Google has certainly gone downhill, but after trying to use Perplexity a couple months ago after a friend espoused how great it was, I quickly gave up on it. I was mostly using it to figure out specific technical terms related to architecture. Google was feeding me so much SEO'ed crap from random construction companies that just repeated the same stuff.
At first Perplexity seemed great, even providing links to the the sources it was drawing its answers from. But after the fifth or so time of not finding the terms it was claiming as real in any of its cited sources, I gave up. And the sources it was citing weren't even particularly good, just the same ones Google was surfacing.
Since everything could just be a hallucination, I was wasting even more time using it than Google. And even when Perplexity isn't hallucinating, I still can't just how trustworthy its sources are without clicking on them, which is another huge problem with searches like that. The context the information is presented in matters as much as the information itself.
Have you tried Kagi? I’ve set all my devices to use it, and I’m at the point where I get annoyed when something goes to Google by mistake and I see crappy results. Kagi just gives much better results.
Did you try any of your queries using Perplexity Pro? (even at the free tier they give you a few 'Pro' queries a day) While it's still far from perfect, the Pro answers are generally higher quality than the free ones.
I'm finding several of the LLM's that can cite sources, including Perplexity, are more useful to me than Google for search these days. The notable exception is Gemini which has been quite bad in my experience compared to the other options.
I've switched to using Claude for most of my "queries" that are not me searching for a particular document and virtually all of my coding questions. It's very good when you have a question about a broader topic that you want to be able to ask lots of specific questions about.
The experience really reminds me of when Google original appeared on the scene in the early 2000s. It was the first time on the web that you could easily find anything and the first search engine that didn't feel overrun by spam.
There have been many topics I've wanted to understand better that essentially require a dive into multiple wikipedia pages. Claude makes wandering around that graph unnecessary and really accelerates the learning/exploration process.
It also allows queries that were previously impossible, for example "tell me if this movie has a happy ending but otherwise don't give me any spoilers".
Initially I wasn't sure I would personally ever use LLMs as an alternative to search, but as I've learned to treat them differently than a standard search engine I find for many cases there's no reason to return to the ever decaying mess that is Google in it's current form.
Append a ? to a Kagi search and you get the best of both worlds: it'll add LLM results (Claude at the moment) to a box at the top of your search. I think it's some sort of RAG model that they built on top of their search engine.
You can try the LLM half of it here without an account:
edit: I tried the perplexity link in the article. For my test query, the text it generated was similar, but it didn't cite any websites to back up its claims. That's useless! I thought having the LLM link to sources was already table stakes.
I asked a speculative question to FastGPT@Kagi and it returned a "Dunno" as the three Wikipedia pages it suggests having checked did not contain an obvious reply.
I asked the same to Perplexity and it returned a plausible speculative logical reply with more credible reference links.
There is a strong difference between "looking for documents" and "asking sensible questions".
FastGPT@Kagi could neither reply to the question nor find web pages that could be interesting to extrapolate the answer.
I've came to use Brave search [1] lately, and find it is super convenient with the auto-AI-based answers based on the top search results (or at the click of a button if it isn't triggered automatically), combined with the no-frills design (once a hallmark of Google(!)).
The ability to ask various questions right from the browser location bar without login is convenient and a surprisingly big deal IMO.
It's insanely bad. I asked Perplexity: "What are some places named after Alexander Hamilton?" It's answers: Fort Hamilton, Ohio (technically correct, except this is now known as simply Hamilton, not Fort Hamilton); Hamilton, Ontaria; Hamilton, NZ; Hamilton, NSW. None of those last three are correct.
I asked Google the same question and it has the Wikipedia page "List of things named after Alexander Hamilton" as the top organic result, with nothing ranked above it (ads included). It also offered an AI summary that isn't totally, flagrantly incorrect, listing Hamilton Place, Hamilton Hall, and Hamilton Heights in New York.
People who write these articles about search quality are blinded by some unstated ideology.
I ran it through Perplexity Pro which gave a very different and detailed answer with this closing note: "It's worth noting that not all places named "Hamilton" are necessarily named after Alexander Hamilton. For example, Hamilton, Ontario in Canada is named after George Hamilton, a Canadian merchant. Similarly, Hamilton in Scotland and New Zealand are named after different individuals."
That seems like a good litmus test question since it has a single correct answer, but a lot of potentially incorrect ones which sound like they might be right (North America has a lot of Hamiltons after all).
AI chatbots as search interfaces will only exacerbate the current issues with google search. What I personally need is a search engine that actually returns me the documents containing my search term and respects the search operators I provide.
Wow, this looks awful. The cited sources are questionable at best, and on my first attempt it's already hallucinating. How can you trust something like this?
Also, how do I get it to return documents? I searched for 'RFC 793' with the intend to read the actual RFC. Instead, the chat bot summarized it for me (which is pointless) and doesn't even use the actual RFC as a source?
Since this is a "Google bad" thread, I feel obligated to mention Kagi which I've been using for nearly a year and it has significantly improved my search experience (better search quality, great customizability).
I mentioned Kagi in another response, but I love having test queries in these search engine threads. Appending ? to your kagi search will add LLM results like these:
For me that gives a link to 793, a newer related RFC that wikipedia mentions, and a draft attempted replacement from 2021. (So, three links to standards and one to wikipedia.)
You could mistrust it since it is based on an uncorrected hallucinating technology, but use it either in contexts in which it is either very difficult to hallucinate, or just to return link to sources containing the sought information - as a search engine.
> how do I get it to return documents
Ask it to provide the link to said document, in literal natural language terms - "provide a link to document D123" -, and it will provide the link in a concise NL sentence as the direct output, plus related links in the "references" section.
> Ask it to provide the link to said document, in literal natural language terms - "provide a link to document D123" -, and it will provide the link in a concise NL sentence as the direct output
I dislike this style of "searching", but it does appear to work. I asked "provide a link to the HTML version of RFC 793" and got back the following URL: https://datatracker.ietf.org/doc/html/rfc793[3. The URL is invalid (for some reason the reference became part of the URL), but it was easily fixable.
This list is different from my previous search yesterday, so I don't doubt that in your case you got different sources than me. It does look like the sources include mirrors of the RFC hosted by arc-it.net and tech-invite.com. I do not recognize these domains and would much prefer to see an ietf.org domain instead (which is also the first Google result in my case).
Looks like Perplexity includes sponsored results at the bottom of your searches, and my UBO isn't blocking them. They seem to be fairly unobtrusive, though, and they are labeled as such. For now, at least.
> They seem to be fairly unobtrusive, though, and they are labeled as such. For now, at least.
This was one of main selling points of Google Search 20-something years ago. Innocent text-only context-relevant ads on the side of the page. How times have changed...
Readit News
At first Perplexity seemed great, even providing links to the the sources it was drawing its answers from. But after the fifth or so time of not finding the terms it was claiming as real in any of its cited sources, I gave up. And the sources it was citing weren't even particularly good, just the same ones Google was surfacing.
Since everything could just be a hallucination, I was wasting even more time using it than Google. And even when Perplexity isn't hallucinating, I still can't just how trustworthy its sources are without clicking on them, which is another huge problem with searches like that. The context the information is presented in matters as much as the information itself.
I'm finding several of the LLM's that can cite sources, including Perplexity, are more useful to me than Google for search these days. The notable exception is Gemini which has been quite bad in my experience compared to the other options.
The experience really reminds me of when Google original appeared on the scene in the early 2000s. It was the first time on the web that you could easily find anything and the first search engine that didn't feel overrun by spam.
There have been many topics I've wanted to understand better that essentially require a dive into multiple wikipedia pages. Claude makes wandering around that graph unnecessary and really accelerates the learning/exploration process.
It also allows queries that were previously impossible, for example "tell me if this movie has a happy ending but otherwise don't give me any spoilers".
Initially I wasn't sure I would personally ever use LLMs as an alternative to search, but as I've learned to treat them differently than a standard search engine I find for many cases there's no reason to return to the ever decaying mess that is Google in it's current form.
You can try the LLM half of it here without an account:
https://kagi.com/fastgpt
edit: I tried the perplexity link in the article. For my test query, the text it generated was similar, but it didn't cite any websites to back up its claims. That's useless! I thought having the LLM link to sources was already table stakes.
I asked a speculative question to FastGPT@Kagi and it returned a "Dunno" as the three Wikipedia pages it suggests having checked did not contain an obvious reply.
I asked the same to Perplexity and it returned a plausible speculative logical reply with more credible reference links.
There is a strong difference between "looking for documents" and "asking sensible questions".
FastGPT@Kagi could neither reply to the question nor find web pages that could be interesting to extrapolate the answer.
The ability to ask various questions right from the browser location bar without login is convenient and a surprisingly big deal IMO.
[1] https://search.brave.com/
I asked Google the same question and it has the Wikipedia page "List of things named after Alexander Hamilton" as the top organic result, with nothing ranked above it (ads included). It also offered an AI summary that isn't totally, flagrantly incorrect, listing Hamilton Place, Hamilton Hall, and Hamilton Heights in New York.
People who write these articles about search quality are blinded by some unstated ideology.
Also, how do I get it to return documents? I searched for 'RFC 793' with the intend to read the actual RFC. Instead, the chat bot summarized it for me (which is pointless) and doesn't even use the actual RFC as a source?
Since this is a "Google bad" thread, I feel obligated to mention Kagi which I've been using for nearly a year and it has significantly improved my search experience (better search quality, great customizability).
https://kagi.com/fastgpt?query=RFC+793
For me that gives a link to 793, a newer related RFC that wikipedia mentions, and a draft attempted replacement from 2021. (So, three links to standards and one to wikipedia.)
You could mistrust it since it is based on an uncorrected hallucinating technology, but use it either in contexts in which it is either very difficult to hallucinate, or just to return link to sources containing the sought information - as a search engine.
> how do I get it to return documents
Ask it to provide the link to said document, in literal natural language terms - "provide a link to document D123" -, and it will provide the link in a concise NL sentence as the direct output, plus related links in the "references" section.
I dislike this style of "searching", but it does appear to work. I asked "provide a link to the HTML version of RFC 793" and got back the following URL: https://datatracker.ietf.org/doc/html/rfc793[3. The URL is invalid (for some reason the reference became part of the URL), but it was easily fixable.
- https://www.arc-it.net/html/standards/standard1117.html
- https://www.youtube.com/watch?v=_WH1Z8htjMo
- https://www.youtube.com/watch?v=7cG-GDpMA4k
- https://www.tech-invite.com/y05/tinv-ietf-rfc-0793.html
- https://www.maths.tcd.ie/~eoin/index/isi_793rfc81.html
This list is different from my previous search yesterday, so I don't doubt that in your case you got different sources than me. It does look like the sources include mirrors of the RFC hosted by arc-it.net and tech-invite.com. I do not recognize these domains and would much prefer to see an ietf.org domain instead (which is also the first Google result in my case).
0: https://www.theverge.com/2024/12/5/24313222/chatgpt-pardon-b...
This was one of main selling points of Google Search 20-something years ago. Innocent text-only context-relevant ads on the side of the page. How times have changed...