Where does LLM pull data from in GPT for Sheets?

[ahtaraut]

Guys, help me out here. I often use GPT for Sheets in my work. When analyzing donor sites for external promotion, I can give it a link and ask it to determine the site's theme. Sometimes the AI makes mistakes, but very often it answers correctly.

This brings me to my question: Where does the AI get this data from if LLM can't follow links (since it's a simple language model)? Do they use their own database that contains all the information? Usually, such databases are updated once a year, which explains why AI makes mistakes in determining some topics (if the site's topic has changed after a change of ownership).

Am I understanding this correctly?

Thanks!

 

[SirLouen]

Nowadays most "LLM" are not providing the raw access to the LLM but just an agent that connects to the actual LLM and also provide some other agents like a search engine agent and a web parser agent. So basically its not like 2 years ago that you got raw data, unless you are using your local LLM, which unless you do all this agent connection, it won't be able to access such.

 

[Wheelgbugai]

I used to wonder the same thing, especially when using GPT to quickly analyze sites. It is true that sometimes it guesses wrong, but most of the time it is reasonable. In my opinion, your speculation about it relying on old or infrequently updated data sounds quite reasonable.

 

[commanday8]

Guys, help me out here. I often use GPT for Sheets in my work. When analyzing donor sites for external promotion, I can give it a link and ask it to determine the site's theme. Sometimes the AI makes mistakes, but very often it answers correctly.

This brings me to my question: Where does the AI get this data from if LLM can't follow links (since it's a simple language model)? Do they use their own database that contains all the information? Usually, such databases are updated once a year, which explains why AI makes mistakes in determining some topics (if the site's topic has changed after a change of ownership).

Am I understanding this correctly?

Thanks!

It’s not actually browsing the site it’s making an educated guess based on patterns in the URL or known data from training. If the site’s changed recently, that won’t be reflected.

 

[ahtaraut]

Yeah, you're pretty much spot on. GPT doesn’t “visit” the link it just tries to guess based on the URL or any text around it. Tools like GPT for Sheets might use cached data or old models, which explains why it's off if a site's changed recently.

That's what I was thinking. And in theory, he gets all the information about the sites I'm asking about from conditional anchor texts or the context as a whole.

 

[ahtaraut]

I used to wonder the same thing, especially when using GPT to quickly analyze sites. It is true that sometimes it guesses wrong, but most of the time it is reasonable. In my opinion, your speculation about it relying on old or infrequently updated data sounds quite reasonable.

And if I understand correctly, such updates occur about once a year. Accordingly, if you make changes to the page, the AI will only learn about it with the next update.

 

[ahtaraut]

Thank you all for your responses. Finally, I would like to ask one more question. If I am not mistaken, the only LLM that currently has direct access to the web is Perplexity. Has anyone heard of this?

To confirm this, just today I ran a test to determine the subject matter of websites in GPT for Sheets, where I compared GPT 5 and Perplexity. Perplexity was almost never wrong, while GPT 5 failed to determine the subject matter several times. The sample was taken from about 100 websites.

In other words, Perplexity may indeed have direct access to the web.

What do you think about this?