• AI visibility

Guide · AI search

Which types of websites do AI search engines rely on most?

AI search engines rely most on established third parties: official bodies, reference sites like Wikipedia, news outlets and specialist review publishers. Which type leads depends on the topic and on the engine. Video, social media and forums matter far more in Google’s AI Overviews than in ChatGPT.

By Underneath Team Guide Updated

The short version

  1. On politics, health and environment questions, government websites took 33.0% of citations among the 25 most-cited domains, in an audit of four AI engines (opens in a new tab).
  2. For 100 health questions put to ChatGPT, institutions such as hospitals, government bodies and journals supplied 75.7% of cited sources, one 2026 study found (opens in a new tab).
  3. For product questions, editorial and product-review sites made up 56.7% of the domains ChatGPT displayed, according to an audit of AI product recommendations (opens in a new tab).
  4. Google’s AI leans on forums far more than ChatGPT: our Reddit study found a Reddit thread in 17.9% of AI Overviews and in no ChatGPT answers.

Which kinds of websites do AI search engines cite most?

Mostly established third parties: official bodies, reference works, news outlets and specialist publishers. Zhang Kai and colleagues (opens in a new tab) analyzed a public dataset of 602 controlled prompts on ChatGPT, Google’s AI Overview and Gemini, and Perplexity. Official, news and vertical (specialist industry) sources accounted for 79.12% to 87.52% of citations across platforms.

Two caveats apply. Researchers wrote those prompts; they were not real user questions. The authors also warn that the source-type labels in the dataset are noisy.

Beyond that broad picture, the mix shifts with the subject. Public-interest topics pull engines toward official sources. Buying questions pull them toward review publishers. Google’s AI features add a large share of video and social media.

Do government and institutional sites really dominate?

On public-interest topics, yes, but not everywhere. Allaham and Diakopoulos (opens in a new tab) put 712 real user questions on politics, health and the environment to ChatGPT, Copilot, Gemini and Perplexity. Among the 25 most-cited domains, government websites took 33.0% of citations and Wikipedia alone took 15.2%. Social media, including Reddit and YouTube, took 8.8%. The same audit also checked how many cited sources were AI-generated.

Health shows the same pattern. Jacques and colleagues (opens in a new tab) coded the sources ChatGPT cited for 100 consumer health questions on one day in January 2026. Institutions with built-in authority supplied 75.7% of citations, while commercial health sites supplied 12.4% and professional or practice websites 11.9%. Ten organizations, led by Wikipedia, Mayo Clinic and Cleveland Clinic, accounted for 52.8% of all citations.

This is not a general rule. In a comparison of 11,500 user queries (opens in a new tab), traditional Google Search was more likely than AI Overviews or Gemini to retrieve government and education sites. Official sources lead when the topic itself is official.

How much do AI engines rely on Wikipedia, YouTube and Reddit?

A great deal, but it varies sharply by engine. Xu and colleagues (opens in a new tab) tracked Google AI Overviews for 40 days in March and April 2026. The five most-cited sites were youtube.com (5.49%), en.wikipedia.org (4.39%), facebook.com (3.68%), instagram.com (3.65%) and usatoday.com (2.80%). Sports alone accounted for 33.6% of all AI Overview references, which may lift the share of video and social sources.

Our own data agrees on YouTube. In our AI Overview citation study, YouTube was cited in 64.2% of AI Overviews. In our Reddit study, Perplexity cited a Reddit thread in 12.5% of 80 buyer questions, Gemini in 2.5%, and ChatGPT and Claude in none.

A study of 11,000 real search questions (opens in a new tab) found the same split. A search-enabled OpenAI model took 0.1% of its citations from social platforms, against 8.5% for AI Overviews and 13.4% for regular Google results. The OpenAI model drew 27.3% of its citations from encyclopedias and reference sites.

What do AI engines cite for product and buying questions?

Independent review and editorial publishers, more than brand or retail sites. Uberti-Bona Marin and colleagues (opens in a new tab) put 117 physical-product questions to ChatGPT, Gemini and Google from the Netherlands in September 2026. Editorial and product-review sources made up 56.7% of the domains ChatGPT displayed and 45.2% in Gemini.

Google’s AI Overviews spread wider. There, editorial and review sources were 25.4% of displayed domains, user-generated and community sources 22.2%, retailers 17.8% and manufacturers and brands 16.1%. The authors labeled source types with an AI model and call these results exploratory.

An August 2025 study by Chen and colleagues (opens in a new tab) points the same way. For US consumer electronics questions, AI search drew 92.1% of its sources from earned media, meaning independent reviews and publications. Google’s regular results gave brand sites 32.9% and social sites 15.4%. That study used developer versions of the engines, not the consumer apps, and thanked a GEO vendor for support.

Do different AI engines favor different types of sites?

Yes, enough to change where a brand should invest. ChatGPT leans toward reference works, news agencies and review publishers. Google’s AI Overviews mix in more video, social media and forums. Engines also differ in how many sources they cite per answer.

News citations show how concentrated one engine can be. Yang (opens in a new tab) analyzed over 366,000 citations from an AI search comparison site in 2025; 9% pointed to news sources. For OpenAI models, the 20 most-cited news outlets took 67.3% of citations in that news sample, led by Reuters and AP News. Google and Perplexity models spread their news citations more widely.

A vendor dataset gives a commercial view. Kumar (opens in a new tab), whose company sells AI visibility tracking, analyzed 149,912 citations for brand prompts. Outside company websites, YouTube led at 4.2%, ahead of tech and business media at 3.8%, Reddit and other forums at 3.3% and Wikipedia at 2.6%.

Do AI engines rely on a few big sites or many small ones?

Both: a short list of repeat sources plus a very long tail. In the Allaham and Diakopoulos audit, the 25 most-cited domains took 23.8% of all citations. Most other domains received very few citations: 59.1% were cited only once.

AI Overviews look similar. In the Xu study, 56.2% of cited sites appeared exactly once in 40 days. The five most-cited sites took 20.0% of AI Overview citations, against 47.0% of links on Google’s first page of results. That long tail is why smaller websites can still get cited.

What should you do about it?

Match your presence to the source types engines use for your kind of question.

  1. Check which sources the engines cite for your buyers’ real questions. The mix differs by topic and by engine, so measure rather than assume.
  2. For buying questions, earn coverage in independent review and editorial publications. They were the largest displayed source type in ChatGPT’s product answers.
  3. If your buyers use Google, invest in video. YouTube was the most-cited site in AI Overviews in the Xu study.
  4. Keep reference entries, such as Wikipedia, accurate where you legitimately qualify for one.
  5. On regulated or public-interest topics, align your claims with official sources, because engines lean on them.

Our generative engine optimization service starts with this kind of source mapping.

What does the research not tell us yet?

The research shows what engines cite, not why, and few studies cover business-to-business questions.

  • Most studies are snapshots from one period, and engines change their sourcing often.
  • Category labels differ between studies, so “earned”, “official” and “editorial” are not identical buckets.
  • Several large datasets come from vendors or used researcher-written prompts rather than real users.
  • Being cited is not the same as shaping the answer. One study found AI Overviews cite Reddit and Quora yet draw 22.1% less content from social platforms than from other sources.
  • We found no study linking source types to sales or leads.

Frequently asked questions

Does ChatGPT cite Reddit?

Rarely, in the studies so far. In our Reddit study, ChatGPT cited no Reddit threads across 80 buyer questions, while 17.9% of Google AI Overviews did.

Which website does Google’s AI Overview cite most?

YouTube, in the largest recent study. Xu and colleagues found youtube.com accounted for 5.49% of AI Overview citations, ahead of Wikipedia at 4.39%.

Do AI search engines prefer government websites?

On public-interest topics, yes. Government sites took 33.0% of top-domain citations for politics, health and environment questions in one audit of four engines.

Are brand websites cited for product questions?

Sometimes, but less than reviews. In Google AI Overviews for product questions, manufacturers and brands were 16.1% of displayed domains, behind editorial and review sources.

Sources

Free strategy call

Some questions are easier to answer about your own business.

Bring the one that matters most. On a free 30-minute call we’ll take a first look at it and send you a short written read afterward.

Ask about my businessHow we work