3 ms·
There’s genuine loss of value going on if we fall into the trap of language-generated SEO - but it’s more subtle than just ‘the internet will be full of crap’
by jameshart 4y ago
There’s genuine loss of value going on if we fall into the trap of language-generated SEO - but it’s more subtle than just ‘the internet will be full of crap’
The thing is that webpages don’t ‘exist’ in the way most people mentally kind of think of them - as ‘documents on the file system of some computer’. And google is not an index over all those ‘documents’.
How search actually works is Google bots go out to a bunch of servers and ask them ‘hey, what kind of documents can you make?’, and the servers respond with dynamically generated lists of links to those ‘documents’, and google it requests them, and the server makes them and sends them to google, which indexes them so that later when someone searches the internet, google can send them to the same link, and the server can make the document again for them.
In an internet that works like this, using an LLM to generate documents that google indexes is a waste of everyone’s time.
This turns google into an index of the output of various LLM prompts that someone has run that they think someone might later search for.
This wastes everybody’s time.
An LLM outputted summary of every earnings report of every company doesn’t need to be pregenerated - we can make that later, at the time the customer asks the query ‘what was in the earnings report of Foocorp in Q4 2013?’
This breaks the ‘document-centric’ conceit of search; and that might not be a bad thing.
- elecush 4y agoeverything accurate but it is only a "waste of some peoples' time" while others profit with their SEO-friendly garbledy-gook
- advisedwang 4y agoInteresting point, but there's a lot of use cases where chat bots still need some novel data, like your "earnings report of Foocorp in Q4 2013" example. So maybe we end up with [smaller] sites of raw data for bots to digest when someone wants a readable form of it. Frankly that might be better then everything being wrapped up in prose.
- distortionfield 4y ago> Google requests a set of documents that turn out to be LLM generated > Other prompts using Google data make more content for Google to crawl and request > Lather, rinse, repeat AI contamination is already starting. There's no way that content written by an LLM hasn't already been used to train an LLM.
- jameshart 4y agoI’m not talking about LLM feedback loops (that is a problem but a different one. Honestly LLMs slurping up low effort human produced SEO docs is just as bad.) This issue is more about the value of search finding ‘documents’ when ‘documents’ are thin layers of low-value-add LLM cruft over raw factual data. The facts are what you’re searching for; search indexing has always involved trying to figure out the underlying informational content of a document; LLM prettification is just adding obfuscation over that data that search indexers then have to try to remove. There has to be a more efficient way for a website to say ‘we have all the earnings data for every US public company’ without them having to dress that up in LLM prettification that google’s language model can index.
- harshalizee 4y agoIt's even worse than that. Sites like MarketWatch auto generates pages for each ticker at the end of trading day to tell you if a stock went up or down and in comparison to broader market. Just take a look at a random stock in apple stocks, the news section is just filled with this garbage.
- bathtub365 4y agoWe’ve reached the point where trees are only falling in the forest because Google is around to hear them.
- mistrial9 4y agoconversely, that which is digital and on the net, is indexable, while so much of intelligent information, factual content and also human communications.. is not digital and not on the net. in the early oughts, I said to anyone that would listen that the powerful popularity searches shown by google will be sort of self-referential given a simple definition of what can be indexed and what cannot..