THE SHORT ANSWER
The typical pipeline expands one question into several sub-queries, retrieves candidate passages for each from an index, reranks them, feeds the survivors to the model, and then attaches citations to whichever passages supported the generated text. Selection therefore happens at passage level, not page level, and it favours a specific, self-contained chunk that answers a sub-question directly. Authority breaks ties between comparable passages rather than deciding the contest.
The mental model most marketers carry is that the assistant runs your query, reads the top ten results and picks its favourites from among them. That is not how any current system works, and the gap between the model and the mechanism matters, because it changes what you should be writing and where on the page you should put it.
The actual sequence has four or five distinct stages, each of which can eliminate you for reasons that have nothing to do with the stage before it. A page can be perfectly authoritative and never survive retrieval; a page can survive retrieval and never get cited because nothing in it was quotable enough to support a sentence.
The numbers, at a glance
Stage one: query fan-out, where one question becomes several narrower sub-queries run independently
Stage two: retrieval, typically hybrid: keyword matching and vector similarity over chunked documents
Stage three: reranking, where a smaller model scores candidates for how well they actually answer the sub-query
Stage four: generation and attribution, where citations are attached to the passages the answer leaned on
Fan-out: your competition is not the question you think
A question like which heat pump suits a 1930s terraced house is not run as a single query. It is decomposed into several: sizing for poor insulation, radiator compatibility, typical costs for that property type, permitted development constraints. Each sub-query goes to retrieval separately, and each has its own set of winners.
This has a direct editorial consequence. A page that covers all four sub-topics shallowly loses each of the four contests to a page that covers one properly. Depth on a narrow question beats breadth across a broad one, which inverts the pillar-page instinct that a decade of content strategy trained into everybody.
Retrieval and reranking: two filters with different tastes
Retrieval is usually hybrid. A lexical component matches literal terms, which is why the specific vocabulary of your trade still matters and why writing around a term instead of using it costs you. A vector component matches meaning, which is why a passage phrased differently from the query can still surface if the semantic content lines up.
Reranking then applies a slower, more discriminating model to the shortlist, scoring how directly each candidate answers the sub-query rather than how similar it looks. This is the stage where a passage that opens with its conclusion beats a passage that arrives at the same conclusion in the last sentence, because the top of the chunk carries disproportionate weight in the scoring.
Attribution: why some cited pages barely influenced the answer
Systems differ on when citation happens. Some ground each generated statement in a retrieved passage as they go. Others generate first and match sources afterwards, which is how a page occasionally ends up cited for a claim it does not quite make. Neither approach is publicly documented in detail by any major vendor, and the behaviour changes between releases.
What survives all the variants is that a sentence-level match is what gets rewarded. If your page contains one sentence that states a fact cleanly, with the subject named and a number attached, that sentence is what gets pointed at. If your page expresses the same fact across three sentences with pronouns linking them, there is nothing compact enough to attach a citation to.
The selection signals you can actually influence
Passage self-containment. Every section readable cold, subject named, no reliance on what came before it.
Specificity. A number, a range, a threshold or a named condition. Hedged generalities give the reranker nothing to prefer.
Vocabulary match. Use the term buyers use, then explain the technical one, rather than the other way round.
Source diversity pressure. Systems avoid citing the same domain repeatedly in one answer, which means breadth of topic coverage across pages beats stacking everything onto one URL.
Freshness on the claim. A dated figure is preferred to an undated one, and both beat a number that is visibly stale.
Corroboration. A claim that agrees with other trusted sources is safer to repeat, which is why contrarian positions need visible evidence to travel.
Engineering a page for citation
Write down the four or five sub-questions your main question would decompose into, and give each its own section.
Open every section with the answer, then the qualification, never the reverse.
Put one checkable number in each section and state the date it refers to.
Delete cross-references such as as described above and replace them with the subject noun.
Read each section aloud in isolation and cut it if it does not make sense without the rest of the page.
Want leads like this in your pipeline?
Flock runs the campaigns, screens the enquiries and hands you only the ones that match your service area, job size and capacity. You pay per lead, not per month.
Book a 15-minute fit check | See lead package pricing
Related answers
Frequently asked questions
Does domain authority still decide who gets cited?
It acts as a tiebreak rather than a gate. Between two passages that answer the sub-query equally well, the more trusted domain usually wins. Between a weak passage on a strong domain and a strong passage on a modest one, the passage frequently wins, which is why small sites can compete here.
Why does the same question return different sources each time?
Fan-out produces slightly different sub-queries between runs, retrieval indexes update continuously, and generation is probabilistic. Treat citation as a probability you are shifting over many runs, not a position you either hold or do not on any given day. This is why a fixed prompt set checked monthly beats any single spot check.
Can I be cited without ranking in classic search?
Yes, though it is less common than vendors imply. Most retrieval indexes overlap heavily with conventional search indexes, so poor ranking usually indicates a problem that hurts both. The clearest exceptions are very long-tail sub-questions nobody else answered at all.
How many sources does a typical answer use?
Usually three to eight, drawn from more candidates than that after reranking discards the rest. The set is deliberately diverse, so being the second or fourth citation is worth a meaningful share of the value rather than a consolation. Readers frequently open two or three of the linked sources rather than only the first one.
NEED A CLEARER PLAN?
Let’s turn your next move into momentum.
Talk to us →