatred for SEO content. I do not want to read "10 Best Practices For" anything. I want to throw my laptop across the room every time I land on a page that is clearly written by someone who has never done the thing they're writing about. LLM-powered search is great at avoiding this, at least in my experience. It finds stuff from the weird old web, the kind of pages Google buried years ago.
The report problem
The actual report the LLM writes? I barely read it.
It is, in a word, slop. Verbose. Overblown. It twists its own sources and forces everything into sections with transitions nobody asked for. All the literary charm of a quarterly report drafted by committee.
The links, though. The links are gold.
My workflow has become: submit question, walk away, come back, skim the report to understand its skeleton, middle-click every link into a new tab, close the report, and read the actual human content. I am using a cutting-edge AI system as a glorified feed reader. Is this efficient? No. Do I care? Also no.
There should be a mode where the LLM just returns a ranked list of links. "Here, I found these. Go read them yourself." I could probably prompt for this, but I keep forgetting, because I am chaotic.
The strange places it takes me
The web pages the LLM digs up are genuinely surprising. I cannot figure out how it decides what to click. What search engine is it using under the hood? What keywords did it search for, and what makes one link worth visiting over another? I have no idea.
But the results are weird in the best way:
- Personal websites that were last updated when George W. Bush was in office.
- Columns from magazines that went defunct before TikTok existed.
- Ancient Blogger and LiveJournal posts from people who have probably forgotten they wrote them.
- Pages hidden inside some megacorporation's labyrinthine support site, impossible to find through normal navigation.
- University lecture notes hosted on public-facing directories.
- PDFs sitting in exposed
wp-contentfolders.
All the stuff search engines are trained to bury. The web that Google has been scrubbing from its index to make room for soulless, SEO-optimized content.
Once it even cited a page that did not exist anymore. The link was dead. How do you cite a page you cannot read? I had to find it on the Internet Archive. I have no explanation for how the LLM knew about a page that was already gone. This is either very impressive or a glitch in the simulation. Possibly both.
On trust and hallucinations
I do not trust LLMs with facts. If I ask one a question and I cannot verify the answer, I assume it is wrong. This is especially true when I am learning something new, because I lack the context to catch errors. If Claude tells me the best way to do something in Rust, and I have written maybe twelve lines of Rust in my life, how would I know if it is lying?
But when the LLM grounds its answer in web pages it actually found and read, and then cites those pages, I can go read them myself. I can check whether the source is a real person with actual expertise, or just another content farm that feeds me its own slop. It's not perfect, but it's a lot better than asking a black box.
This also means I get information that is more current than the training data. That part works well, though not always.
The wishlist I keep in my head
The feature I want most: I want to see the damn search results. See what keywords the LLM decided to search for. Let me edit its research plan before it starts browsing. Tell it "actually, ignore anything from that domain, I hate that website."
Gemini lets you edit the research plan a little bit. That is nice. I want more.
- Give me lenses like Kagi has: "only search personal blogs" or "only academic sources" or "only whatever weird niche forum has the answer."
- Let me uprank and downrank sources.
- If the LLM finds something mid-research that reframes my whole question, let it interrupt itself and ask for clarification. I want recursive meta-research. A research dream.
- And for the love of god, stop hiding web search behind a menu. Let me click "New Search" the way I click "New Chat."
I know Perplexity exists. I have tried it. It is okay. Kagi Assistant exists too. Also okay. Neither has clicked for me the way this janky Claude workflow has. Maybe I am just attached to the drama of waiting seventeen minutes for a report I will not read. It makes me feel like I am doing something.
So is it worth it
I like LLM-powered search. It is broken and slow and verbose and often unreadable, but it is the only way I have found to reliably reach the good parts of the web that traditional search gave up on.
I am paying twenty dollars a month for a machine that browses the web like a curious, somewhat deranged librarian from the early 2000s. And honestly? I think that is money well spent.