How GPTree searches the web

A model only knows what it was trained on. Ask about this morning's news, a share price, a score, or how a particular page reads right now, and training data cannot help. GPTree closes that gap by letting the assistant go and look.

Two ways a search happens

The assistant decides. When your question needs something current or external, the assistant runs a web search on its own, mid answer, then writes the answer from what came back. You do not have to ask for it and there is nothing to approve: searching is a read-only action, so it runs without interrupting you for permission.

You decide. Click Web Search before you send a message and that message is searched for, whatever it says. This route fetches the results before the model starts, so it is a little quicker. Reach for it when you already know the answer turns on something current. The button is not shown in the onboarding Guide, which searches on its own whenever a question needs it.

The two do not double up. If the button already fetched results for a message, the assistant starts from those and only searches again when part of your question is still not covered.

How much it searches

The assistant sizes the effort to the question. A single fact gets one search. A question with several parts, or one that compares things, gets a search for each part, so asking about two products does not leave one of them unanswered. Open research gets a few searches, and then the assistant answers and tells you what it could not cover.

There is a cap on how much page text one answer takes in. Once it is reached, the assistant finishes the remaining parts from shorter snippets and says which parts are thinly sourced, instead of searching on and on. The cap applies to a single answer, not to you: every new message starts fresh, so you can always ask it to dig further.

What comes back

A search returns ranked results with the text extracted from the pages, not just a list of links. Every result comes back with its title, link and a short summary, plus an excerpt of the page around the names and numbers in your question. The assistant answers from that material and names the sources it used, so you can check the claim against the page rather than taking the answer on trust.

When a result looks like it holds the answer but the excerpt stops short, the assistant can open that page and read the part that matters, for example the list of files for an install.

Names and versions you give it

If you name a specific product, model or version, the assistant searches for that exact name and keeps it as you wrote it. A version newer than the model's training is treated as new, not as a mistake. If the search does not confirm the name, it tells you the search did not confirm it, rather than telling you it does not exist.

What it will not do

It will not describe a search it did not run. Before this existed, a question about current events left the model with two bad options: narrate a fetch that never happened, or claim it had no internet access at all. Both were misleading in different directions. The assistant is now instructed to call the tool or say plainly that it cannot, and if search is unavailable or returns nothing it tells you that instead of guessing.

It only opens links you shared in the conversation, or links that came back from an earlier search. It will not open a link it made up or one it found written inside a page, because text on a web page or in an email should never be able to send your conversation somewhere else. If you want a page read, paste the link.

It is also not a live feed. A search happens when a message needs one. Nothing runs in the background, and an answer reflects what the web said at the moment you asked.

Availability

Web search is on every plan and needs no configuration from you. Searches count toward your usage the same way the rest of a turn does, because fetching and reading pages is work the model does on your behalf.