A page can rank well on Google, get retrieved by ChatGPT, match the topic perfectly — and still never be cited. The reason sits in a stage most GEO advice doesn't mention: between retrieval and answer-writing, AI search runs a reranker, a model whose only job is to score each candidate passage against the query and decide which ones deserve to feed the answer. Understanding this stage explains the most frustrating pattern in AI visibility: being "right there" and losing anyway.
This is the GEOly AI practitioner's guide to the full source-selection pipeline — the four stages, where pages actually lose, and a diagnosis-first workflow that stops teams from rewriting paragraphs to fix problems that live in a different stage.
Key takeaways
- AI search selects sources in four independent stages — query planning, retrieval, reranking, generation — and a page can pass three and fail the fourth. One "visibility score" hides which gate killed you.
- Rerankers score passages, not pages. A topically relevant page loses to a competitor whose paragraph is shaped like the answer — direct claim first, entities named, self-contained.
- The "listicle effect" isn't bias: lists win because one passage answers many fan-out sub-queries at once. The fix is answer-shape alignment, not converting your site to listicles.
- Diagnosis before editing: if the page was never retrieved, no rewrite for the reranker can save it. Separate retrieval exposure, passage strength and citation rate — three metrics, not one.
- Don't "chunk-optimize." Write self-contained sections with explicit subject names; that survives any chunking scheme an engine uses.
The four-stage pipeline — and where you lose
Stage 1 — query planning. The engine rewrites the user's question into multiple hidden sub-queries (fan-out): a "best CRM for small teams" prompt spawns searches on pricing, integrations, alternatives, complaints. Lose here and nothing downstream matters — the topic facets simply never point at you.
Stage 2 — retrieval. Keyword search (BM25-family) and vector search each pull candidates, typically fused into one pool. This is where classic SEO still rules: crawlability, indexation, rankings. Google states openly that its AI surfaces retrieve from the same index with the same systems. Lose here through blocked crawlers, JS-only rendering, or simply not ranking.
Stage 3 — reranking. A cross-encoder reads each candidate passage the query and scores it. This stage is blunt about format: in published experiments, a descriptive product paragraph scored near zero for "what are the best AEO tools" while a passage that directly named a shortlist scored 99%+ across almost every model family tested. Same topic, same facts available — different answer shape, opposite outcome.
