AI Search Is RAG on the Open Web

If you have heard of RAG, the pipeline from the last lesson may look familiar. It should. AI search is a form of retrieval-augmented generation: find relevant text first, then have a language model write an answer from it.

This lesson connects the two ideas. It also explains a part that is easy to miss: the model still brings its own memory to every answer, and the two can mix.

What You'll Learn

• What RAG means in one sentence
• How AI search is the same as RAG, and how it is different
• When an AI tool searches and when it answers from memory
• How the model's memory can leak into a cited answer
• What this means for how much you trust an answer

RAG in one sentence

A language model learns from a huge amount of text during training. After that, its knowledge is frozen. It does not know about anything that happened after its training ended, and it may remember details wrong.

RAG fixes this by giving the model fresh, relevant text at the moment you ask. The model is told to answer from that text. Retrieval brings in the facts. Generation writes them up.

The Embeddings Explained course explains how the retrieve step finds passages by meaning. If you want to see a small RAG system from the inside, Local RAG for Beginners walks through one on your own computer. This lesson does not repeat either. It focuses on what changes when the documents are the whole web.

Same idea, different library

The steps match. What differs is the collection being searched.

That last row matters most. In a document tool, the sources are usually trusted, and the main problem is finding the right part. On the web, the sources themselves can be wrong, old, or written to mislead. The AI has to judge quality, not only relevance, and it does not always do that well.

When does the AI search at all?

In search-first tools, every question triggers a search. In chat assistants, it depends. The tool, or the model inside it, often decides whether a question needs the web.

A question like "explain photosynthesis" may be answered from memory, because the facts do not change. A question like "what is the weather in Lisbon today" needs a search. Many questions fall in between.

Here is how to tell what happened:

• Sources are shown. The tool searched. You can open the links.
• No sources, no search indicator. The answer most likely came from the model's memory alone. Treat it like any chatbot answer.
• You are not sure. Look for the tool's own signals: a "Searching the web" step, a sources panel, or links under the answer. Do not ask the bot whether it searched. Models often make up that answer, including page lists.

Many tools also let you turn search on yourself. For anything recent, local, or where details matter, turn it on.

Memory leaks into cited answers

Here is the part most people miss. Even when the AI searches, the model still knows everything it learned in training. When it writes the answer, it uses the retrieved passages and its own memory together. There is no wall between them.

Most of the time this helps. Memory lets the model explain a term, add background, or write smooth sentences. But it can also cause problems:

• An old fact from memory beats a new fact from the page. The page says a price went up. The model "remembers" the old price and writes that instead.
• A detail from memory gets a citation it did not earn. The model adds a fact it remembers, and a nearby citation makes it look like the page said it.
• Memory fills a gap. The pages did not answer one part of the question, so the model fills it in from memory without saying so.

This is why a cited answer is not the same as a checked answer. The citations show what was retrieved. They do not mark which sentences came from memory.

Why companies build it this way

You might ask: why not force the model to use only the pages? Some tools try, with strict instructions. But a model that can only copy from pages writes stiff, broken answers. It cannot explain, connect, or rephrase. So builders allow some freedom, and accept some risk.

Builders also tune how much the model should trust the pages over its memory. When the pages and memory disagree, a well-built system should follow the pages and say there is a conflict. Not every tool does this every time.

What this means for you

Knowing that AI search is RAG gives you three useful rules:

1. The answer is limited by what was retrieved. If the right page was not found, the answer cannot use it.
2. The answer is limited by source quality. Retrieval does not make a weak page strong.
3. Some sentences may come from memory, not the sources. Specific facts without a clear match on the cited page need checking.

The next lesson looks at the exact ways these limits turn into wrong answers.

Key Takeaways

• RAG means retrieving relevant text first, then having a model write an answer from it. AI search is RAG over the open web.
• In document tools the sources are usually trusted. On the web, sources vary in quality, and some are wrong or old.
• Chat assistants decide whether to search. No sources shown usually means the answer came from memory.
• The model mixes retrieved text with its own memory. Memory can add old facts or fill gaps, and still sit next to a citation.
• A citation shows what was retrieved. It does not mark which sentences came from memory.