How AI Engines Choose What to Cite: The Mechanism, From the Research

Ask why a particular page ended up under an AI answer and most explanations reach for ranking language: authority, backlinks, position. That framing is wrong in a way that costs money. Citation is a retrieval decision made when the question is asked, from whatever the engine currently holds, and it responds to different inputs than a search ranking does.

The short version

An engine assembles an answer by retrieving passages that fit the question, then writing over them. The primary research from Aggarwal et al. at Princeton measured this directly: adding citations, quotations and statistics to a passage moved its visibility inside generated answers by up to 40% in controlled tests. That is a passage-level effect, not a domain-level one. Our own 62-query panel adds an uncomfortable second finding: the retrieval pool is barely an agency-website pool at all. Across 301 citations it drew on 178 different domains, 131 of them cited exactly once, and the single largest source was LinkedIn at 6%.

What the engine is doing

Retrieval, not ranking

The engine gathers candidate passages that answer the question. Nothing is pre-sorted into positions. A page that answers one sub-question precisely can be pulled while a stronger domain that answers it vaguely is skipped.

Passage-level, not page-level

Selection happens below the page. A 2,000-word article with one clean, self-contained answer block can be cited for that block alone, which is why structure beats length.

Assembled per query

The same page can be cited for one phrasing and ignored for a near-identical one. That is not inconsistency; it is the mechanism working as designed, and it explains the churn we measured across 52 days.

What the research says moves it

Passage change	Measured effect	What it means for a page
Adding cited sources	Up to 40% visibility lift	Link the evidence, not the summary
Adding statistics	Large positive effect	Quantify claims rather than asserting them
Adding quotations	Positive effect	Attributed, specific, checkable
Keyword stuffing	Negligible or negative	The old lever does not transfer

Those figures come from controlled tests on a benchmark, not from a Dubai agency panel, and we cite them as mechanism rather than as a promise about your results.

Where our own data complicates the picture

The research describes what makes a passage citable. It does not tell you whose passages are in the pool. We measured that separately. Across our 62-query panel (16-18 September 2026), YouTube, LinkedIn, Reddit and Quora hold 54 of the 301 citations (18%) - a bigger bloc than any single website, and present in 27 of the 56 answers that cited anything. No agency or vendor domain held more than 2%.

The practical consequence is blunt. Making your page maximally citable raises your odds inside the pool you can reach. It does nothing about the majority of answers assembled from surfaces you do not own. That is the away game, and it is bought with placement, not with markup.

What to do with the mechanism

1. Put a direct answer in the first 130 words of any page targeting a question. The engine retrieves passages; give it a clean one.
2. Quantify every claim you can defend. "Most brands" retrieves worse than "3 of 10 brands in our panel".
3. Cite your sources inline rather than in a bibliography. The research effect attaches to the passage, not the page footer.
4. Write one self-contained block per sub-question, rather than one long argument that only makes sense in order.
5. Audit which of your target answers are sourced off-domain before spending another hour on your own pages.

Questions buyers ask

Does this mean backlinks do not matter? They matter for the older ranking systems that still feed some of the retrieval pool. What the research shows is that passage-level evidence moves generated-answer visibility independently, which is a lever most sites are not pulling.

If I add statistics, will I get cited? It raises the probability of the passage being selected. Nothing guarantees selection, and our own pages carry heavy evidence while scoring 0% across three readings on our 8 Dubai questions.

Why does the same page get cited for one question and not another? Because retrieval runs per query. Small phrasing differences change which passages clear the bar, which is the same mechanism behind the churn we measured.

Is Entity clarity part of this? Yes, upstream of it. The engine has to resolve who you are before your passages are candidates for questions about your category. Retrieval selects passages; Entity decides whether you are in the conversation.

Does freshness matter to retrieval? It appears to, and our own decay checks treat it as a lever. A page that has not moved in two years competes against pages written since.

How do I know which of my pages are being retrieved? Ask your buyer questions and read the source list under each answer. Your cited pages are the ones the engine currently finds useful. As Firoz Azees puts it: "a measured zero is a work order; a flattering estimate is a story you tell yourself."

What is the single highest-leverage change? An answer block at the top of the page, quantified and sourced. It is cheap, it is passage-level, and it is the change the research measured most directly. The limits of this discipline are worth reading alongside it.

At Ivanooo, Firoz Azees runs Distinctiveness Engineering for the AI-answer era: measuring who the engines name, cite and recommend, then engineering the gap between listed and chosen. Start with a free AI visibility check and see which of your pages the engines currently retrieve.