Where AI gets its answers: the sources behind AI recommendations

AI assistants such as ChatGPT, Gemini, Claude and Google AI Overviews build their answers from live web sources (review sites, Reddit threads, YouTube videos, editorial coverage and brand websites) retrieved at the moment you ask. Those citations are heavily concentrated: the top 1% of domains capture 64% of all AI citations, and the top 10% capture 84%, according to industry analysis compiled by ZipTie. When a buyer asks an assistant what to buy, the brands named in the answer are drawn largely from what those retrieved sources say. This guide explains how retrieval works, which types of website dominate AI citations, how the patterns differ by engine, and why the source map worth acting on is the one for your own category.

Key takeaways

• AI assistants answer commercial questions by searching the live web and synthesising what the retrieved sources say, not purely from training data.
• Citations concentrate at the top, but the list churns: 40–60% of AI-cited sources change from month to month, per the Semrush AI Visibility Index.
• Each engine has its own source diet. Google AI Overviews stay close to organic search results; the URLs ChatGPT cites overlap only slightly with Google's top rankings.
• Global "most cited websites" lists collapse at category level: cited domains change significantly by industry, so the useful map is the one for your category, not the internet-wide average.

Where does ChatGPT get its information?

ChatGPT gets its information from two places: its training data, and live web search performed at the moment of the question. For anything commercial ("best accounting software for a small UK charity", "which running shoes last longest") live search is common: an October 2025 Nectiv analysis found commercial-intent prompts trigger web searches in ChatGPT 53.5% of the time, versus 18.7% for informational queries. Gemini, Claude and Google AI Overviews work in the same two layers: a retrieval step, then a synthesis step.

The retrieval step is simple to describe. The assistant turns your question into one or more searches, fetches a set of pages, and writes an answer grounded in what those pages say ("grounding" simply means tying the generated answer to retrieved documents rather than to memory alone), often with the sources cited inline or listed underneath. The brands it can name are largely the brands those retrieved pages name. That is why the question "where does AI get its answers?" is really a question about which websites the engines choose to retrieve, and it is measurable.

Recent content is over-represented in what gets cited. An AirOps study found that 95% of ChatGPT citations come from content published or updated within the last 10 months. That is a pattern in citation data rather than a proven rule about how retrieval ranks pages, but it suggests a page that described your category well in 2023 is unlikely to be doing much work in an answer written today.

What sources do AI assistants cite?

AI assistants cite a recognisable set of source types: review and comparison sites, retailers and marketplaces, Reddit and other forums, YouTube, editorial and press coverage, reference sites, and brand websites. On buying questions, independent third-party pages account for a large share: "best of" list posts alone account for around 43.83% of ChatGPT citations for commercial queries, according to Glen Allsopp's analysis published by Ahrefs. When a buyer asks for a recommendation, the engine is often reading other people's shortlists.

Skavora classifies every citation captured in a scan into a taxonomy of source types. The table below groups that taxonomy into the categories that matter most for buying questions, with what each type typically offers a buyer's question:

Community sources feature more heavily than their size would suggest. A Semrush AI Visibility Index snapshot found Reddit, LinkedIn and YouTube among the top-cited sources across the leading AI engines in October 2025. Presence across several communities is associated with being named more often: an Evertune study of 75,000 brands found that brands present on four or more non-affiliated forums are 2.8x more likely to appear in ChatGPT responses. That is a correlation rather than a demonstrated cause, but it points at where to look.

What are the most cited websites by AI?

The most cited websites by AI are a short list of giants, but the list is unstable, and treating any snapshot as permanent is a mistake. In Google AI Overviews, a Surfer SEO study measured the top five cited domains as youtube.com (23.29%), wikipedia.org (18.41%), google.com (16.38%), reddit.com (9.37%) and linkedin.com (8.80%).

The volatility is the finding most people miss. A three-month Semrush study (230,000 prompts, three engines, 13 weekly snapshots, over 100 million citations) watched ChatGPT's citation of Reddit collapse from roughly 60% of responses in early August 2025 to roughly 10% by mid-September, while Wikipedia fell from about 55% to under 20%. More broadly, the Semrush AI Visibility Index found that 40–60% of AI-cited sources change from month to month.

Two practical conclusions follow. First, any "most cited domains" league table is a photograph of a moving object, useful for orientation, useless as a plan. Second, when the source mix swings, a brand's AI visibility (how often AI assistants mention and recommend it) can swing with it, which is why measurement is worth repeating rather than doing once. (If the term is new to you, start with our explainer on what AI visibility is.)

Do different AI engines cite different sources?

Yes, each AI engine draws on a measurably different pool of sources, so visibility on one engine tells you little about the others. The clearest split is between Google AI Overviews and everything else.

Google AI Overviews stay close to organic search: Authoritas and Optimizely studies found 40–76% of AI Overview citations also appear in the top 10 organic results. Of the four engines, AI Overviews are the one where strong Google rankings are most likely to show up in the answer.

ChatGPT and its peers are far less aligned with Google rankings. An Ahrefs study of 15,000 prompts found only 12% of URLs cited by ChatGPT, Gemini and Copilot appear in Google's top 10 for the same query. Ahrefs also reports that around 80% of citations by large language models (LLMs, the models behind assistants such as ChatGPT, Gemini and Claude) come from pages that don't rank in Google's top 100 at all. Ranking well and being cited are related but clearly not the same thing.

This is the practical case for measuring across engines rather than extrapolating from one. A brand with strong Google rankings can be well represented in Google AI Overviews and still be absent from ChatGPT and Claude, where the review sites and forums being retrieved may never mention it. Comparing the cited sources per engine is how that gap becomes visible, which is also why understanding how AI assistants decide which brands to recommend starts with understanding what they read.

Why do global "most cited" lists tell you little about your category?

Global most-cited lists tell you little because AI engines change their source diet by industry: the domains that dominate the internet-wide average can be nearly irrelevant in your category. The Surfer SEO and Amsive data shows it starkly: in health queries, the top domains cited by Google AI Overviews flip to nih.gov (38.89%), youtube.com (27.91%) and healthline.com (15.06%), a completely different hierarchy from the general-purpose top five.

The same logic runs all the way down to niches. For project management software, the influential sources might be G2, a handful of "best tools" roundups and two subreddits. For garden furniture, they might be retailer category pages and YouTube reviews. Knowing that Wikipedia is heavily cited across the whole internet gives you no outreach target; knowing that one comparison site appears in half of the AI answers about your category gives you a precise one.

This is the difference between reading about AI citations and mapping your own. A category-level source map answers the questions that convert into action: which specific sites do the engines cite when buyers ask about products like mine, which of them mention me, and which mention only my competitors?

How can you see the sources behind your own category?

The direct way to see your category's source map is to ask the engines your category's buying questions and record what they cite. Skavora's free check does exactly this: it researches your brand live, generates 25 realistic buyer questions, and runs them through Gemini with live web grounding (grounding means the answer is written against pages retrieved at that moment), capturing every source cited in every answer. Around 80% of those questions are non-branded: they never mention your brand at all, because those are the ones where a buyer has not yet decided who to ask for. The split, and the reasoning behind it, is set out in how Skavora builds its question set. The questions also never name a platform or any brand other than your own, and the engine answering them is never told which brand is being measured. You can run a free AI visibility check with no signup; the free report includes the full cited-source breakdown, classified by the taxonomy above and ranked by influence.

That ranked source list doubles as an outreach target list. It is the practical loop behind generative engine optimisation (GEO), the practice of earning a place in AI-generated answers: measure which sources AI cites in your category, earn presence on those sources (a listing, a review profile, a genuine community presence, coverage), then re-scan to see whether the answers changed. Given the month-to-month churn in cited sources, one scan is a single reading; re-scanning is how you tell a durable position from a lucky snapshot.

One honest caveat: a scan is a snapshot. Live, grounded answers vary from run to run, and Skavora measures AI outputs rather than controlling them. No tool can guarantee that an assistant will name your brand. What a source map gives you is a direction rather than a guarantee: the best-evidenced list of places where earning coverage is most likely to matter in your category. The £7.99 report extends the same picture across all four engines (Gemini, ChatGPT, Claude and Google AI Overviews) with per-engine source ecosystems and the measured overlap between them, which is where the engine-by-engine differences described above become visible for your own brand.

Further reading: for the step-by-step version (how to build the question set, count mentions and classify sources yourself), see our guide on how to measure your AI visibility step by step.