How AI search retrieves and cites sources

The six-stage path from question to citation

The most useful working model is a pipeline. Different engines combine or rename stages, and their exact systems are proprietary, but the model shows where a good page can disappear before a citation is produced.

1. Interpret and expand the question

A question may be rewritten into several searches. A comparison query can trigger searches for definitions, features, prices, limitations and alternatives. Google calls this query fan-out in its public explanation of AI Overviews and AI Mode.

2. Retrieve candidates

The system searches an index or another retrieval source for relevant documents and passages. If the page is blocked, non-canonical, unindexed or only understandable after unsupported interaction, it may never enter the candidate set.

3. Form passages or chunks

Long pages are represented in smaller units. A chunk needs enough local context to make sense by itself. A heading plus an answer, definition, condition and evidence is more portable than a paragraph full of 'it', 'this' and references to a chart several screens away.

4. Rerank the evidence

Initial retrieval aims for recall; reranking narrows the candidates. Specificity, relevance, source quality, freshness and consistency can all matter. This is why mentioning every related term is not enough: the passage must be the best support for the exact sub-question.

5. Generate the answer

The model composes a response from selected context and its instructions. It may combine several sources, omit a retrieved source or decide that the evidence is insufficient.

6. Attribute sources

The interface attaches one or more links to support claims. Retrieval and citation are related but not identical: a page can influence an answer without receiving the visible citation, and a citation can be attached at a broader sentence level than the precise passage used.

The anatomy of a citable passage

A citable passage does one job completely. It identifies the subject, answers the question immediately, states the conditions or scope, and provides evidence or a method when needed.

Citation rate is the share of observed prompt responses in which a domain or URL is named as a source. Measure it over a fixed prompt set, engine set and time window; otherwise the number cannot be compared.

That definition can travel because the noun is explicit, the metric is bounded and the measurement condition is included. It does not require the preceding paragraph to repair its meaning.

Why query fan-out changes content planning

A page about 'the best GEO tools' may be competing across hidden searches for monitoring tools, crawler diagnostics, citation tracking, analytics, pricing and alternatives. One thin list is unlikely to be the strongest source for every branch. A durable content architecture uses a clear hub and focused supporting pages connected by descriptive internal links.

How to improve selection odds

• Use descriptive headings that mirror real questions and sub-questions.
• Lead with the answer, then explain, qualify and prove it.
• Prefer original definitions, procedures, comparisons and data over generic summaries.
• Name the author, publication date, update date and method where trust depends on them.
• Link to primary sources and distinguish observed facts from inference or opinion.
• Keep the same fact identical in page copy, metadata, structured data and feeds.
• Provide stable canonical URLs and meaningful internal links so each passage has a discoverable home.

A crucial limit: citation is probabilistic

No page earns a permanent citation slot. Results vary with the user, language, location, model, index freshness, interface and exact wording. The honest objective is to improve eligibility and citation frequency across a representative sample—not to promise that one optimisation will force one engine to quote one URL.