How AI Decides Which Sources to Cite

Most teams treat AI citations like a black box, so they end up optimizing for a mechanism they have never actually watched run. We have watched it run 37,230 times.

How AI models decide which sources to cite

AI models pick sources in two stages. First the engine decides whether to fetch the live web at all. If it does, it runs a search, pulls a pool of candidate pages, and re-ranks them on authority, freshness, and how cleanly a passage answers the exact question. The answer then cites the handful of pages it leaned on. If the engine answers from training data instead, it often cites nothing.

That two-stage split is the whole game, and it explains almost everything buyers find confusing about AI citations, including why the same brand shows up on one engine and vanishes on another.

The numbers below come from the CITE Index by Cite Solutions, our corpus of 37,230 real AI answers collected daily across ChatGPT, Gemini, and Google AI Mode over 10 consumer verticals from May 19 to June 12, 2026. Those answers cited 5,239 distinct domains. It is one of the few datasets built from what the engines actually returned to real queries, not what a vendor guesses they do.

Retrieval vs generation: the split that decides everything

There are two ways an AI model can produce an answer.

It can retrieve. The engine runs a live web search, fetches a set of pages, reads them, and writes an answer grounded in what it just read. Because the answer is built from fetched documents, the engine can point at them. That pointing is the citation.

Or it can generate. The engine answers straight from its training data, the patterns it absorbed during pretraining. There is no live document to point at, so there is often nothing to cite. The model knows the shape of the answer without knowing a current URL for it.

Every engine sits somewhere on that spectrum, and you can read its position directly off how often it cites anything at all.

Engine	Answers that cite any source	Avg sources per answer	What this tells you
Google AI Mode	97.9%	5.2	Retrieval-grounded by default. Almost always pulls live pages.
ChatGPT	87.4%	5.2	Retrieves often, but answers a real share from memory.
Gemini	74.0%	4.4	The least source-transparent of the three. One answer in four shows nothing.

Google AI Mode cites a source in 97.9% of its answers, averaging 5.2 sources each. ChatGPT cites in 87.4%, also at 5.2 sources when it does. Gemini cites in just 74.0%, at a lower 4.4 sources. So on Gemini, roughly one buyer answer in four arrives with no visible source at all, which means there was no slot for your brand to win even if you were the obvious answer.

The lesson is not that one engine is better. It is that retrieval-grounded engines give you more shots on goal. On Google AI Mode, nearly every answer is a competition for one of about five citation slots. On Gemini, a quarter of answers never open that competition. If you are deciding where to spend first, spend where the doors are open most often.

The operator implication: your citation strategy is really a retrieval strategy. You are not trying to be remembered by a model. You are trying to be fetched and quoted at the moment the question is asked. Those are different jobs, and the second one is the one you can actually influence.

What gets a source pulled into the pool

Once an engine decides to retrieve, it runs a search and assembles a candidate pool, then re-ranks it. Four signals do most of the sorting. None of them are mysterious once you have seen enough answers.

Authority comes first. The engine leans toward domains it already treats as trustworthy for the topic. This is partly inherited from training (it knows what a reliable source looks like) and partly live (it weighs the domain's standing on the open web). A page on a domain nobody references rarely survives the re-rank.

Freshness matters more than most B2B teams expect. For anything that changes (pricing, product comparisons, "best X for Y" lists), the engine prefers recently updated pages. A definitive guide from 2023 loses to a thinner page updated last month, because the engine is trying not to quote stale facts back at the user.

Then there is extractability. The engine is not citing your page. It is citing a passage on your page. It wants a clean, self-contained chunk that answers the question in a few sentences, with the claim and the context in the same place. Walls of marketing copy with the answer buried in paragraph nine get skipped for a page that states it plainly up top. We have written before about why passages beat pages in retrieval, and the CITE data keeps confirming it.

Last is corroboration. Engines are more comfortable citing a claim that shows up in more than one place. A number that appears only on your own site reads as a marketing assertion. The same number echoed in a review site, a forum thread, and a news write-up reads as a fact. The engine treats third-party agreement as a confidence signal, and it cites with more confidence.

The operator implication: you cannot bolt these on after the fact. A page that is authoritative, current, cleanly structured, and corroborated elsewhere is a page that gets pulled. Miss any one and you fall down the re-rank, no matter how good the prose is.

The source mix: why your own site is never enough

Here is the part that breaks the most strategies. The citation pool is long-tail and it is not dominated by brand sites.

Across the 37,230 answers, the engines cited 5,239 distinct domains. Even the most-cited domains each take only a small slice. There is no single site that "owns" AI answers. The pool is wide, and it rewards a spread of source types rather than one hero page.

Two source types show up far more than their size would suggest. Reddit appears in 21.9% of all answers, roughly one in five. YouTube appears in 8.4%. Community and video are not side channels here. They are load-bearing parts of how these engines build answers, because forum threads and video transcripts are dense with the plain-language, corroborated, recently-posted passages retrieval loves.

In our India consumer dataset the recurring source types were app stores (Google Play), YouTube, Reddit, large news publishers like Times of India, and brand-owned domains. The specific domains are vertical-specific and will not match a B2B SaaS query in the US. The portable lesson is the mix, not the names: community, video, news, and owned. Buyers see all four types blended into a single answer, and the engine pulls the best passage from whichever type has it.

That mix is exactly why an owned-only strategy underperforms. If you only invest in your own pages, you are competing for a minority of the citation slots and ceding the community, video, and news slots to whoever shows up there. The brands that get cited often are present across the mix: their own pages are clean and current, and they also appear in the threads, the videos, and the coverage the engine reaches for. We see the same source-type spread when we break down where AI search actually cites brands across verticals.

This is also the honest answer to "does Reddit help." At 21.9% of all answers, a relevant Reddit thread is one of the highest-probability places for your brand to surface inside an AI answer. Not because you can spam it (you cannot, and it backfires), but because genuine presence in the right threads puts you in the corpus the engine reads from.

Why the same brand gets cited by one engine and not another

Put the two halves together and the most common buyer complaint explains itself.

Engine A retrieves for a given question and Engine B answers from memory. You get cited on A and ignored on B, with no change on your end. Or both retrieve, but they run different searches, assemble different candidate pools, and re-rank on slightly different weightings. A Reddit thread that surfaces on one engine never enters the pool on another. This is not your site being inconsistent. It is two different retrieval systems looking at the same web through different lenses.

That is why we tell teams to stop asking "am I cited" and start asking "cited where, for which questions, on which engine." A single visibility number hides the mechanism. The useful view is per-engine, per-question, which is the whole point of measuring share of voice across AI search rather than a single vanity score. When you see it that way, the gaps become a worklist instead of a mystery.

The operator implication: do not optimize for an average. Find the engines and questions where you are absent, check whether the engine even retrieves for those questions, and then go fix the specific source type that is winning the slot you want.

What to actually do with this

The mechanism gives you a clear order of operations. First, make your own pages retrievable: state the answer plainly near the top, keep the facts current, and structure passages so a single chunk stands on its own. Second, get corroborated off-site, because a claim that lives only on your domain reads as marketing and a claim echoed elsewhere reads as fact. Third, show up in the community and video layer that Reddit's 21.9% and YouTube's 8.4% prove the engines lean on. Fourth, measure per engine and per question, because the averages lie.

None of this is a one-time push. Freshness decays, threads move, and the engines re-weight constantly. The teams that stay cited treat it as a standing program, which is most of what a GEO agency is actually for. If you want the underlying framework we run this against, it is laid out in our CITE framework.

FAQ

How do AI models decide which sources to cite when answering buyer questions?

In two stages. The engine first decides whether to fetch the live web or answer from training data. If it fetches, it runs a search, builds a pool of candidate pages, and re-ranks them on authority, freshness, how cleanly a passage answers the question, and whether the claim is corroborated elsewhere. It then cites the few pages it leaned on. Answers built from memory often cite nothing.

Why does my brand get cited by one AI engine but not another?

Because the engines retrieve differently. In the CITE Index, Google AI Mode cited a source in 97.9% of answers while Gemini cited in only 74.0%, so Gemini simply opens fewer citation slots. Even when both retrieve, they run different searches and re-rank with different weightings, so a source that enters one engine's pool may never reach another's. Measure per engine, not on average.

Does getting cited on Reddit help AI visibility?

Yes, materially. Reddit appears in 21.9% of all answers in the CITE Index, roughly one in five. That makes a relevant thread one of the higher-probability places for your brand to surface inside an AI answer. It works through genuine presence in the right discussions, not promotion. Spammed threads get ignored or backfire, but real, corroborated mentions feed the corpus engines read from.

How many sources does a typical AI answer cite?

When an answer cites at all, it leans on about four to five sources. In the CITE Index, Google AI Mode and ChatGPT both averaged 5.2 sources per cited answer, and Gemini averaged 4.4. So you are usually competing for one of roughly five slots, and those slots span source types: community, video, news, and brand-owned pages mixed into a single answer.

Is being cited the same as ranking on Google?

No. AI engines do not rank pages for the user, they extract passages and ground an answer in them. A page can rank well on Google and still never get pulled into an AI answer if its passages are hard to extract, stale, or uncorroborated. The reverse also happens. Treat AI citation as a separate retrieval problem with its own signals, and run an AI visibility audit to see where the two diverge.