What Content Does AI Cite? — How Generative Engines Choose Citations
How generative engines like ChatGPT and Perplexity pick the sources behind an answer, explained as a three-step process — retrieval, grounding, and synthesis — plus the conditions that make content citable: extractable chunks, semantic density, source credibility, and freshness.
You can pour effort into your content, but if it never gets cited in a ChatGPT or Perplexity answer, that writing is invisible in the world of generative search. Many operators ask, "We wrote something good — so why doesn't AI pick us up?" The problem isn't only the quality of the writing; it's that the content's structure is out of step with the way AI chooses sources. This article works through how generative engines select the evidence behind an answer, starting from first principles, and then lays out what you have to change to raise your odds of being cited.
The three steps a generative engine takes to choose sources
A generative engine's answer doesn't arrive all at once; it is usually built through three steps. Once you understand the order, you can see where content drops out.
| Step | What it does | Effect on citation |
|---|---|---|
| Retrieval | Gathers documents semantically close to the question as candidates | If you don't make the candidate set, you never even enter the later steps |
| Grounding | Narrows the candidates down to the passages directly used for the answer | An answer that closes within a single paragraph is more likely to be selected |
| Synthesis | Stitches the chosen evidence together into answer sentences | Citation marks attach only to the sources adopted as grounding |
The key point is that the citation appears in the final step, attached only to the sources actually adopted as grounding. Making the candidate set in the retrieval step is not enough; the grounding step has to judge "this passage is the answer to the question." Because gathering candidates in retrieval starts from whether an AI crawler can read the page properly, the rendering and accessibility described in Google's crawler documentation are the precondition.[5] An article that covers this process more broadly from a cost and strategy angle continues at What Is GEO.
The smallest unit of citation is the "extractable chunk"
AI does not cite a page whole. It extracts the part that corresponds to the answer — that is, the chunk — and uses it. Here a chunk generally maps to a single paragraph. So even for the same information, citation likelihood varies greatly depending on how it is arranged.
Why self-contained paragraphs have the advantage
If the question is "What's the difference between AEO and SEO?" but the answer is scattered across a definition in the paragraph above, an example in the paragraph below, and one cell of a table, the engine struggles to form a single clean unit of grounding. Conversely, when the answer closes within one paragraph — "the difference is X, the reason is Y, and for example Z" — that paragraph itself becomes a citation candidate. This is exactly why tables, definitions, and FAQs are recommended in GEO. They structurally separate question-answer pairs and make extraction easy. Google's recommended FAQ structured data is a way to spell out that separation in a machine-readable form.[3]
What makes a chunk citable — four attributes
Whether a chunk gets adopted in the grounding step comes down to roughly four attributes. We look at each cause, its effect on citation, and the action to take, together.
| Attribute | Cause (why it works) | Action |
|---|---|---|
| Extractable chunk | The answer must close within one paragraph to be cut out as a unit of grounding | Split paragraphs by question and put the answer in the first sentence |
| Semantic density | A claim with no evidence is hard to judge as trustworthy | Attach figures, definitions, and sources to each claim in the same paragraph |
| Source credibility | The more verifiable the source, the more it's preferred during synthesis | Link primary materials and standards documents inline |
| Freshness | For changing information, older documents get excluded | Keep date metadata accurately updated |
These four attributes are not abstract advice. Semantic density — the combination of a clear claim and the evidence attached to it — is especially decisive. Repeating the same keyword does not raise semantic density. Rather, when "a claim + the statistics, definitions, and sources that support it" come together in one place, the engine treats that paragraph as safe grounding. On the credibility side, using Schema.org Organization or structured data to spell out the publisher and the nature of the content in machine-readable form works as a supporting factor.[2] In addition, practices like the llms.txt proposal, which point AI to the documents it should reference first, contribute to securing candidates in the retrieval step.[4]
What the GEO paper showed experimentally
This mechanism is backed not by intuition but by measurement. The paper by Aggarwal et al. (arXiv 2311.09735), which formalized the term GEO, tested which content edits raise visibility within answers in a generative-engine setting. The gist of the result was that edits that add citations, statistics, and sources lift visibility within generated answers by up to roughly 40%.[1] Conversely, plain keyword stuffing had little effect or even hurt.
This result lines up exactly with the four attributes above. Adding statistics and sources is an action that raises both semantic density and source credibility at once, while repeating keywords improves none of the attributes. That said, this figure is the average effect under particular experimental conditions, so it does not mean a 40% lift is guaranteed identically across every domain and engine — keep that in mind. The stable takeaway is the direction (edits that add evidence are advantageous) rather than the magnitude of the effect. How to translate this direction into actual document structure continues in Content Structure That Gets Cited in AI Answers, and reinforcing trust at the entity level continues in Entity and Knowledge Graph Optimization.
Wrapping up
A generative engine gathers candidates through retrieval, picks the passages to use for the answer through grounding, and then, in synthesis, attaches citations only to the sources it adopted. Citation therefore happens at the level of an "extractable chunk" rather than the whole article, and whether that chunk gets adopted comes down to four attributes: an extractable structure, semantic density, source credibility, and freshness. As the GEO paper showed, edits that add citations, statistics, and sources meaningfully raise visibility, whereas repeating keywords does not. In the end, the path to more citations lies not in more words but in structural design that closes the answer at the level of a single question and places the evidence right alongside that answer.
Frequently asked questions
- A generative engine first retrieves documents that are semantically close to the question to gather candidates (retrieval), then picks the passages that directly support the answer (grounding), and synthesizes the response (synthesis). At that point, the more clearly an answer closes within a single paragraph, comes from a trustworthy source, and is relatively fresh, the higher its chance of being adopted as grounding.
- Only part of it. Rather than reading a page in full, the engine extracts the chunk (paragraph-level) that corresponds to the answer and uses that. So when the core answer is scattered across several paragraphs and a table, extraction gets hard, whereas a passage that closes self-containedly within one paragraph is more likely to become a citation candidate.
- No. A generative engine looks not at keyword frequency but at semantic match to the question and the clarity of the evidence. Rather than repeating the same word, putting the claim and its evidence (figures, sources, definitions) together in one paragraph raises citation likelihood.
- It depends on the topic. For information that changes — prices, versions, policies — recent documents are strongly preferred, whereas for stable topics like definitions or principles, credibility and clarity weigh more. Keeping date metadata accurate helps in every case.
- First, split your content by the questions you most often get, and make each paragraph close as an answer to that question. Then attach evidence to every claim (statistics, definitions, source links), and check whether the page is readable by AI crawlers (rendering, crawler access, structured data).
Q.On what basis does AI choose the sources for an answer?
Q.Does AI cite the whole page, or only part of it?
Q.Does stuffing in lots of keywords help get cited?
Q.How important is freshness?
Q.Where should I start if I want more citations?
Sources
Related documents
- Content Structure That Gets Cited in AI Answers — Writing for ExtractabilityThe writing AI cites is not the same as writing that reads well. How to raise extractability through citable units, answer-first placement, question-answer structure, and tables, lists, and definitions — grounded in GEO research and a practical checklist.
- What Is GEO — The Definition of Generative Engine Optimization and How It Differs From SEOGEO (Generative Engine Optimization) is the strategy of getting your content cited in answers produced by generative engines like ChatGPT and Perplexity. Here is the definition, how it differs from SEO, and how it works.
- Entity and Knowledge Graph Optimization, Explained — How AI Recognizes Your BrandEntity SEO and knowledge graph optimization make AI recognize your brand as one clear 'entity.' How sameAs, schema.org Organization, and Wikidata connections shape AI trust, and the steps to put them in place.