Retrieval is the step where an answer system fetches live or indexed pages and passages to ground a response, rather than relying only on parameters learned in training. It is useful because findability at answer time still depends on crawlable, clear, citable URLs — even when the interface looks like a chat box.
Whilst retrieval sounds automatic, in practice buried answers, blocked content, and conflicting facts mean the right page never enters the context. For example, the proof lives in a PDF image or a campaign landing that rotated off URL, so the model cites an older third party. We often recommend stable answer pages, clear claims near the top, and internal links between entities and proof so retrieval has something solid to pull.
Generative answers mix memory and retrieval. When retrieval runs, HTML that states claims plainly outperforms locked PDFs, image-only text, and vague homepage copy. Technical SEO basics (indexation, canonicals, performance) still affect whether a page can be fetched and trusted. Retrieval is also the bridge to RAG-style architectures inside enterprises — same idea, different boundary. For GEO, retrieval is why “we ranked once” is not enough: the passage must be selectable as a source when the prompt is asked.