How do AI engines decide what to cite?

AI engines decide what to cite through a three-stage process: retrieval pulls candidate sources relevant to the query, ranking scores those sources for relevance and trust, and generation decides which retrieved passages the model actually quotes or paraphrases. Each platform weighs these stages differently, which is why ChatGPT, Perplexity, and Google AI Overviews often cite almost entirely different sources for the same question.

TL;DR: Profound's analysis of more than 680 million citations found that only 11% of domains get cited by both ChatGPT and Perplexity for matching prompts — each platform runs on different retrieval logic. What you can influence: structured content, quotable definitions, original data. What you cannot influence: a platform's retrieval algorithm, its training data cutoff, and which competitors it has already indexed. Spend your effort on the first list.

Every AI engine that cites sources runs some version of the same pipeline: retrieve, rank, generate. Retrieval pulls a set of candidate documents that might answer the query, either from a live web index, a pre-built vector store, or the model's training data. Ranking scores those candidates for relevance, authority, and freshness. Generation is the model deciding, sentence by sentence, which retrieved passage to quote, paraphrase, or ignore entirely.

Where engines diverge is which part of that pipeline they lean on. Perplexity performs real-time web retrieval on most queries, so new content can surface in its citations within hours of being indexed. ChatGPT blends a static training corpus with a Bing-powered retrieval layer for current-events and commercial queries, so its citation behavior shifts depending on whether a query needs live data. Google AI Overviews is grounded directly in the Google Search index, which tends to favor pages that already rank well organically.

The practical consequence: a page that earns Perplexity citations will not automatically show up in ChatGPT's, and vice versa. Treating "AI search visibility" as one undifferentiated target is the most common mistake teams make — each engine needs its own evaluation, not a single generic fix.

Five signals show up consistently across the research on AI citation behavior, in roughly descending order of leverage.

Entity recognition is the process by which a language model identifies a named thing in text — a company, a product, a person, a concept — and resolves it to a single, disambiguated identity rather than treating it as an arbitrary string of words. It matters for citation because a model can only credit you correctly if it can tell, reliably, that "Boring Magic," "boringmagic.ai," and "the AI marketing automation agency mentioned on this page" all refer to the same entity.

Inconsistent naming breaks this. If a company is referred to as "Acme," "Acme Inc," and "Acme Marketing" across different pages, a model may treat these as three separate, weaker signals instead of one strong one. The fix is unglamorous: use the same name, the same way, everywhere — your site, your schema markup, your social profiles, your directory listings — so every mention reinforces the same entity instead of fragmenting it.

This is also why citation consistency, the traditional local-SEO discipline of matching your name, address, and category across listings, has quietly become relevant to AI visibility too. It is the same problem in a new context: help the model resolve who you are with zero ambiguity.

The pattern is consistent: everything in the high-control column is something you do once, on your own content, with no platform dependency. Everything in the low-control column is something you can monitor but not move — a platform's retrieval logic changes on its own schedule, and Profound's research shows 40 to 60% of cited domains rotate out monthly across major platforms, regardless of what you do. Spending a quarter reverse-engineering one platform's algorithm is the equivalent of chasing a single Google update in 2012 — directional work pays off across all of them, and platform-specific tricks decay within weeks.

Pull the controllable factors together and a citation-ready page has a consistent shape: it opens with a direct, quotable answer; it names itself the same way it is named everywhere else; it backs claims with a specific number rather than an adjective; and it is marked up so a machine can parse its structure as easily as a human can read it.

None of this requires guessing what a specific model wants this month. It requires writing clearly, structuring honestly, and being consistent about who you are across the internet — which is also, not coincidentally, what makes content good for human readers.

Before scoping any work, the AI search visibility checker shows you whether ChatGPT, Perplexity, and Google AI Overviews are already citing your brand — free, in minutes. For the broader discipline this sits inside, our GEO and AEO guide covers how the two relate to traditional SEO and where to start. If you want a working system after that, that is what our AI search visibility service installs — book a 30-min scope call and we will bring a citation audit of your current pages to it.