2026-08-03 · Jack Stovell
How AI assistants decide which sources to cite
AI assistants decide what to cite through retrieval: the assistant turns your question into search queries, takes the top results from a search index, fetches those pages, extracts the passages that actually answer the question, and cites the pages the surviving passages came from. A citation is therefore a survivor of four filters — ranked, fetchable, readable, quotable — and a page can fail at any one of them. The ranking factors inside these systems are not public, but the pipeline itself is observable, and every stage of it is something your page passes or fails.
Where do AI citations come from?
From live retrieval at answer time, not from the model's trained memory. There are two separate paths by which an AI system can know you: training data baked into model weights, which accrues slowly and has a cutoff, and live retrieval, where the assistant searches, fetches pages, and answers from what it can read right now. When an answer arrives with sources attached, you are looking at the second path. That distinction matters practically, because retrieval is the path you can influence this week. The pipeline runs roughly: query formulation, then a search index, then a fetch of the top candidates, then extraction of relevant passages, then synthesis of an answer with attribution. Each step is a filter — a page that never ranks is never fetched, a fetch that gets blocked returns nothing, a fetched page that is a JavaScript shell reads as empty, and a readable page with no passage that answers the question gets read and then ignored.
What makes a page quotable?
A page is quotable when it contains a self-contained passage that directly answers the question being asked. In practice that means:
- A direct answer high on the page — the opening paragraph, not a conclusion buried under preamble. Extraction rewards pages that answer first and elaborate second.
- Question-shaped headings that match how people actually phrase the query, so the extractor lands on the right section.
- Self-contained passages: a paragraph that keeps its meaning when lifted out of context, with the subject named rather than pronoun-referenced.
- Clear entity identity — one canonical URL, consistent naming, identity markup — so the system is confident about who is speaking and does not conflate you with someone else.
- Served HTML that is readable without JavaScript, because most AI fetchers never execute it.
What stops a page being cited?
Most non-citations are mechanical failures, not quality judgements. The recurring disqualifiers:
- Login walls. Content behind authentication cannot be fetched at answer time — LinkedIn profiles are the canonical example, walled off from most crawlers.
- JavaScript shells. The fetch succeeds but the returned HTML is near-empty, because the content only exists after client-side rendering.
- Blocked crawlers. A robots.txt rule or an edge-level bot block — sometimes one the site owner never chose — refuses the fetch outright.
- Ambiguous identity. If the system cannot tell which same-named person or business a page describes, citing it becomes risky and it often simply won't.
- Never ranking at all. Retrieval starts from a search index, so a page invisible in ordinary search never reaches the fetch stage — this is where getting seen in ChatGPT and classic search visibility overlap.
How do you write content AI wants to cite?
Answer the question in your first paragraph, in prose that stands alone, on a page any crawler can fetch and read. That single habit addresses more of the pipeline than anything else, because it serves the extraction stage directly. Beyond it: use question-phrased headings, keep passages self-contained, state your identity unambiguously, and make sure nothing mechanical — blocks, shells, walls — disqualifies you before content quality is even considered.
Honesty about the limits: nobody outside these companies knows the ranking weights, and anyone quoting precise citation factors is guessing. What we can say is what the mechanism observably rewards, and that removing disqualifiers is necessary but not sufficient — after that, retrieval, ranking and model behaviour decide. We track this against our own pages in public: our scoreboard at hilyt.it/geo currently records that models do not yet cite hilyt profiles for the queries we track. We publish the misses because the mechanism deserves honest reporting, not just success stories. The per-platform specifics differ enough to warrant their own guides — see how to get cited by Perplexity — and our free checker at /tools/ai-visibility tests the mechanical disqualifiers against your site in one pass.
Check your AI visibility — free
Related guides
- How to get seen in ChatGPT — the mechanics, not the myths
- How to get cited by Perplexity
- Generative engine optimisation (GEO), explained