Searching for information has become the most common way Americans use AI. According to an AP-NORC survey, 60% of U.S. adults—and nearly three-quarters of adults under 30—say they use AI to find answers to their questions.
Whether they're asking "Should I lock in my mortgage rate?" "Is this rash something I should worry about?" or "What's the best laptop under $1,000?," millions of people are increasingly turning to ChatGPT, Gemini and Claude instead of a Google search. In fact, earlier this year, Federal Reserve researchers reported that 41% of U.S. workers now use generative AI for work and 50% of Americans use generative AI outside of work, up 26% and up 31% year-over-year, respectively.
Federal Reserve Vice Chair Michael Barr said in February that generative AI is being adopted at an unprecedented pace. "The speed of AI adoption may be much faster than previous general-purpose technologies, boosting productivity growth, but also allowing less time for workers, businesses, and the economy to adapt," he told the New York Association for Business Economics.
As more people rely on AI to make everyday decisions, one question has become increasingly important: Where do those answers actually come from?
The answer isn't as simple as "the internet." Ask ChatGPT, Gemini and Claude the same question and you're likely to get different responses because each model draws from a different mix of sources. When all three were asked, "Who makes the best pickup truck?" ChatGPT recommended the Ram 1500 and cited Cars.com. Claude chose the Ford F-150 without citing a source. Gemini declined to name a winner, responding that the best truck "depends entirely on what you prioritize: maximum towing capacity, daily luxury, or long-term reliability."
25 Million AI Citations Analyzed
While the answers are often instant and remarkably confident, the differences aren't necessarily a sign that one AI is right and another is wrong. They reflect the information each model chooses to retrieve, prioritize and synthesize.
To better understand those differences, Muck Rack's Generative Pulse analyzed more than 25 million citations generated by ChatGPT, Gemini and Claude across thousands of real-world prompts. The analysis found that AI models rely on different sources to build their responses, and the type of question someone asks can dramatically change the information an AI model chooses to surface.
ChatPGT, Claude, and Gemini Don’t Use the Same Sources
One of the clearest findings was that there is no single information ecosystem powering generative AI. ChatGPT cites sources in nearly every response and frequently draws from Wikipedia, news outlets and reference sites. Claude cites less often but relies heavily on scientific research and academic literature. Gemini turns to Reddit far more often, giving community discussions a larger role in many of its answers.
The Type of Question Changes What AI Cites
The research also found that AI changes its sourcing strategy depending on what people ask. If someone asks, "Why are egg prices rising?" or "What's happening with the housing market?" AI is much more likely to rely on recent news coverage. If they ask, "Which robot vacuum should I buy?" or "How do I remove a wine stain?" it's more likely to draw from product reviews, reference sources and corporate sites.
One of the clearest examples comes from "best of" questions. When people ask ChatGPT for recommendations such as the best restaurants, hotels or attractions, it frequently cites Google Maps. In Muck Rack's analysis, ChatGPT cited Google Maps search results at a rate of 188 Maps citations per 1,000 queries. It included an average of seven Google Maps pins each time Maps appeared in a response. Claude cited Google Maps only once across the entire study.
Does AI Search Still Rely on Journalism?
Despite some concern that AI is replacing journalism, the study found that news reporting continues to play a significant role in shaping AI answers. More than one quarter of all citations came from journalistic sources. More than half of those articles had been published within the previous year. Rather than concentrating on a handful of prominent national publications, AI models cited more than 20,000 outlets including local newspapers, trade publications and specialized expert sites reporting on issues like healthcare, travel, personal finance and technology.
The next time AI recommends a financial product, summarizes a breaking news story or explains a medical condition, remember that the answer is shaped by the information it chose to retrieve. Asking another AI assistant, checking the cited sources or reading the original reporting can help provide a fuller picture. As AI becomes a more common way to discover information, understanding where an answer came from may become just as important as the answer itself.
Muck Rack
Methodology
What Is AI Reading? is published by Muck Rack's Generative Pulse team and examines how generative AI systems cite sources when responding to realistic consumer prompts. The May edition analyzed more than 25 million links from responses generated by ChatGPT, Claude and Gemini across 17 industries and multiple query types. Because generative AI systems evolve quickly, citation behavior may shift as models are updated or retrained.
This story was produced by Muck Rack and reviewed and distributed by Stacker.
