Website search retrieves and ranks documents for a visitor to choose from. RAG retrieves documents and then has a language model write an answer from them. RAG is not a replacement for search — it is a layer on top of it, and its quality is capped by the retrieval underneath.
| Website search | RAG | |
|---|---|---|
| Output | Ranked results | Written answer |
| Best for | Browsing, comparing, choosing | Factual and support questions |
| Latency | Milliseconds | Seconds |
| Cost per query | Negligible | Model inference |
| Failure mode | Wrong results, visibly | Confident wrong answer |
Generation cannot compensate for retrieval. If the wrong documents are fetched, the model produces a fluent, well-structured, incorrect answer — which is more damaging than an empty result page, because nothing signals to the visitor that it is wrong.
Investing in retrieval quality is therefore the highest-leverage work in any RAG deployment, and it is the part most often skipped in favour of prompt tuning.
Use generation for questions with one correct answer — policies, sizing, shipping, order state. Use results for anything involving choice. Route by intent rather than committing the whole surface to either.