Re-ranking is the step in which an AI system no longer assesses pages but passages. After the initial retrieval, there is a candidate pool of pages. In re-ranking, the individual text sections of these pages are checked against the question, and only the most relevant ones advance a round. Whatever stays below the relevance threshold drops out, even if the page itself ranks well.
How does re-ranking work?
The system splits the candidate pages into sections (chunking), calculates the semantic proximity of each section to the question (cosine similarity) and adds further signals: Does the section answer the question directly? Is it specific or general? Is it understandable without context? Then a threshold is applied, the “cut-off”. Sections above it become grounding snippets, sections below it are discarded.
What makes a passage relevant can be summed up in five properties: it answers the query directly. It has a strong topical and semantic match. It is clear and easy to read, for machines too. It is highly specific and focused. And it is high-quality and self-contained.
Why does re-ranking matter for AI visibility?
Re-ranking is the reason why good rankings do not guarantee a citation. “Being retrieved doesn’t guarantee being cited”: a page can be in the pool and still not get a single passage over the threshold, because its paragraphs are too general, too promotional or only understandable in context. Conversely, a page in position six can be cited with a single precise paragraph.
What does this mean for your website?
Optimise at paragraph level. Every section that answers a customer question should begin with the term being asked about, deliver the answer in the first sentences and be understandable without the rest of the page (semantic chunking). Cut hedging and introductions, add figures, names and evidence (information gain). The page has to rank, the passage has to convince.