AI answer engines decide what to cite through retrieval-augmented generation: the model pairs its trained knowledge with a live search step, pulls passages from an index, then attributes the supporting links that best match the answer it wrote. Citation favours authoritative and community sources and only partly follows Google rankings, so being cited is a distinct layer on top of SEO.
Every B2B marketer now asks the same question: when ChatGPT, Perplexity, or Google answers a buyer in my category, why does it name some sources and not others? It is not random, and it is not pure ranking. There is a mechanism, it is measurable, and it rewards different work than the old playbook. Understanding it is the difference between guessing at AI visibility and engineering it, the foundation of AI search engine optimization and the heart of winning the traffic collapse.
How do AI engines actually decide what to cite?
Through a process called retrieval-augmented generation, or RAG. A language model on its own cannot reliably cite anything, because during training it absorbs patterns from billions of documents without keeping addressable links back to them. So when an engine needs to answer with current, traceable information, it adds a live retrieval step. The system turns your query into a search, pulls the most relevant passages from an index, feeds those passages to the model as context, and the model writes an answer grounded in them. Because it knows which passages it used, it can attribute the supporting links you see, as the research on RAG explains.
The practical consequence is a two-stage gate. First your content has to be retrieved, which means it must be indexed, relevant, and well-structured enough to surface for the query. Only then can it be selected as one of the handful of links the engine cites. For Google AI Overviews, Botify documents that the text is generated by Gemini while the cited links are pulled from Google's organic search index, the same index behind the ten blue links. Google's own guidance confirms there is no special markup: a page just needs to be crawlable, indexed, and snippet-eligible.
Do AI citations follow Google rankings?
Only loosely, and that is the most important thing to understand. Ahrefs analysed AI Overview citations and found that only 37.9% of cited URLs appeared in the top 10 organic results for the query. Roughly 31% came from positions 11 to 100, and another 31% came from beyond the top 100 entirely. The reason is query fan-out: when AI triggers, Google splits the question into many related sub-questions, runs them, and aggregates pages that appear across that wider net. A page that ranks for a specific sub-question can be cited even if it never ranks for the head term.
That does not mean rankings are irrelevant. BrightEdge finds that most AI Overviews still cite sites ranking in the top 35, with a strong bias toward positions 1 to 12, so strong SEO remains the eligibility floor. The nuance matters for strategy: ranking gets you into the retrieval pool, but covering the full fan-out of buyer sub-questions is what earns the citation. Optimising for one high-volume keyword is no longer enough, which is the practical core of AEO versus SEO.
Key Takeaway
Citation is a retrieval problem, not just a ranking problem. Rank well enough to be eligible, then cover the cluster of sub-questions an engine explores when it builds an answer. Breadth of relevant coverage beats a single strong keyword.
Which sources do AI engines cite most?
A surprisingly concentrated set. Pew Research found that Wikipedia, YouTube, and Reddit together made up 15% of all sources cited in Google AI summaries. Profound's cross-engine study shows each engine has its own taste: Wikipedia is ChatGPT's single most-cited domain at 7.8% of all citations and nearly half of its top-ten sources, while Reddit leads on Perplexity at 6.6% and on Google AI Overviews at 2.2%. Ahrefs reports that YouTube is now the most-cited domain in AI Overviews, and that brand mentions in video titles and transcripts were the strongest single factor correlating with AI visibility across 75,000 brands.
| Engine | Citation philosophy | Signature source |
| ChatGPT | Authoritative knowledge bases and established media | Wikipedia (7.8% of citations) |
| Perplexity | Community and peer-to-peer discussion | Reddit (6.6% of citations) |
| Google AI Overviews | A blend of professional content and social platforms | YouTube and Reddit |
Sources: Profound, 2025; Semrush, 2025.
These patterns also shift. Semrush found that in late 2025 ChatGPT abruptly cut its citations of Reddit and Wikipedia to avoid over-relying on a few manipulable sources, while PRNewswire, Forbes, and Medium gained share. The lesson for B2B is blunt: a large part of whether you are cited depends on your presence on third-party surfaces, not just your own site. Your YouTube explainers, your LinkedIn posts, your reviews on G2 and Capterra, and your coverage in earned media all feed the corpus these engines mine.
Want to see which sources AI engines cite for your category, and whether you are one of them?
Run the free AI Visibility CheckHow do ChatGPT, Perplexity and Google AI Overviews differ?
They share the retrieval pattern but differ in how they choose and show sources. ChatGPT search decides when to run a web search and returns answers with embedded link cards. Semrush found ChatGPT cites pages ranking in position 21 or beyond about 90% of the time, meaning it reaches deep into the long tail rather than mirroring page one. Perplexity is the most transparent: it places numbered sources beside each paragraph and lets users choose whether to search the open web, their own uploaded files, or both. Google AI Overviews, as covered in our piece on how AI Overviews work, frame citations as supporting links to a Gemini-written summary drawn from the search index.
For a B2B brand this means there is no single AI to optimise for. ChatGPT rewards authoritative, well-structured explainer content; Perplexity rewards community presence and clean, extractable pages; Google rewards the same fundamentals plus video and fan-out coverage. The common thread is that each engine assembles an answer from many sources and attributes only a few, so the goal is to be the clearest, most citable source for specific sub-questions, not to win one ranking.
Are AI citations even accurate?
Often not, and this is a risk you have to manage. The Tow Center for Digital Journalism tested eight AI search engines and concluded they were all bad at citing sources, commonly citing the wrong article, misattributing quotes to the wrong outlet, and linking to outdated coverage. Retrieval-augmented generation improves factual accuracy over pure model recall, but it is not a fix: if the retriever surfaces weak or competitor-authored documents, the model will faithfully summarise them into a confident, well-cited, wrong answer.
For B2B vendors the stakes are concrete. A misattributed security certification, an outdated pricing model, or a competitor's framing presented as fact can shape a shortlist before a buyer ever reaches your site. The defence is source hygiene: make sure authoritative, current, clearly structured content about your products and category exists and is indexable, so that when retrieval happens the engine has good inputs to draw from. This is also why structured data for AI citation matters, since machine-readable claims are easier for a model to attribute correctly.
Watch Out
If you are not the cited source for your own category, someone else is, often a competitor or a third-party reviewer. AI answers fill the gap with whatever they can retrieve. Silence does not keep you neutral; it hands the narrative to the sources that did show up.
How do you get cited by AI?
You earn citation by being eligible, being comprehensive, and being present where the engines look. None of it is a trick; it is the disciplined construction of citable presence.
Keep the SEO foundations that make you eligible
Crawlable, indexed, snippet-eligible pages with genuine authority are the retrieval floor. If you cannot be retrieved, you cannot be cited. Rankings still get you into the pool.
Cover the fan-out, not just the keyword
Answer the cluster of sub-questions an engine explores: definitions, comparisons, pricing, implementation. Build "X vs Y" and "best tools for Z" pages that AI can lift directly.
Write answer-first for machine extraction
Lead with a clean, quotable answer, name the relevant entities, and use clear headings. Models cite passages, so make the citable passage easy to find and lift.
Earn presence on the sources AI trusts
Get represented on YouTube, Reddit, LinkedIn, review platforms, and earned media. A large share of citations points to third-party surfaces, not your own domain.
Measure Share of Model and citation drift
Track which URLs AI engines cite for your category and watch when citations drift to competitors. Citation analytics turns AI visibility from a guess into a managed metric.
Run as a connected system rather than scattered tactics, this compounds, which is the logic behind answer engine optimization and an AEO content system. And the upside is real: Semrush found that an AI-search visit converts 4.4 times better than an average organic visit, so even a smaller volume of cited, high-intent traffic can outweigh the clicks you lost. For the specific tactics, see our guide on how to get cited by ChatGPT.
Find out whether AI is citing you or your competitors
Knowing how citation works is step one. The free AI Visibility Check shows which sources the major answer engines cite for the questions your buyers ask, where competitors are winning the citation, and what to build to earn it yourself.
Run the free AI Visibility CheckFrequently asked questions about how LLMs decide what to cite
How does ChatGPT decide which sources to cite?
ChatGPT search runs a live web retrieval step, pulls relevant passages, writes an answer grounded in them, and attributes the supporting links that best match. It reaches deep into the long tail, citing pages ranked 21 or beyond about 90% of the time, and historically leans on authoritative sources like Wikipedia.
Do AI citations come from Google rankings?
Only partly. Ahrefs found just 37.9% of AI Overview citations appear in the top 10 organic results; the rest come from deeper rankings surfaced through query fan-out. Rankings make you eligible, but covering related sub-questions is what earns the citation.
What sources do AI engines cite most?
A concentrated set. Wikipedia, YouTube, and Reddit make up about 15% of AI-summary sources. ChatGPT favours Wikipedia, Perplexity favours Reddit, and Google AI Overviews blend professional content with YouTube and community sites. Presence on these third-party surfaces strongly affects whether you are cited.
Are ChatGPT citations real or fake?
Sometimes fabricated or wrong. The Tow Center found that all eight AI search engines it tested frequently cited the wrong article, misattributed quotes, or linked to outdated coverage. Always verify a citation before relying on it, and keep accurate, indexable content so engines have good sources to draw from.
How many sources does an AI answer cite?
Usually several. Pew found 88% of AI summaries cite three or more sources and only 1% cite a single source, with a median length of 67 words. Each citation often supports just one sentence, so being the best source for a specific point matters more than covering the whole topic.
How do I get my brand cited by AI?
Stay eligible with strong SEO foundations, cover the full fan-out of buyer sub-questions, write answer-first for machine extraction, earn presence on cited third-party surfaces like YouTube and Reddit, and measure Share of Model. Start with a free AI Visibility Check.
Resources
- Pew Research Center: Do people click links in Google AI summaries? (2025)
- Ahrefs: 38% of AI Overview citations pull from the top 10
- Profound: AI platform citation patterns across ChatGPT, Google AI and Perplexity
- Semrush: The most-cited domains in AI, a 3-month study
- Columbia Journalism Review: AI Search Has a Citation Problem
- Google Search Central: AI Features and Your Website
- Perplexity Help Center: Internal Knowledge Search and source selection
