What are the five stages of the AI search pipeline?

You own an input to stages three, four and five, and nothing at all at stages one and two. Interpretation runs before any document is touched, on the user’s words, the conversation history and the engine’s own priors alone. Fan-out is generated from that intent by the engine’s own model of the subtopics. Neither stage reads your site, so any tactic sold as improving them is either mislabelled or empty.

At stage three your inputs are index membership, fetchability and relevance. Google states the eligibility rule for its own surfaces directly: a page “must be indexed and eligible to be shown in Google Search with a snippet, fulfilling the Search technical requirements”, and adds that no special files or schema.org markup are needed to appear.¹ Its optimization guide names the mechanism behind that rule, describing retrieval-augmented generation as improving answers “by relying on our core Search ranking systems to retrieve relevant, up-to-date web pages from our Search index”.⁶ At stage four your input is passage structure, meaning whether the answer to a specific sub-question sits in one liftable chunk. At stage five your input is wording.

There is also a large part of the pipeline you own no input to, and it is not a stage: it is everybody else’s pages. Stages three and four run over an index built from the whole web, so on most commercial questions most retrieved candidates are third-party reviews, roundups and discussion threads. One 2026 vendor study of four brands, covering 47,097 citations across three engines, put the share of cited material a brand owns at roughly 2%; four brands is illustrative rather than a constant, but the order of magnitude is the point.⁹

A stage-five tactic loses at stage three whenever a rewrite that makes a passage more quotable simultaneously makes the page less relevant to the query that would have retrieved it, because the pipeline multiplies the two effects rather than adding them. A KDD 2026 benchmark built an end-to-end environment of 171,003 documents and 2,700 queries with retrieval and reranking reinstated, and tested ten body-only strategies. Averaged across them, top-20 presence fell from 0.58 to 0.53, post-rerank top-10 presence from 1.00 to 0.84, and the citation rate from 0.50 to 0.47; the worst single strategy took top-20 presence to 0.37 and the citation rate to 0.39.³ The survey’s reading is exact: a rewrite “may therefore perform well once injected while making the document less retrievable or less competitive upstream.”²

Two other benchmarks land in the same place from different directions. A NeurIPS 2025 datasets-and-benchmarks paper ran “a total of ten methods” over a corpus it describes as “more than 1.9k queries and 16k documents”, and concluded that these methods are “not only largely ineffective but can actually have the opposite expected effect by decreasing document ranking”: “Out of 54 cases, we uncover only three where the ranking improvements are statistically significant”, and none at all for question answering.⁴ On position it stays qualitative and prints no multiple: making a document the first one in the context window “leads to far greater citation ranking gains in the LLM response” than any of the content methods it tested.⁴ The 2024 KDD paper that founded the field reported much larger gains, but on a custom metric inside a purpose-built two-step engine, for a page already sitting in the five-document context; the survey calls the popular generalisation of that number “a relative maximum on one metric under a specific configuration”.⁵² Both results are honest; only one of them ever ran a retrieval step. Where GEO advice goes wrong takes both benchmarks apart in full.

Measure the stages separately, because a single blended visibility number cannot tell you which one failed, and the fix for each is different work by a different person. The survey states the sequence plainly: “produce a relevant, comprehensive, verifiable, clearly structured, and technically retrievable page; then measure retrieval, citation, and fidelity separately.”²

The benchmark that restored the full pipeline reports three numbers per strategy instead of one: presence in the top 20 after retrieval, presence in the top 10 after reranking, and the citation rate in the finished answer. The three move independently. Adding structural information to the same documents raised average top-20 presence from 0.58 to 0.71 while the citation rate barely moved, going from 0.50 to 0.52.³ A single blended score would have shown a small positive and hidden both halves of that result.

The founder version needs no instrumentation and takes an afternoon. Stage three: fetch the page with a plain request, confirm the answer text is in the HTML, then confirm it is indexed and snippet-eligible.¹ Stages two and three: search the sub-questions your topic fans out into and see whether anything of yours ranks. Stage four: read the passages engines quoted from whoever they cited instead. Stage five: compare that wording with yours. Four checks, four different failures, and only the last is a writing problem.