The retrieval pipeline is the whole chain of steps that runs between a user asking a question and being shown an answer with links. Understanding this chain is essential for optimisation, because each step is a separate place where content can drop out – and most content drops out before it even reaches the model. The first step is query processing. The system does not take the question literally but breaks it down into sub-queries, fills in missing context from the previous conversation, and estimates the intent. A question about the suitable solution for a small company thus splits into sub-queries about prices, about features, about alternatives, and about user experiences. This is already where the first opportunity arises: the more sub-queries your content can cover, the more entry points you have into the next steps. The second step is coarse retrieval of candidates. For each sub-query, a set of documents is pulled from the index, usually through a combination of keyword search and semantic-similarity search. What matters here is whether your content is even in the index, whether it is accessible to bots, and whether the key information loads without a script running. Content hidden behind an interactive element does not exist at this step. The third step is reranking. From the dozens or hundreds of candidates, a more precise model selects the passages that best answer the specific sub-query. What decides this is not the authority of the whole website but the match between a specific passage and a specific question. This is precisely why small websites appear in answers alongside large brands. The fourth step is context assembly. The selected passages are inserted into the model's input, and their number is limited – on the order of single digits to tens. Everything else does not make it into the answer. At the same time, the model works more reliably with information at the beginning and end of the inserted text, so a claim buried in the middle of a long paragraph has less chance of getting through. The fifth step is answer generation. The model formulates the text so that it draws on the inserted passages and attaches sources to its claims. It favours passages that confirm a claim unambiguously and without any need for inference – that is, specific figures, one-sentence definitions, clear conditions. Vague marketing phrasing cannot be anchored, and so it drops out of the answer even if the page ranks well in classic search. The sixth step is verification and display. The system checks the consistency of claims with the sources, adds links, and displays the answer. It is also at this stage that the order of the brands listed is decided, which is not random and is influenced by agreement with what other sources state. The practical conclusion from the whole chain is simple. Optimisation for generative search is not a single activity but the removal of obstacles at six different points: technical accessibility for bots, the existence of passages for individual sub-questions, phrasing the answer right in the first sentence, the concreteness of the data, consistency with other sources, and an unambiguous identity for the company. Neglecting any one of them means that work on the others is wasted. When planning work, it is therefore worth proceeding in the order in which the chain runs. First, technical accessibility is verified, that is, whether bots can reach the content at all and whether the text loads without a script. Next, the missing answers to sub-queries that emerge from question mapping are filled in. After that, the phrasing is adjusted so that the first sentence of every section answers directly and contains a specific piece of data. Finally, information about the company is aligned across external sources so that no contradictions arise during verification. The reverse order – that is, writing new content on websites that bots cannot reach – is the most common reason why investment in content brings no result in generative search at all. It is likewise worth tracking at which step of the chain you are losing the most – this can be estimated from whether you are completely absent from answers, mentioned without a link, or cited only on peripheral questions.
See also: Query fan-out, Reranking, Top-k retrieval.
