What is the Prompt-to-Citation Funnel?

At a glance

• The funnel has four stages: the prompt is asked, sources are retrieved, some are selected, a few are cited.
• Being retrieved and being cited are different outcomes, and the gap between them is where content work belongs.
• Each stage fails for a different reason, so one visibility number cannot tell you what to fix.

From a typed question to a named source

Someone asks an assistant a question. Four things happen in order, and you can be eliminated at any of them.

First the prompt is expanded. The assistant splits one question into several searches, which decides what gets looked up at all.

Second, retrieval. Each sub-search pulls back candidate passages. If your page cannot be crawled, or nothing on it matches the sub-topic, you are out and nothing downstream matters.

Third, selection. The model keeps the candidates it will actually write from. A retrieved page that buries its answer in the fifth paragraph loses to one that states it first.

Fourth, citation. The answer gets written and a few sources get named. You can be read and used without being credited, which is the most common frustration.

Knowing which stage is leaking

“We are not showing up in AI search” is not a diagnosis. It describes four problems with four different fixes, and guessing wrong costs a quarter.

Never retrieved means the problem is technical or topical: crawlers blocked, content rendered in JavaScript, or no page on that sub-topic at all. Better copy will not help.

Retrieved but not selected means the content exists and is not usable. Too long, too hedged, answer buried, marketing language where a plain sentence belongs.

Selected but not cited is closest to a win. The model used your material and named someone else, usually a source it treats as more authoritative. That is an evidence and authorship problem, not a structure one.

Three different briefs, and one vague metric would have sent you to the wrong one.

Find your weakest stage first

You can map this by hand for a handful of prompts, and it is worth doing before buying anything.

1. Pick ten prompts that matter. Real buyer questions, not brand searches. Prompt tracking across a fixed list keeps this comparable month to month.
2. Run each one and read the citations. Cited, mentioned without a link, or absent.
3. Test retrieval separately. Search the same sub-question on an ordinary search engine. If your page appears nowhere, you have a retrieval problem, not a writing problem.
4. Read the pages that beat you. An LLM audit does this systematically. Look at how directly they answer, not how long they are.
5. Fix one stage across all ten prompts before moving on. Mixing fixes teaches you nothing about which one worked.