Extraction

Modern engines don't read pages the way people do. A question gets expanded into multiple sub-queries (query fan-out), candidate pages are retrieved, and then individual passages are scored for how directly they answer. The unit of competition is the paragraph. This is why a page can rank well in classic search and still contribute nothing to AI answers: it wins on page-level signals while containing no individually liftable passage.

Extractable writing has a recognizable shape: a heading phrased as the question people actually ask, followed immediately by a self-contained 40–60 word answer — no 'it depends' throat-clearing, no dangling references to earlier paragraphs — then supporting detail beneath. Self-containment matters because the passage will be read out of context; a paragraph that begins 'As mentioned above' dies at retrieval time.

Example: a moving company's pricing page opens with three paragraphs about their commitment to service. The engine skips them. Restructured to open with what a local three-bedroom move typically costs and the four factors that change it, that passage becomes the extractable answer for every cost prompt in the market. The Princeton GEO research showed content carrying statistics, quotations and source citations gains 22–41% visibility in generative answers (Aggarwal et al.) — evidence-dense passages are extractable passages. Making pages pass this test is the core of AEO; its counterpart on the model's side is synthesis.