LLMO

How LLMs Pick Sources: The Retrieval, Scoring, and Citation Mechanism

LLMs do not search the web in real time the way SEO writing often implies. The actual mechanism is closer to information retrieval than to search ranking, and the implications for LLMO are direct.

By Dmitry Paranyushkin · Updated

Most explanations of how LLMs pick sources flatten the actual mechanism into slogans — “LLMs cite trusted sources,” “build E-E-A-T,” “write for entities.” The slogans are not wrong; they are not specific enough to drive structural decisions. The mechanism underneath has identifiable stages, each with implications for how to structure content.

Stage 1: Retrieval

Most production AI search systems use a retrieval-augmented architecture. The query gets encoded as a vector by the same kind of model that encoded the passages in the index. Cosine similarity (or a closely related metric) returns the top N passages whose embeddings sit closest to the query embedding.

The implications for content structure are immediate. The retrieval system matches at the passage level, not the page level. A page with one well- structured passage and a lot of surrounding text may be retrieved for that passage; a page with the same information spread across diffuse paragraphs may not. Headings, definitional sentences, and tight self-contained passages are the retrieval-friendly shapes.

Embedding-based retrieval also makes the entity layer load-bearing. Strings that differ in surface form but map to the same entity tend to embed near each other; strings that share words but refer to different entities embed further apart. Writing about the entity, with the entity named explicitly in headings, raises retrieval recall on queries about that entity.

Stage 2: Scoring

The retrieved candidate set goes through a scoring pass before synthesis. The scoring signals vary across engines, but four categories appear consistently in published research and observed behavior.

  • Source authority. Domain-level signals comparable to but not identical to Google’s E-E-A-T. Sources with consistent topical coverage and verifiable authorship rank higher.
  • Freshness. For time-sensitive queries, recent passages rank higher. For evergreen queries, freshness contributes less and may not contribute at all.
  • Structural quality. Passages with clear extractable structure (definition, comparison, named entity, FAQ format) score higher than passages buried in narrative prose. The structural signals are observable even when the engine does not publish its scoring.
  • Citation graph (in some engines). Perplexity and several others maintain explicit citation graphs that weight passages by how often they get cited across the engine’s sessions. Self-reinforcing once the engine starts citing a source consistently.

Stage 3: Synthesis

The synthesis stage compresses the top passages into a single answer. Citations appear when the engine can confidently attribute a claim to a specific passage; when the model paraphrases without confident attribution, the citation drops.

Two implications follow. First, structurally distinctive passages (a 40–60 word definition, a comparison table, a misconception block) get cited more often than generic prose because the model can attribute the claim cleanly. Second, network-level consistency across a source raises citation rate because the model can attribute multiple claims to the same source confidently.

The paraphrase-without-citation pattern is the canonical failure mode for publishers. The model uses the information; the citation goes elsewhere or nowhere. Structural extractability is the main lever against the failure.

What this implies for content structure

  • Surface the extractable passages early. The 40–60 word definitional answer in the first 100 words of the page is the single highest-leverage structural move.
  • Name entities in every heading. Headings populate the embedding space the retrieval system searches; entity-named headings get matched against a wider set of related queries.
  • Hold definitions consistent across the network. Definitional drift weakens the trust signal at the source level.
  • Use comparison tables where queries span two entities. Tables are the structural shape that gets cited most often in cross-cluster answers.
  • Name common misconceptions explicitly. Where the model has training-time uncertainty, an explicit misconception block raises the source’s authority on the disambiguating claim.

Common misconceptions

Frequently asked questions

How do LLMs decide which sources to cite?
Three stages: retrieve candidate passages by embedding similarity, score them by relevance and trustworthiness and structural quality, synthesize the top passages into an answer with citations attributed where the model can confidently link a claim to a passage.
Do LLMs prefer high-authority domains?
Authority contributes. It does not dominate. Structural extractability is the swing factor; well-structured mid-tier domains regularly outcite poorly structured high-DA domains.
How does ChatGPT pick which links to cite?
ChatGPT runs its retrieval-augmented variant on top of Bing’s search index. The mechanism is the three-stage pattern above; the index source is Bing-derived.
What is retrieval-augmented generation?
RAG is the architecture pattern where an LLM is paired with a retrieval system; the retrieval pulls candidate passages from an index, and the LLM synthesizes an answer from those passages. Most production AI search runs on RAG variants.
Can I get my content into an LLM's training data?
Difficult to influence directly. The retrieval-and-cite path is the accessible optimization layer for most publishers. Training-data influence is possible at scale but indirect.
LLMO is what happens when content stops being optimized for search engines and starts being structured for retrieval systems. Read the full guide or run a free LLMO audit.