More and more people are having questions answered by ChatGPT, Google AI Overviews or Perplexity instead of doing their own research. But where do these answers actually come from? A new study with 250,000 evaluated answers provides concrete figures for the first time about the sources of AI models.
The search for information on the Internet has changed fundamentally thanks to artificial intelligence. A while ago it worked like this: If you wanted to know something, you typed it into Google, clicked through the results bar and looked for the relevant websites.
Basically, many users still approach the matter in the same way. Instead of a list of links, they now receive AI-supported answers via Google. Many also turn directly to one of the major language models (LLMs) for information.
Instead of comparing results themselves, users are served ready-made answers, often with a few source links in the side. Whether these are correct and how they come about is usually unclear.
Study shows: These sources use AI models for their answers
To provide clarity as to which sources AI models use for their answers, the software company blinq conducted a study. To do this, between January 1 and June 30, 2026, it evaluated daily responses to around 2,000 specified prompts that went to the leading LLMs ChatGPT, Google AI Overviews and Perplexity. A total of 250,000 responses were collected and evaluated.
The central result: Around two thirds of all sources cited by the three systems (66.2 percent) come from company websites, i.e. company blogs, official websites and landing pages.
A look at the most frequently cited websites shows that YouTube (15.18 percent) and Reddit (14.55 percent) lead by a significant margin. This is closely followed by Wikipedia (9.05 percent) and LinkedIn (7.12 percent).
When it comes to German news sites, Welt is ahead with 17 percent, ahead of Handelsblatt (12.2 percent) and Chip (8.9 percent).
Corporate websites therefore provide the largest share of content. However, when it comes to individual addresses that are used particularly often, platforms with real user discussions and reviews make the difference.
Why ChatGPT cites sources other than Google AI Overviews
The study breaks down its results separately. The differences between the three language models are clear. ChatGPT is particularly based on Reddit, Wikipedia, Welt, LinkedIn, Google and BILD. However, highlighting the two Axel Springer media should not be a coincidence. The publisher maintains a highly controversial partnership with OpenAI.
Google AI Overviews, on the other hand, cites user generated content (UGC) noticeably often: YouTube is in first place here, but does not appear in the top sources at ChatGPT. Google also ranks significantly higher for AI Overviews than for ChatGPT. Instagram, Handelsblatt and OMR reviews also play a role.
Perplexity also relies heavily on YouTube and Reddit, but also uses LinkedIn and Instagram as additional user content sources. There are also Handelsblatt, Chip and Wikipedia for specialist authority as well as OMR and Stern.
3 things to check for every AI response
For anyone who regularly uses artificial intelligence for their internet searches, the study findings have a few practical consequences:
- First: An AI answer is not tested factual knowledge, but rather a summary of what is currently circulating on the internet about a topic. This includes company advertising, Reddit discussions and YouTube videos. Many companies are also aware of how they can influence their own visibility. Precisely because two thirds of the sources come from companies themselves, it is worth having a healthy doubt.
- Second: Reddit and YouTube content features heavily in the answers. This can be helpful because there are many honest reports on the platforms. But it can also mean that individual opinions or specifically placed contributions in an AI response are given more weight than they actually deserve.
- Third: Anyone who receives an AI answer usually does not see at first glance how many different and trustworthy sources actually lie behind it. The study itself indicates that the citation style is changing “highly dynamically”. For example, in February 2026, LinkedIn played practically no role as a source. Today the portal is one of the most important sources for all three systems.
And don’t forget: question!
A few well-known basic rules still apply. When making important decisions, such as health, politics, money or major purchases, it is worth looking at the actually linked sources of an AI answer. Simply reading the summary is often not enough. Classic research can also help to form a well-founded opinion.
If you notice that an answer sounds remarkably positive about a particular product or company, you should note that this information most likely comes directly from the company website itself. For topics on which there are controversial discussions, it may be worthwhile to additionally search for independent test reports or specialist press instead of relying solely on the first AI answer.
Also interesting: