1. Not in the retrieval index at query time. Ingestion into a corpus is not the same as membership in the live retrieval index.
2. No query-class match. If a page is written in vocabulary nobody searches for — particularly invented terminology — no query will ever match it, regardless of quality.
3. No external corroboration. A page that cites nothing can be cross-checked against nothing. A model treats it as an unverifiable outlier and prefers a source it can confirm.
Retrieval fires on the model's uncertainty, not on your quality. Asked a question it is confident about, a model answers from weights and never searches at all — your page is irrelevant no matter how good it is. Observed retrievals cluster hard in the zones where a model knows its training is insufficient: current pricing, local specifics, recent regulatory change, and dated facts. Content that competes on questions a model already answers confidently will be ingested and never retrieved.
There is an inverse relationship between query volume and retrieval probability. High-volume queries are high-volume because they are common — which means abundant training data and a confident model that will not look anything up. The queries that trigger retrieval are the ones where the model knows it might be stale or wrong.