What's the Difference Between RAG and an AI That Browses the Web?

In kind, nothing: an AI assistant that browses the web is a RAG system. It runs the same three steps — look something up, put what it found in front of your question, write an answer from it. Calling one “RAG” and the other “web browsing” makes them sound like rival technologies. They aren’t. The only real difference is who chose the material being searched: with a typical RAG setup, somebody at your organisation chose it; with a browsing assistant, a search engine chose it, live, for that one question.

That sounds like a technicality. It isn’t. It changes four things you can feel in the answers.

First, be clear that the recipe is identical

If you’ve read How does RAG work?, the sequence is familiar: retrieve, augment, generate. A browsing assistant does exactly that. Your question becomes a search query, some results come back, the text gets pasted in front of your question, and the model writes.

Swap the library and every step still applies. That’s why it’s more useful to ask “what is this thing searching?” than “is this RAG?” — a question taken apart in Does RAG search the internet?.

The one real difference: who picked the library

With an internal RAG assistant, the collection is a decision. Someone put the current handbook in and left the superseded draft out. That decision was made once, in advance, by a person who could be asked about it.

With a browsing assistant, the collection is decided per question, in a fraction of a second, by ranking. Nobody vetted those pages for your purpose. Nobody vetted them at all.

There’s a second, quieter part of that choice: the assistant isn’t reaching for “the web” as one undifferentiated thing. Search results come in kinds, and they’re fetched separately — Serply’s documentation lists separate endpoints for news, images, maps and academic results alongside the general web one, each returning its own set of fields. So something upstream is deciding which slice your question gets pointed at, and a question routed to general web results will come back with different material than the same question routed to news. You don’t see that decision either.

What that changes: four things

1. Freshness

The browsing assistant wins outright and it isn’t close. Anything published this morning is reachable; a curated collection knows only what’s been added to it, which might be last quarter. If your question has a date in it, or an implied one (“current,” “latest,” “still”), a closed collection is the wrong tool.

2. Whether anyone can vouch for the source

Here the curated collection wins outright, for the same reason it lost the first round. An internal assistant quoting the handbook is quoting a document with an owner, a version and an approval. A browsing assistant quoting the fourth result is quoting a stranger. Both sound equally confident, and confidence is a writing style rather than evidence — a point worth internalising from How do I know if an AI answer is accurate?.

This is the axis people underestimate most. “It found a source” and “the source is any good” are unrelated facts.

3. Whether you get the same answer twice

Ask a browsing assistant the same question tomorrow and the results underneath may have moved: rankings shift, pages get edited, news arrives. The answer can change without anything being wrong. A closed collection changes only when somebody changes it, so a differing answer there is a signal worth chasing rather than background noise.

If repeat answers matter to you — a policy, a number you’ll act on — this alone is a reason to prefer the curated side, or at least to save the answer you got.

4. What a citation proves

Both kinds of system can show sources, and the sources mean different things.

• A document reference (“Travel Policy, section 4”) proves the passage exists in material your organisation stands behind. What’s left to check is whether the assistant read it correctly.
• A web link proves only that this page was retrieved and read. Whether the page is right, current, or selling something is entirely open.

So the verification move differs. For an internal answer, you’re checking the summary against the source. For a web answer, you’re checking the source itself. Same habit, different target.

Which one do you actually want?

As a reader choosing which tool to open, the rule is short:

• Something about the world, recent, public — news, prices, whether a thing still exists, what happened. Use the browsing assistant, and open a link.
• Something about your organisation — policy, process, “what did we agree with this client.” Use the internal assistant. The browsing one physically cannot know, and if it answers anyway it is inventing.
• Something stable and general — how a concept works, what a term means. Either will do; often neither needs to retrieve at all, because it’s already in the model’s training.
• Something consequential either way — verify before acting, regardless of which side answered. Neither kind of retrieval makes an answer true; it makes it checkable.

When one product is both

Increasingly the distinction lives inside a single assistant rather than between two products. The same chat window may search your files for one question and the web for the next, deciding which per question. Which is fine, right up until it does both and merges them.

Two habits cover it. First, notice what the citations look like — outside links or internal document names — because that tells you which library answered. Second, when the stakes justify it, ask outright: “Was any of that from the web?” A system that retrieved can generally say. A system that answered from memory and cited nothing is a third case entirely, and the one to be most careful with — see Why does ChatGPT cite sources sometimes and not others?.

The takeaway

There is no technical rivalry between RAG and web browsing, because browsing is one of the ways RAG gets done. The question worth asking is never “which technology is this?” but “who chose what it read?” A person choosing in advance buys you accountability and repeatability. A search ranking choosing on the spot buys you freshness and reach. Knowing which one is behind a given answer tells you what that answer is worth — and what you still need to check yourself.