How do selection and reranking decide which passages an AI answer uses?

Selection and reranking turns several candidate lists into one short, ordered list of evidence that fits inside a fixed context budget. Four operations happen in sequence. The candidate sets returned for each fan-out sub-query are merged; near-duplicates and syndicated copies are collapsed so the same claim does not occupy three slots; what remains is reordered by a model scoring how well each passage answers the specific question rather than how authoritative its domain is; and the tail is cut until the evidence fits. Google describes the same shape in first-party language: grounding retrieves pages “by relying on our core Search ranking systems”, after which “our systems then review the specific information from those retrieved pages to generate a more reliable and helpful response”.⁹ That review is stage four.

That a model reads the passages, rather than a formula scoring a link graph, is not a guess about closed systems. It is the published technique: an EMNLP 2023 study found properly instructed language models “can deliver competitive, even superior results to state-of-the-art supervised methods on popular IR benchmarks” as re-ranking agents.¹⁰ A reranker is a reader, and it reads chunks.

The cut is the part people underestimate. Everything upstream produced candidates in the hundreds or thousands; what reaches the writing stage is a handful. In a 602-prompt academic sample published April 2026 the mean number of citations on a finished answer was 6.88 for ChatGPT, 12.06 for Google’s AI surfaces and 16.35 for Perplexity.³ Almost all of the pool was discarded here, and nothing discarded can be recovered by better prose further down. The how AI search works stage places this stage inside the full pipeline.

A buried answer loses because the reranker scores the chunk, and the eight paragraphs surrounding your answer are not evidence for it, they are noise around it. A 3,000-word page whose answer sits in paragraph nine presents one relevant passage embedded in a lot of material about something adjacent. A shorter page whose second sentence answers the question presents a passage that is almost entirely on-target. Page-level authority does not rescue the first case, because authority is not what is scored here.

Position within the page appears to matter as well, though the evidence is vendor-published and self-declared correlational. A June 2026 study of 10,000 AI Overview responses and 100,000 citation placements, matched to source locations via scroll-to-text fragments, found 38% of citations were pulled from the first 100 words of a page, up from 20% in a smaller comparison set a year earlier; its authors state plainly that the data is correlational and that they “can show where AI models cite from, not definitively state why”.² The practical reading is not “put everything at the top”. It is that preamble, brand throat-clearing and “in this article we will cover” occupy the most valuable real estate on the page.

Position in the model’s context, once a passage is selected, matters too, and the support is stronger than for anything about page shape. The 2026 critical survey grades position in context Strong, conditional on the document already having been retrieved, and query-document relevance Strong in controlled settings.⁴ The NeurIPS 2025 C-SEO Bench puts numbers on it: on GPT-4o-mini in retail, the best traditional intervention, improving the source’s position in the LLM context, moved the target 2.77 rank places on average, against 0.36 for the strongest of the ten content-rewriting methods it tested. Its own wording carries no multiple: making the target document “the first one in the LLM context window leads to far greater citation ranking gains in the LLM response than any C-SEO method”.⁵ Why has its own literature: performance “is often highest when relevant information occurs at the beginning or end of the input context, and significantly degrades when models must access relevant information in the middle of long contexts”.⁸ You do not set that position. You influence it by being the passage that most exactly answers the sub-question.

Only negatively, and only on the surfaces that document a control. Google’s guidance names four: “to limit the information shown from your pages in Search, use nosnippet, data-nosnippet, max-snippet, or noindex controls”, and its AI features run off the same snippet eligibility.⁶ That gives you a way to exclude a specific block of text from being quoted, or to cap how much of a page can be shown, both of which reduce your exposure rather than steer it.

There is no positive equivalent. No engine offers a way to mark a paragraph as the one you would like quoted, and no markup, file or directive nominates one. The only positive influence is compositional: make the passage that answers the sub-question the clearest and most exactly on-target thing on the page, and give it a heading that states the question. That is not a trick, which is why it keeps working.