Ask any answer engine a question in your category and it produces one confident paragraph built from somewhere — but most people never ask where. This guide is a map of that somewhere. It starts from the distinction that governs everything downstream: the two data sources behind any AI answer are the model's training data (the baked-in memory it was built from) and live retrieval (the surfaces it fetches at answer time), and only the second is a game you can play this quarter. From there it catalogs what the retrieval layer actually reaches for — the handful of source types that recur at the top of nearly every large-scale citation study (community forums led by Reddit, reference pages led by Wikipedia, video led by YouTube, professional networks led by LinkedIn, plus news, review sites, and product pages), why each engine draws from a largely different set so there is no single list to optimize against, and why community and video became first-class sources rather than afterthoughts. Then it turns from surfaces to signals: within the pool of pages an engine could pull from, what actually gets one selected — freshness, corroboration across independent surfaces, real authority markers, and a passage that answers the question cleanly enough to lift. It closes on the uncomfortable structural fact a creator has to plan around: you cannot be the source an engine cites on a surface you never publish to, so AI-search sourcing is as much a coverage problem — being legibly present across every source type the studies name — as it is a writing problem.
An answer engine hands you one confident paragraph and, if you are lucky, a few footnotes. The paragraph came from somewhere — but the somewhere is treated as a black box by almost everyone reading it, and by most of the advice written about it. This guide opens the box. It is a map of the data sources behind AI search: what an engine like ChatGPT, Perplexity, Google's AI Overviews and AI Mode, or Gemini actually reaches for when it answers a question in your category, and what makes it reach for one source over another. If generative engine optimization is the work of getting cited, this is the terrain map you need before you start — because you cannot optimize toward sources you have not identified. The companion piece on the craft of getting quoted is generative engine optimization and AI-answer citations; this page is the layer beneath it, the where rather than the how.
Everything downstream depends on a single distinction, so it is worth being precise about it. Every AI answer draws on two fundamentally different kinds of data source. The first is training data: the enormous corpus of text the model was built from, which becomes its baked-in memory. It is why a model can talk fluently about your industry without looking anything up — and it has a cutoff date, is frozen until the next model version, and is effectively impossible for you to influence directly at the scale that would matter. The second is live retrieval, often called grounding: at the moment you ask, the engine fetches current sources — from its search index, the open web, or a licensed data feed — and composes the answer from what it just pulled. This is the layer built on retrieval-augmented generation, and it is the entire reason a page you publish today can be cited tomorrow.
The practical takeaway is blunt. Training data shapes what the model knows in general and how it talks; retrieval decides which specific pages get named for a specific question, and it re-runs for every query against the live web. You spend your effort on retrieval, because it is the source layer that reflects what exists now rather than what was true at training time, and it is where citations are actually assigned. When someone asks 'how do I get into the model's training data,' the honest answer is that you mostly do not, and you should not want to — the retrieval layer is faster, measurable, and yours to compete in. The rest of this guide is about what that layer reads.
Over the past year a run of large-scale studies has analyzed AI citations at serious volume — hundreds of millions of them across the major engines — and the encouraging thing is how consistently the same source types surface at the top. The retrieval layer is not reaching for a random slice of the web. It reaches, again and again, for a recognizable handful of source categories.
Reddit is the standout, ranking first or near it in almost every multi-engine study, and its dominance is not an accident. Google signed a content-licensing deal with Reddit in February 2024, reported at roughly sixty million dollars a year, giving it continuous access to Reddit's data API and regular data transfers — so the single most-cited source across AI search is one an engine is paying for structured access to. The deeper reason forums win is that they capture something a polished article cannot: real people comparing real options in their own words, which reads to an engine as authentic consensus. The strategy implications of this are worked through in Reddit visibility in AI search.
Wikipedia is the other pillar, and it is unusual in being strong on both ChatGPT and Google's AI Overviews when almost nothing else is. It functions as the engines' default definitional and entity source — the neutral backbone an answer is built around before it layers on more specific citations. What Wikipedia's traffic patterns reveal about the click economy of AI answers is examined in what Wikipedia reveals about AI Overviews and web traffic.
YouTube is among the most-cited and fastest-rising source types, especially for how-to, demonstration, and product queries, because a video with a real transcript is a rich, quotable source an engine can read. LinkedIn rounds out the top tier, and notably the citations skew toward individual profiles and posts rather than company pages — the network is read as a source of practitioner expertise, not corporate copy. The three-surface play across product pages, community, and video is laid out in earning AI citations across product pages, Reddit, and YouTube.
Below the top four sit the major news outlets (Forbes, Reuters, Business Insider, the wire services), review and comparison sites, and retail/product pages — Amazon among them. A large study of AI answers found that listicles, articles, and product pages together account for more than half of what gets cited, which tells you the format an engine prefers as much as the domain: structured, comparative, and specific.
Here is the finding that quietly wrecks the tidy 'top 50 domains' listicles: the engines do not share a source set. Studies find that the overlap between the domains ChatGPT cites and the domains Perplexity cites is small — in one large analysis, only around a tenth. Each engine runs its own retrieval stack, its own index or search partner, and its own licensing arrangements, so they are reading largely separate slices of the web. Wikipedia is strong on ChatGPT and AI Overviews; Reddit and YouTube are structurally central to AI Overviews and Perplexity but comparatively weak on ChatGPT. The consequence for your planning is that being a source in one engine tells you almost nothing about the others, and a coverage strategy has to treat each engine as its own ecosystem — a point developed in AI visibility across the major AI assistants. It also means the concentration is real but not absolute: the top sources capture a large chunk of citations, yet no single domain typically commands more than a low single-digit share of the total, so the long tail is genuinely alive.
The most consequential shift in the source landscape is that surfaces once treated as pure marketing are now surfaces engines quote from directly. Research in 2026 found Google's AI Overviews citing social platforms at enormous scale — Facebook posts referenced millions of times, and a meaningful fraction of U.S. searches now surfacing a social post inside the AI answer itself (the underlying finding is here). Pair that with YouTube's position as a top and fast-rising source, and the picture inverts the usual creator assumption. Your video library and your social feed are not the promotional layer that drives people toward the 'real' citable asset on your website; they are themselves among the data sources an engine pulls the answer from. The condition is legibility: a video is a source only if it has a real transcript and captions an engine can read, a title and description that state the answer, and a claim consistent with what you say elsewhere. A wall of pixels with no readable text is invisible to the retrieval layer no matter how good the video is.
Identifying the surfaces is half the map. The other half is understanding that being in the eligible pool is necessary but not sufficient — from the set of pages that could answer a question, the engine selects, and selection runs on signals. Four matter most, and they are the same whether the source is your blog, a forum post, or a video.
Freshness is the first and most underrated: engines skew toward recent sources, so a page that answered the question well a year ago quietly loses its citation to a competitor who keeps theirs current. A source is not a set-and-forget asset. Corroboration is the heaviest lever: before committing to a claim, an engine tends to cross-check whether independent surfaces agree, because a lone unsupported page is a liability and a claim echoed across a forum, a video, a third-party article, and your own site is safe to repeat. This is why presence across many source types beats one perfect page — the sources reinforce each other. Authority markers are the third: the founding GEO research (presented at the KDD 2024 conference) found that adding cited statistics, quoting named experts, and answering the question directly measurably raised a source's odds of being pulled, while keyword density — the reflex carried over from old SEO — did nothing. And the passage has to be liftable: structured so a discrete chunk answers one question cleanly and can be attributed without ambiguity, backed by schema that mirrors the visible content. The passage-level craft is covered in AI search content optimization and AI search citation optimization; here the point is only that these signals decide selection among sources that already exist.
It is easy to read the domain lists and conclude the game is entirely offsite. It is not. Your own website remains a source an engine retrieves and grounds against, particularly for questions specifically about you, your product, or your category — and it is the one source you fully control. The requirements are the ordinary ones raised to a higher stakes: it has to be crawlable, structured for answer engines rather than only for human skimming, and internally consistent with everything you say elsewhere. Some sites now also publish an llms.txt file to guide how AI systems read them. But the owned surface is one node in the corroboration graph, not the whole graph — its job is to state your canonical facts cleanly so that when an engine cross-checks a claim it found on a forum or a video against your site, the story matches. A site that contradicts your other surfaces does not just fail to help; it actively makes you harder to cite. The full measurement-and-strategy loop around all of this is in AI visibility and GEO.
Put the surfaces and the signals together and a hard conclusion falls out that most GEO advice steps around: you cannot be the source an engine cites on a surface you do not publish to. If the retrieval layer for your category leans on YouTube and you have no readable video, you are not competing for those citations — you are absent from that source pool entirely, and no amount of on-page optimization fixes an absence. If it leans on community threads and you have no presence there, likewise. AI-search sourcing is therefore as much a coverage problem as a writing problem: being legibly present, with consistent facts, across the specific source types the studies name for your category. For a large publisher that is a staffing question. For a creator or small brand it is the real bottleneck, because producing consistent, machine-legible presence across video, social, community, and owned pages — and keeping it fresh, since freshness is a selection signal — is exactly the work that does not scale by hand.
The through-line of this guide is that AI answers are built from a knowable set of source surfaces, and that a creator's binding constraint is coverage — you have to actually exist, legibly and consistently, on the surface types the engines pull from. That is the specific problem Kompozy addresses, and it is a different angle from optimizing any single page. Kompozy is a content generation and multi-platform publishing engine; it does not measure your citations, run your source audit, or influence a model's training data, and this guide has pointed you at the pages that cover the strategy and measurement. What it removes is the production ceiling that makes multi-surface presence impractical for a small team.
Concretely: the surface types the citation studies rank highest map almost one-to-one onto what Kompozy generates and publishes. Video, the fast-rising source, is where the fit is sharpest — a Persona Short is produced with auto-captions and a real transcript, which is precisely the machine-legible video the retrieval layer can read and quote, rather than the pixel wall it skips. The text posts, carousels, and quote graphics fan out across the eight social platforms plus blog and email, so you are actually present on the social and owned surfaces engines cite rather than absent from them. And because a single Persona Brief governs voice and the canonical facts for everything generated, the same claim carries identically across all those surfaces — which is the corroboration signal, made a property of production instead of a discipline you enforce by hand across a dozen assets that drift.
The two selection signals this guide flagged as decisive — freshness and corroboration — are where an automated engine earns its place rather than just saving time. Autopilot keeps the multi-surface set refreshed on a cadence, so your sources stay recent instead of decaying into the version an engine stops preferring; and because output originates from one governed brief and clears quality gates that reject invented statistics and off-brand claims before anything ships, the surfaces reinforce each other instead of contradicting — a fabricated number on one surface is exactly what gets a source dropped from consideration everywhere. The content-repurposing workflow is the mechanism, but the point for AI-search sourcing is coverage and consistency, not reach. Kompozy does not decide what is true or which questions to own — that judgment stays yours. It makes it feasible to be a legible, consistent, current source on every surface an answer engine actually reads, which is the part that otherwise does not scale.
From two distinct places. The first is the model's training data — the large corpus of text it was built from, which becomes its baked-in memory and has a cutoff date. The second is live retrieval, also called grounding: when you ask a question, engines like Perplexity, ChatGPT search, Google's AI Overviews and Gemini fetch current sources from the web, their search index, or licensed data feeds, then write the answer from what they just pulled. Training data shapes what the model knows in general; retrieval decides which specific pages get cited for a given question. The retrieval layer is the one you can actually influence, because it is re-run for every query and reflects what is on the web now, not what was true at training time.
Across the large-scale citation studies published over the past year, the same source types recur at the top: community forums (Reddit consistently ranks first or near it), reference sites (Wikipedia), video (YouTube), professional networks (LinkedIn), and then major news, review, and retail/product pages such as Forbes, Reuters, and Amazon. Two caveats matter. No single domain typically commands more than a low single-digit share of all citations, so the field is concentrated at the top but still wide. And the exact ranking differs sharply by engine — Wikipedia is strong on both ChatGPT and Google's AI Overviews, Reddit and YouTube are structurally central on AI Overviews and Perplexity but far weaker on ChatGPT — so there is no one list to optimize against.
No, and this is one of the most useful things to know. Studies analyzing hundreds of millions of citations find the overlap between the domains ChatGPT cites and the domains Perplexity cites is small — often only around a tenth. Each engine has its own retrieval stack, its own index or search partner, and its own licensing deals, so they read largely separate slices of the web. The practical consequence is that appearing as a source in one engine does not mean you appear in the others, and a coverage strategy has to account for each engine's distinct source ecosystem rather than assuming a single win propagates.
Yes, and increasingly so. Google's AI Overviews cite social platforms at large scale — research in 2026 found Facebook posts cited millions of times and a meaningful share of U.S. searches now surfacing a social post inside the AI answer — and YouTube is among the fastest-rising and most-cited source types, especially for how-to and demonstration queries. Community threads on Reddit are cited so heavily in part because Google licenses Reddit's data directly. The upshot is that a creator's video library and social presence are not just marketing that supports the 'real' work of getting cited; they are themselves among the surfaces an engine pulls answers from, provided they are legible to a machine — real transcripts, captions, and text an engine can read.
Being in the eligible pool is necessary but not sufficient — selection is decided by signals. Freshness matters: engines skew toward recent sources, so a page that answered the question last year loses to one updated this month. Corroboration matters most: engines cross-check whether a claim is echoed across independent surfaces before committing to it, so a fact repeated consistently across your site, a forum, a video, and a third-party article is safer to cite than a lone page. Real authority markers — cited statistics, named experts, quotable specifics rather than adjectives — raise selection odds; keyword density does not. And the passage has to be directly answerable, structured so it can be lifted and attributed cleanly.
AI search data sources fall into two layers: the model's training data — its baked-in memory, fixed at a cutoff — and live retrieval, the surfaces an engine fetches at answer time. Retrieval is where you can compete. Large studies consistently find community forums (Reddit), reference pages (Wikipedia), video (YouTube), and professional networks (LinkedIn) among the most-cited sources, with each engine drawing from a largely different set. Within that pool, engines select the sources that are fresh, corroborated across independent surfaces, and answer the question cleanly enough to quote.
Get started → · ← All guides · Compare Kompozy vs other tools