Guide
How AI Search Works: How LLMs Choose What to Cite
The short answer
How do LLMs choose what to cite?
AI answer engines choose citations in 3 stages: retrieval, where a search index returns candidate pages; chunk selection, where self-contained passages are pulled from those candidates; and synthesis, where the model writes an answer and attributes claims to the passages it used. A page can only influence the stages it participates in. The Princeton GEO paper measured visibility gains of up to 40% from optimizing the passage layer.
Every AI answer engine — ChatGPT, Perplexity, Google AI Overviews and AI Mode, Copilot, Gemini — assembles answers the same basic way: fetch evidence, select passages, write prose, attribute claims. The trade press wraps this in mystique; the pipeline itself is short, and each stage is a place your pages either stay in the running or drop out. This page walks the pipeline in operator terms — what each stage does, what the evidence says it rewards, and where you have no influence at all.
What happens between a question and an AI answer?
Three stages: retrieval, chunk selection, and synthesis. Retrieval-augmented generation — RAG — is the umbrella term: the model pulls current documents before answering rather than relying only on what it memorized in training.
First, the engine turns the user's question into search queries. Google documents this as query fan-out: "a set of concurrent, related queries generated by the model to request more information and fetch additional relevant search results." One question becomes several searches, which is why a page targeting a narrow sub-question can appear in the answer to a broad one — the mechanics on Google's side are detailed in what Google AI Mode is.
Second, the engine breaks retrieved pages into chunks — self-contained passages — and scores them for relevance to each sub-query. Third, the model synthesizes an answer from the selected chunks and attaches citations to the sources it drew from. iPullRank's AI Search Manual frames the discipline that emerges from this as relevance engineering: you are optimizing passages for a retrieval system, not pages for a ranked list.
Where do AI engines get their candidate pages?
From search indexes — each platform builds or borrows one, and being absent from the index ends your candidacy before any model sees you. Google's AI features retrieve from Google's core index; its guide states the generative features "are rooted in our core Search ranking and quality systems." ChatGPT search retrieves from the index OAI-SearchBot builds, and OpenAI states that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers. Perplexity's index is built by PerplexityBot, which its docs describe as used for search indexing, not model training.
| Stage | What the engine does | What the operator controls |
|---|---|---|
| Retrieval | Fan-out queries hit a search index; candidate pages return | Crawler access, indexation, topical relevance |
| Chunk selection | Pages are split into passages; passages are scored per sub-query | Self-contained sections, answer-first structure |
| Synthesis | The model writes the answer, attributing claims to used passages | Specific, dated, attributable claims worth quoting |
| Memory answers | The model answers from training data; nothing is fetched | Nothing at answer time |
These crawlers are visible infrastructure, not theory: ClaudeBot, GPTBot, and PerplexityBot all appear in our fleet's server logs across our three production builds [our data]. The full roster of documented bots, token by token, is in the complete AI crawler list.
How does chunking decide which passage gets quoted?
Engines quote the chunk, not the page — so the unit you are really publishing is the self-contained passage. When a page is split for retrieval, each passage is scored on its own, stripped of everything around it. A paragraph that opens with "as we saw above" or leans on a previous section's definitions scores as noise, because the context it leans on is gone.
The passages that survive extraction share a shape: they open by answering a question, they carry a specific fact or number, and they still make sense read alone. That shape is what this library calls a citable passage, and it is the reason answer-first pages exist as a discipline rather than a style preference.
This stage is also where page structure quietly becomes retrieval infrastructure. Question-shaped headings give the chunker clean boundaries; 2–4 sentence paragraphs keep a thought from being truncated mid-claim; a table gives the engine a fact it can lift without paraphrase risk.
What did the Princeton GEO paper actually measure?
The paper that named the field — "GEO: Generative Engine Optimization" (arXiv 2311.09735, KDD 2024) — reports that content-side optimizations "can boost visibility by up to 40% in generative engine responses." The tested tactics that performed best were textual: adding source citations, quotations, and statistics to the content itself. Markup and technical tricks were not the levers; the winning changes made passages more attributable.
Read the number with its limits attached. It is a benchmark result — measured across a large query benchmark against generative engines in a controlled setting — not a field guarantee, and the authors state plainly that "the efficacy of these strategies varies across domains." A tactic that lifted visibility for one domain's queries did less, or nothing, for another's. Up to 40% is the ceiling observed in an experiment, not a rate anyone should project onto their own site.
The honest translation for operators: the paper is strong evidence that passage-level content changes move retrieval-stage outcomes, and weak evidence about how much they will move yours.
Does freshness change what LLMs cite?
The best available data says yes, meaningfully. Kevin Indig's State of AI Search Optimization 2026 (January 2026) reports that content less than 3 months old is roughly 3 times more likely to be cited. One analysis, not a replicated law — but it is consistent with how retrieval works: answer engines re-crawl aggressively and prefer current evidence for questions where currency matters.
This is also a claim we can sanity-check against infrastructure we watch: AI crawlers revisit our published pages on an ongoing basis in our fleet's logs — the fetching never stops after first indexation [our data]. Recency bias only functions because the engines keep coming back.
What can't you influence at all?
The model-memory layer — and it is larger than most coverage admits. Indig's 2026 analysis found 24% of ChatGPT answers involve no fetch: the model answers from training data, cites nothing, and no page structure, schema, or freshness signal can touch that answer at generation time. Whatever the model absorbed about your topic in training is what those answers contain.
Two more limits belong on the table. Platform weighting is opaque: which sources an engine trusts, and how much, is not documented anywhere you can act on. And synthesis is probabilistic: the same question can produce different citations run to run. This is why we treat every "get cited in ChatGPT, guaranteed" pitch as disqualifying — nobody controls answer-engine output, including the platforms' own documentation teams. What an operator controls is candidacy: present in the indexes, extractable at the passage level, specific enough to be worth attributing. The full operating pattern we run across our builds is in our generative engine optimization guide, and the platform-specific mechanics are in how ChatGPT selects its cited sources.
Frequently asked questions
How do LLMs decide which sources to cite?
Through a 3-stage pipeline: a search index retrieves candidate pages, self-contained passages are extracted from them, and the model synthesizes an answer while attributing claims to the passages it actually used. Pages compete at every stage — retrievability, extractability, and having specific claims worth attributing.
What is RAG in plain terms?
Retrieval-augmented generation — RAG — means the model fetches current documents before answering instead of relying only on training memory. All major AI search products work this way: retrieval supplies the evidence, the model writes the prose, and citations point back at the retrieved sources. It is 1 pipeline with different indexes behind it.
Does ChatGPT search the web for every answer?
No. Kevin Indig's State of AI Search Optimization 2026 found 24% of ChatGPT answers involve no fetch — the model answers from training memory. Those answers cite nothing, so no page optimization can influence them. The remaining answers retrieve from the index OAI-SearchBot builds.
Does fresh content really get cited more?
Kevin Indig's 2026 analysis reports content under 3 months old is about 3 times more likely to be cited. That is one study's finding, not a law — but it matches the retrieval logic: engines re-fetch often and prefer current evidence for time-sensitive questions.
Can you make an AI engine cite your page?
No — nobody controls answer-engine output, and no one can promise a citation. What you control is candidacy at each of the 3 stages: being in the indexes engines retrieve from, publishing passages that survive extraction, and carrying specific attributable claims. The rest is the engine's decision.
Sources
- GEO: Generative Engine Optimization — Princeton University et al.
- Google's Guide to Optimizing for Generative AI Features — Google
- State of AI Search Optimization 2026 — Growth Memo (Kevin Indig)
- OpenAI crawlers documentation — OpenAI
- Perplexity Crawlers — Perplexity
- The AI Search Manual — iPullRank