A new study from researchers at Ruhr University in Bochum, Germany, and the Max Planck Institute for Software Systems sheds light on the significant differences between traditional search engines like Google and the newer wave of AI-powered generative search engines. The research, presented in the pre-print paper “Characterizing Web Search in The Age of Generative AI,” systematically compares link-based search results to those generated by AI systems, including Google’s AI Overviews, Gemini-2.5-Flash, and OpenAI’s GPT-4o in both its standard and web-search-enabled forms.
The researchers set out to quantify a phenomenon that many internet users have noticed since the introduction and sometimes controversial rollout of AI-generated search summaries: generative AI searches often return information from sources and websites that are far less prominent than those found in traditional search results. This divergence goes beyond just the order of links—it extends to the fundamental nature of the sources being cited.
To assess this, the team compiled a diverse set of queries. These included specific user questions from the WildChat dataset (which logs real queries submitted to ChatGPT), broad political topics curated by AllSides, and a collection of products from Amazon’s top 100 most-searched list. They then compared the results returned by traditional Google search (the familiar list of ranked links) with those produced by AI-powered systems.
One of the most striking findings was that generative AI search engines cited sources from significantly less popular websites compared to traditional Google search results. Popularity was measured using the Tranco domain tracker, which ranks websites based on their overall web traffic and prominence. While traditional Google searches typically pulled from highly ranked (and thus, more visited) domains, AI search engines were much more likely to reference sites outside even the top 1,000 or top million domains. This was especially pronounced with Google’s own Gemini search, where the median cited source fell outside Tranco’s top 1,000 domains for almost all types of queries.
In addition to relying on less popular sites, the study found that the overlap between sources cited by AI-generated results and those found in Google’s top search results was surprisingly small. Over half (53%) of the sources referenced by Google’s AI Overviews did not appear in the top 10 Google links for the same query, and 40% were completely absent from the top 100 links. This means that users who rely on AI-generated search summaries may be exposed to information from websites they would never encounter through traditional search methods.
However, the research does not suggest that these AI-generated results are necessarily inferior. In fact, the AI systems—especially those based on GPT technology—tended to draw more frequently from corporate websites and encyclopedias, while rarely citing content from social media platforms. This difference in source selection could have both positive and negative implications, potentially filtering out unreliable or sensationalist information but also missing timely or firsthand reports that sometimes break on social networks.
The study also analyzed the depth and diversity of information provided