Generative AI shifts online search from popular sites to broader, synthesised sources

A new study reveals that generative AI is transforming search engines by sourcing information from a wider array of less-known websites, raising questions about accuracy and authority in the digital information landscape.

A recent academic study conducted by researchers from Ruhr University Bochum and the Max Planck Institute for Software Systems offers a revealing look into how generative AI is reshaping the architecture of online search. The investigation, published a...

s a preprint titled “Characterizing Web Search in the Age of Generative AI,” highlights a fundamental shift from traditional search engines’ reliance on indexing and ranking popular web pages to AI-driven systems’ synthesis of information from a broader, and often less popular, range of sources.

The researchers analysed thousands of queries drawn from diverse datasets, including ChatGPT user interactions, politically charged topics from the media-bias monitoring site AllSides, and trendy product searches from Amazon. They compared results from conventional Google search with those generated by AI-powered search systems such as Google’s AI Overviews, Gemini 2.5 Flash, and two variants of OpenAI’s GPT-4o.

Their findings challenge assumptions about authority and relevance in search results. Using Tranco, an independent ranking system that measures the popularity of internet domains, the study showed that AI-generated answers frequently draw from websites outside the top 1,000 most-visited domains, with more than half of the sources cited by Google’s AI Overviews not appearing in the first 10 organic Google results. Gemini’s AI search results exhibited a similar pattern, often referencing sites less frequented or indexed than those prioritized by traditional search engines. Even GPT-4o’s web-enabled variants—which selectively retrieve external data—tended to cite institutional or encyclopaedic domains rather than social or discussion forums common in traditional search.

Rather than deeming these less conventional references inferior, the authors emphasised that AI search reflects a distinct model prioritising synthesis and summarisation. Utilizing Stanford University’s LLOOM evaluation tool, which assesses conceptual coverage, the study noted that AI systems tend to cover as broad a range of concepts as the top 10 Google results but moderate this breadth by condensing and integrating information. This approach, while beneficial for creating cohesive summaries, may gloss over nuances or alternative interpretations that traditional search preserves, particularly when queries are ambiguous—such as names shared by multiple people.

The integration of pre-trained large language models gives these AI systems an edge in background knowledge. GPT-4o’s Search Tool, for instance, can sometimes deliver comprehensive summaries drawn from its internal knowledge base without seeking additional data online. While this works well for established topics, it limits responsiveness to recent or breaking news, as evidenced by the AI’s tendency to produce unclear or generic replies when prompted with trending queries from mid-September 2025.

This emergent divergence between AI and traditional search holds implications for how users interact with information online. While traditional search engines like Google remain dominant—still used by 95% of Americans monthly, according to recent clickstream data—there is a visible increase in AI tool usage, with 21% of U.S. users accessing AI-driven search tools frequently. A survey cited by Innovating with AI reveals that 83% of users find these AI-powered search tools more efficient than conventional methods, echoing a shift in user preferences even as Google’s market share dipped beneath 90% for the first time since 2015.

However, concerns around accuracy and reliability persist. Separate research from Columbia Journalism Review’s Tow Center for Digital Journalism exposed significant citation errors, with AI models incorrectly attributing news sources in over 60% of cases. This trend of confidently delivering plausible but incorrect answers—the so-called “hallucination” or confabulation problem—is a persistent challenge across AI search platforms. Alarmingly, paying for premium AI search services has sometimes correlated with higher error rates.

From an SEO and content strategy perspective, these developments suggest a new landscape. AI search engines may prioritise authoritative content differently, with overlaps between traditional and AI-generated sources ranging widely—from as little as 33% up to 73% depending on the platform. For content creators, establishing robust, verifiable authority remains crucial, but adapting to AI’s unique ranking criteria and synthesis methodology will be vital.

The authors of the Ruhr and Max Planck study call for fresh evaluation frameworks tailored to generative AI search, advocating for metrics beyond conventional ranking systems. These should account for diversity of sources, conceptual breadth, and the efficacy of AI-generated summaries in delivering coherent, reliable responses.

In sum, generative AI is ushering in a pluralistic approach to web search—moving from retrieval of popular, high-ranking webpages to dynamic synthesis drawing on a wider, sometimes less-trodden web. As AI search tools evolve, balancing breadth with accuracy and enhancing user trust will be key challenges in navigating this next generation of information discovery.

Source: Noah Wire Services