Does RAG Search the Internet?

Usually not. A RAG system searches whichever collection of documents somebody deliberately pointed it at — a company’s handbooks, a folder of PDFs, a product manual, a set of support tickets. The open internet is not included by default, and in most business setups it is excluded on purpose. But nothing in the idea of RAG requires a private collection, and some products do point the retrieval step at the live web. So the honest answer is: it searches what it was given, and what it was given is a choice somebody made.

That choice is invisible from the outside, which is exactly why this question keeps coming up. Here’s how to think about it — and how to find out which kind you’re using.

The acronym names the step, not the source

RAG stands for Retrieval-Augmented Generation: retrieve some relevant material, add it to your question, and let a language model write the answer from it. If that’s new, What is RAG in simple terms? walks through the whole loop.

Notice what the word retrieval doesn’t say. It doesn’t say retrieve from where. It only says that before answering, the system goes and fetches something. The “where” is a separate decision, made by whoever built the thing, and it can be almost anything: a shared drive, a wiki, a database of past orders, a single uploaded PDF, or yes, a search engine’s index of the web.

This is the piece beginners are rarely told, and it explains most of the confusion. Two products can both be “doing RAG” and have nothing in common in terms of what they actually know.

So what is it usually searching?

In the great majority of real deployments, the answer is a corpus — a fixed, deliberately assembled collection of documents. Somebody sat down and decided: these files, not those. The HR handbook and the benefits summary, but not the draft that was never approved. The current product documentation, but not the 2019 version.

That collection then gets prepared for searching — split into passages, converted into a searchable form, and stored somewhere the system can query quickly. None of that preparation invents new material; it only makes the chosen material findable.

The important word is assembled. A RAG system’s knowledge is not the sum of human writing. It is one shelf, and somebody chose the books on it.

Why the collection is usually closed on purpose

If the web is so much bigger, why would anyone deliberately shut it out? Four reasons, and all four are about control rather than about size.

• You know where an answer came from. When the assistant says “the refund window is 30 days,” you can point at the exact internal document that says so. A web result gives you a page, and pages disagree.
• The material is not public in the first place. Your handbook, your contracts, your ticket history — none of that is on the internet, so a web search could never find it. The whole value of most internal assistants is exactly this.
• The web is optimised, not curated. Plenty of pages exist because somebody wanted them to rank, not because somebody wanted them to be right. Once those pages become retrievable, they become quotable.
• It stays still. A closed collection changes when somebody changes it. The web changes constantly, which makes yesterday’s correct answer hard to reproduce today.

Note that none of these say the closed collection is better. They say it is known. That’s the trade being made.

But it can be the web — and sometimes is

Some products absolutely do point retrieval at the live internet. Consumer assistants that answer about last night’s news are doing precisely this, and it works: the model’s own training stopped at some point in the past, so a live lookup is the only way to know anything newer. How do AI chatbots search the web? traces that path step by step.

Worth knowing, because it demystifies things: when a system retrieves from the web, it usually isn’t crawling the internet itself. It sends a query to a search service and gets results back as data — a list of titles, links and short descriptions it can hand to the model. This is an ordinary, purchasable service rather than anything exotic; Serply’s published description of what its search endpoint returns lists organic results with their titles, links and descriptions as the fields that come back from a single request.

That detail matters to a beginner for one reason only: it tells you that a “web-connected” assistant’s knowledge is really a search engine’s ranked results, not the internet. Whatever that ranking put in the top handful is what the assistant will read. Nothing else on the web exists for the purposes of that answer.

How to tell which one you’re using

You don’t need documentation. Three questions, about a minute:

1. Ask about something that happened yesterday. If it answers correctly with a link, something is retrieving live. If it says it can’t know, or confidently gives you something stale, it isn’t.
2. Ask about something only in your own material. “What does our travel policy say about rail fares?” A web-only system cannot answer this, no matter how fluent it sounds. If it tries, it’s guessing.
3. Look at what the citations point to. Links to outside websites mean the web. Document names, page numbers or file references mean a private collection. No citations at all usually means it answered from memory and retrieved nothing — the case covered in Does ChatGPT use RAG?.

That third test is the most useful habit to build, because it also tells you when a system that usually retrieves quietly didn’t bother this time.

Most products have both

The tidy either/or breaks down in practice. Plenty of assistants can search your files and search the web, and decide per question which to do — sometimes both, then blend the results into one answer.

Which is why the two aren’t really rivals, and why the useful comparison isn’t technical at all — it’s about who chose the material, laid out in What’s the difference between RAG and an AI that browses the web?.

Blending is convenient and occasionally hazardous, because a blended answer rarely labels which sentence came from where. A paragraph can open with your internal policy and close with a claim lifted from a vendor’s marketing page, in the same calm voice. If an answer matters and it might have mixed sources, ask directly: “Which of that came from our documents?” A system that retrieved can usually tell you.

The takeaway

RAG does not mean “searches the internet.” It means the system looks something up before answering, and the thing it looks in was chosen by whoever built it — most often a private, curated collection, occasionally the live web, sometimes both. Since that choice decides everything the assistant can possibly know, it’s the first thing worth establishing about any tool you’re asked to trust. Ask what it searches. If nobody can tell you, that’s your answer about how much weight to put on it.