How AI engines choose what to cite

AI engines choose what to cite in three steps: they retrieve pages that seem relevant, extract the passage that most cleanly answers the question, and prefer sources they can trust enough to name. A page wins a citation when it is findable, liftable and credible, not simply when it ranks well. Optimise for all three and you make the citation easy.

How do AI engines decide what to cite?

An AI engine does not read a page the way a person does. It works in three stages, and a page can fall out at any one of them. First it retrieves a set of candidate pages that look relevant to the question. Then it extracts the passage that best answers the query. Finally it decides which of those sources it trusts enough to name in the answer. Rubric's three pillars, Known, Findable and Trusted, map onto exactly these stages, so a weak pillar tells you which stage you are losing at.

Step 1: retrieval, can the engine find your page?

Retrieval is the engine gathering candidate pages to answer from. If a crawler cannot reach or read your page, nothing else matters, because the page never enters the candidate set. Retrieval depends on the AI bots being allowed in robots.txt and not blocked at the firewall, the page being indexable and reachable through internal links, an XML sitemap, and a fast server response. Several engines do not run JavaScript, so a page that only assembles its content after scripts run may be retrieved as an empty shell.

Step 2: extraction, can it lift a clean answer?

Extraction is the engine pulling the passage that answers the question. Engines lift the first self-contained passage that cleanly answers the query, so a page that buries its answer under introduction is hard to quote even when the answer is present. Extractable pages lead with a direct 40 to 70 word answer, use question and claim-shaped headings that match how the query is phrased, keep each section self-contained, and use tables and lists for comparative or sequential facts. See answer-first-content for the structure.

Step 3: trust, will it put your name to the claim?

Trust is the engine deciding whether the evidence is strong enough to attribute a claim to you. Engines prefer sources they can stand behind, so pages with sourced facts and clear authorship are named more often than confident but unsupported prose. This is the Trusted pillar: statistics that carry a source, external source links, a named author with real credentials, and content that has been kept fresh. It also depends on the engine being able to resolve who published the page, which is the Known pillar.

What the research says moves citations

The Princeton GEO study (KDD 2024) tested content changes against generative engines and found clear, directional effects on how often a page was surfaced in answers.

These are directional findings, not guarantees, but they line up with what Rubric rewards: evidence, sources and clean structure, not repetition.

How Rubric turns this into a score

Rubric turns the three-step decision into a score by running 44 checks across the Known, Findable and Trusted pillars and estimating citability from 0 to 100. A low Findable score means engines struggle to extract an answer; a low Trusted score means they can find the answer but hesitate to name you; a low Known score means they cannot resolve who you are in the first place. Read the-three-pillars to see which checks sit where, then work the weakest pillar first.