RAG

RAG means retrieval-augmented generation. Rather than relying only on what it learned in training, the model retrieves documents at answer time and writes from them.

Why does RAG matter to publishers?

Because it makes your live pages part of the answer. Under pure training, your influence is frozen at the moment the model was trained. Under RAG, a page you publish today can shape an answer today, which is precisely why query-time crawls exist.

What makes a page usable by a RAG system?

Retrievability and quotability. The page must be fetchable, and its passages must make sense in isolation, because the model typically retrieves a fragment rather than the whole document.

Does RAG mean my page is quoted verbatim?

Sometimes. More often the model paraphrases and cites. Either way the retrieved passage is the raw material, which is why self-contained writing matters more than clever page-level structure.

What are the stages of a RAG pipeline?

Four, and a publisher is visible in three of them. The corpus is built by crawling; the question is turned into a query; passages are retrieved and ranked against that query; and the model composes an answer from the passages it was handed. A page absent from the first stage cannot appear in any of the others, which is why blocking an index crawler removes you from answers even though the crawler itself never writes a word.

Why do answers cite sources at all?

Because the passages are right there. A model working purely from training weights has no document to point at, so a citation would be a guess. A retrieval system has the URL the passage came from, so attribution is close to free — which is also why the citation can be correct while the sentence around it is not. The engine is vouching for where it read something, not for the reasoning it built on top.

Does RAG replace training data?

No, and treating the two as interchangeable is the most expensive mistake in this subject. Training fixes what a model knows in general; retrieval supplies what it needs for one specific question. A publisher’s relationship to each is different: training is a one-way transfer you cannot audit, while retrieval is a live dependency on pages you can change this afternoon.

What can go wrong between retrieval and the answer?

Plenty. The retrieved passage can be correct and stale, correct and about a different product with a similar name, or correct and then paraphrased into something the source never said. Publishers who watch answers about themselves usually find that the failure is not misattribution but compression: the nuance in the third sentence of a passage rarely survives into a two-line summary.