An AI citation is the output of three separate steps, not one decision - retrieval finds candidate passages, reranking narrows them under context constraints, and generation decides what to quote and how to attribute it. None of the major engines publish the exact weights inside that pipeline, and anyone claiming otherwise is guessing. But the pipeline's shape is public, documented, and consistent across engines - and understanding it explains something the pillar page already established: why a page can rank #1 and still never get cited. It can lose at any one of these three steps, independent of how well it ranks.
The Three-Step Mechanism, in Plain Terms
This is retrieval-augmented generation (RAG), the architecture behind essentially every AI answer engine's citation behavior, first described in a 2020 research paper and now the default pattern across the industry. Strip away the machine-learning terminology and it's three steps: retrieval pulls a set of candidate passages that are semantically relevant to the query - not just keyword-matched, but matched on meaning, using dense vector representations rather than exact text - from across the engine's index. Reranking narrows that candidate set further under real constraints: how much context the model can hold at once, how the passages compete against each other for relevance, how recent or specific they are. Generation is where the model writes the actual answer and decides which of the surviving passages to quote or paraphrase, attaching a citation to the ones it uses.
Microsoft's own documentation for Bing-grounded AI agents shows this concretely: a query goes out, Bing returns a set of relevant chunks, the model generates its answer from those chunks, and URL citations get attached as annotations pointing back to the source. That's not a metaphor - it's a description of how one major real-world implementation actually works, and the same basic shape holds across ChatGPT, Perplexity, and Google's AI surfaces even though none of them expose this level of detail publicly.
Why a Technically "Correct" Passage Can Still Lose
Retrieval and reranking are competitive, not additive - your passage isn't scored in isolation, it's scored against every other candidate passage retrieved for that query, and only a handful survive to the generation step. Two things move the needle here, and one popular claim doesn't. What helps: topical match at the passage level, not just the page level - a peer-reviewed study running 252,000 controlled citation trials across six LLMs found topical relevance and list position were the strongest predictors of being cited first, with a recent timestamp and explicit specifics (like pricing) helping consistently. What doesn't reliably help, according to that same controlled research: formatting changes made in isolation, without a corresponding improvement in the substance underneath them. Academic chunking research backs this up structurally - passages that preserve real document hierarchy and semantic coherence retrieve and rerank better than ones chopped into arbitrary fixed-size blocks, which is a different claim than "add more headers" or "shorten your sentences."
One more honest caveat worth stating plainly: even when a citation does appear, it isn't guaranteed to faithfully represent what the source actually said. Research on RAG attribution has found that model-generated citations can be malformed, pointed at the wrong passage, or not actually reflect how the model used the source's content. A citation is evidence the system selected your content - it isn't a guarantee of perfect representation, and treating it as one oversells what's actually been measured.
One Citation or Several? It Depends on the Engine and the Question
There's no fixed number. OpenAI's own documentation for ChatGPT search describes web retrieval producing inline citations and a sources panel that can include multiple links per answer; Google's AI Overviews and AI Mode surface varying numbers of supporting links depending on the query. Independent measurement backs this up with real numbers: one large comparative study found AI Mode averaging around 7 unique domains cited per query, AI Overviews averaging closer to 3, with ChatGPT and Perplexity falling in between - and Perplexity showing the highest domain overlap with Google's own top 10 of any engine measured (over 91%), while ChatGPT showed the weakest. The practical point: how many sources an engine tends to cite, and how closely that set tracks traditional rankings, is a property of the specific engine and query type - not a fixed rule you can plan around universally.
How Chunking Actually Works - and What It Doesn't Require From You
"Chunking" is something that happens on the retrieval system's side, not a formatting task for a writer to perform manually. Peer-reviewed chunking research shows that retrieval systems that preserve a document's real structure and hierarchy - rather than slicing it into naive, fixed-size blocks - retrieve and rank better. That's evidence for writing clearly structured content with real logical sections. It is not evidence for manually splitting a page into artificially tiny fragments in the hope of matching some AI-specific format. Google has said this outright, not just implied it: its own AI-optimization guidance states there's no requirement to break content into tiny pieces for AI to understand it, and no new machine-readable files or AI-specific markup are needed to appear in Google Search. The chunking happens after your page is published, inside the engine - your job is a well-structured page, not a pre-chunked one.
How to structure content so AI engines can actually quote it →
Does structured data actually help you get cited? →
The Technical Floor: What a Page Actually Needs to Be Machine-Readable
Before any of the above matters, a page has to clear a basic technical bar - and this is where the mechanism page genuinely overlaps with CoreAEX's other cluster on AI crawler access. In short: your content has to be publicly crawlable, indexable or snippet-eligible, and it has to render reliably for whichever crawler is trying to read it. OpenAI's OAI-SearchBot has to be able to reach a page for it to be eligible for normal inclusion as a cited source in ChatGPT search; Perplexity publishes the same expectation for PerplexityBot. If a page depends on client-side JavaScript to render its actual content and the crawler in question doesn't execute that JavaScript, none of the mechanism described above ever gets a chance to run - there's no passage to retrieve in the first place. That's a deep enough topic that CoreAEX has an entire separate cluster on it.
The full JS-crawler-access pillar - why this happens and how to fix it →
Confirm your own site clears this bar →
Chat Answers vs. AI Overviews: Not the Same Mechanism
Treating "AI search" as one system undersells how differently these surfaces actually behave. Google's AI Overviews and AI Mode are built substantially on top of Google's own search index and ranking signals - which is part of why the pillar page's ranking-correlation numbers for Google's own AI surfaces are meaningfully higher than ChatGPT's. But even within Google's own AI surfaces, that correlation is looser than a static ranking would suggest: one large Ahrefs study covering 863,000 SERPs and 4 million AI Overview URLs found that as of March 2026, just 38% of AI Overview citations came from pages in the traditional top 10 - down sharply from 76% a year earlier. A separate volatility study found that roughly 70% of the pages cited in AI Overviews change within any given 2-3 month window, moving largely independently of organic rankings (the two showed a correlation coefficient of just 0.09 in that research). Chat-based engines add a further layer of difference: ChatGPT's citation pool showed the weakest overlap with Google's top 10 of any engine measured in a large comparative study, while Perplexity's showed the strongest - meaning "optimize for AI search" as a single strategy is already an oversimplification before you even get to writing a sentence.
What This Actually Means for What You Write
Don't chase a formatting checklist as a substitute for topical substance - the strongest controlled evidence available says formatting alone barely moves the needle, while topical relevance and specificity do. Don't assume your page is even in the running before checking that it's technically crawlable and renderable in the first place - that's a precondition, not a nice-to-have. And don't treat "AI search" as one target: what gets a citation from Google's AI Overviews and what gets a citation from ChatGPT are shaped by measurably different mechanisms, with different, and moving, degrees of overlap with traditional ranking.
← Back to the AEO fundamentals pillar
Next: how to structure content so AI engines can actually quote it →
What makes a specific claim trustworthy enough to cite →
Sources: The three-step retrieval/reranking/generation mechanism draws on Lewis et al., "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks" (2020, the foundational RAG paper) and Microsoft Learn's documentation of a concrete Bing-grounded retrieval pipeline. Passage-level retrieval mechanics from Karpukhin et al., "Dense Passage Retrieval for Open-Domain Question Answering" (EMNLP 2020). The 252,000-trial finding on topical relevance, list position, freshness, and the weak effect of formatting-only changes is from Vishwakarma, Kumar, and Jamidar's peer-reviewed study (ACM SIGIR 2026) - the same source anchoring this claim on the pillar page. Chunking-and-structure research from Jain et al., "AutoChunker" (ACL Industry 2025) and "Equipping Retrieval-Augmented LLMs with Document Structure" (EMNLP Findings 2025). Citation-faithfulness limitations from "Model Internals-based Answer Attribution for Trustworthy RAG" (EMNLP 2024). Google's own no-special-markup statement from Google's AI optimization guidance. Engine-specific product mechanics from OpenAI's ChatGPT search documentation, OpenAI's crawler overview, and Perplexity's crawler documentation. The 38%-of-citations-from-top-10 figure (down from 76% a year earlier) is from Ahrefs' study (863,000 SERPs, 4 million AI Overview URLs, published March 2, 2026) - verified directly at the primary source. Domain/URL overlap-by-engine figures (AI Mode ~54% domain overlap with Google's top 10, Perplexity over 91%, ChatGPT weakest) are from a separate Semrush study (150,000+ citations, published July 21, 2025) - verified directly and confirmed as a genuinely distinct study from the one already cited on the pillar page, not duplicate reporting of the same research. The ~70% AI Overview turnover and 0.09 correlation-with-organic figures are from Authoritas' volatility research (published February 19, 2025) - verified directly at the primary source.
About the author
Zarko Zivkovic is the founder of CoreAEX, building technical SEO, AEO, and AI-visibility systems for B2B SaaS companies. Connect on LinkedIn.