Retrieval-Augmented Generation (RAG)
Retrieval-augmented generation is a technique where a language model first retrieves relevant documents from an external source, then generates its answer grounded in that retrieved text, so the response reflects specific, citable material rather than the model recalling facts from its training weights alone.
RAG is how most modern answer engines work. When you ask a question, the system runs a retrieval step (usually a semantic search over an index of documents), pulls back the passages it judges most relevant, and feeds them to the model as context. The model then writes an answer that draws on those passages, which is why tools like ChatGPT, Perplexity, and Google AI Overviews can name sources and link to them: the source is literally in front of the model at generation time. Without retrieval, a model can only paraphrase whatever happened to be in its training data, which may be stale, generic, or simply absent for a niche product.
The pipeline has two stages, and each can fail independently. Retrieval decides which documents enter the context window; generation decides what the model does with them. A page can be perfectly written and still lose at the retrieval stage because it was never indexed, never chunked into passages a retriever can match, or never corroborated by other sources the system trusts. Conversely, strong retrieval can still produce a weak answer if the underlying passages are thin or contradictory. Optimising for AI search means caring about both stages, not just the prose a human would read.
Consider a Shopify store selling merino base layers. A shopper opens ChatGPT and asks which brand holds up best after repeated washing. The assistant retrieves whatever it can find on that exact behaviour: forum threads, a couple of editorial round-ups, and any review text that has been published in a readable, structured form. If the store keeps its 900 reviews locked inside a JavaScript widget that renders after page load, the retriever often sees an empty shell and pulls nothing. A competitor whose reviews sit in crawlable HTML, marked up with Product and Review schema, gets quoted instead, even with fewer reviews overall.
RAG does not guarantee correctness, and it is worth being honest about that. The model can still misread a passage, blend two sources, or cite a page that does not actually support the claim. Retrieval quality also varies by query, so a page that earns a citation for one phrasing can be invisible for a close paraphrase. For commerce, where buyers increasingly ask AI assistants to compare and recommend products, the recurring gap is that genuine customer reviews exist but are not readable, corroborated, or indexed in a form a retriever will pull. Being retrievable is the precondition for everything else: if your content never enters the context window, the quality of your answer never gets a chance to matter. Closing that gap is the work BeyondReviews focuses on.