How the LLM merges results (reranking)
After retrieving dozens of passages, an AI system has to decide which ones to keep. This sorting step, called reranking, largely decides who gets cited.
After the initial retrieval, an AI system typically ends up with far more candidate passages than it can use in an answer. A sorting mechanism, called reranking, evaluates and prioritizes these passages to keep only the most relevant ones. This discreet but decisive step largely determines who gets cited and who doesn't. Unlike an SEO ranking algorithm, which stays broadly stable over time for a given query, the outcome of reranking can vary from one question to another even on a related topic, which makes it even trickier to observe.
Why sorting is needed after retrieval
The first document-retrieval step, whether through a classic search engine or a vector database, generally produces a fairly broad list of potentially relevant results, without much fine-grained precision. This first list is deliberately wide so as not to miss useful content, but it also contains noise: passages barely related to the topic, redundant with one another, or of uneven quality. Reranking exists to refine this list, evaluating each passage more precisely against the specific question the user asked. This step remains largely invisible to a website: unlike a position in Google's results, there's no "page 2" of reranking you can check to understand why a passage got dropped. That's what makes the topic hard to diagnose without a dedicated measurement tool.
The general principle behind reranking
In a typical RAG architecture, a specialized model, often smaller and faster than a full LLM, re-evaluates each candidate passage by comparing it directly to the question asked, rather than through simple keyword matching. This model assigns a relevance score to each passage, and only the top-scoring passages are passed on to the language model for drafting the final answer. Some systems also combine several sources of results, for instance keyword search and semantic similarity search, then merge the two rankings into one; researchers often describe this kind of merging as "reciprocal rank fusion," a method that favors passages ranked well across several lists at once rather than those that dominate a single list. This choice illustrates a fairly general principle of reranking: a passage's robustness, meaning its ability to stay relevant across several different evaluation methods, often matters more than a perfect but isolated match on a single criterion.
What makes a passage more likely to rank well
Without claiming to know the exact, weighted criteria of each proprietary system, some general characteristics can be observed that tend to favor a passage during reranking: a clear, direct match with the intent of the question, phrasing that answers head-on rather than circling the topic, and enough standalone clarity to be understood without having read the rest of the page. Conversely, a passage buried in generic content, one that requires outside context to make sense, or one that vaguely restates a question without precisely answering it, statistically has less chance of being kept at this stage. Take an example: for the question "what's the legal withdrawal period for an online purchase," a passage that opens directly with "the legal withdrawal period is 14 days from receipt of the product" has a better chance of ranking well than a paragraph that starts with "the question of deadlines is complex and depends on several factors" before arriving, a few sentences later, at the same information. The gap between the two isn't about accuracy, since both can be equally correct, but about how quickly the useful information surfaces for a system that has no patience for a slow build-up.
The link with chunking and content structure
Reranking operates on passages, not whole pages. That's why how your content is split up beforehand, a process called chunking, has a direct influence on what makes it to this sorting step. An article where each section forms a self-contained unit of meaning gives the system more "good candidates" to evaluate, rather than one long block of text where the useful information gets diluted. This splitting mechanism is explained in detail in the lesson Chunking: how the LLM splits up your content. In other words, quality content that stays poorly segmented can end up penalized not on substance, but on form: reranking simply doesn't have the right units to evaluate.
What this means for how you write your content
In practice, this encourages writing that gets straight to the point: answering a question within the first sentences of a section, clearly structuring each subsection with an explicit subheading, avoiding paragraphs that require having read the whole article to make sense. These structuring principles are detailed in the lesson Structuring an article for GEO, which offers a template designed to maximize the chances that as many of your passages as possible survive the reranking step.
Keep in mind that reranking is an invisible but decisive filter: your content can be perfectly retrieved at the first step and still never appear in the final answer, simply because it wasn't judged relevant, clear, or self-contained enough at sorting time. This observation makes the case for regularly reviewing your own content, section by section, honestly asking whether each passage, taken in isolation, would give a clear answer to someone who had read nothing else. None of this requires guessing blindly: publishing content, checking whether it gets cited for related questions over time, and adjusting its structure based on what does and doesn't get picked up is a workable, iterative way to learn what a given reranking system tends to favor.