![[Adobe Stock]](https://sp-ao.shortpixel.ai/client/to_webp,q_glossy,ret_img,w_750,h_375/https://www.psypost.org/wp-content/uploads/2026/11/google-ai-overview-search-750x375.jpg)
Google’s AI-generated search summaries frequently rely on highly credible sources, but they still produce unverified information in about one out of every nine factual claims. A massive new analysis reveals how these summaries bypass traditional search rankings, fundamentally altering how users receive information and how web publishers make money. The research was published on the preprint server arXiv.
When a person types a query into Google, they traditionally receive a ranked list of web pages. Recently, the company introduced AI Overviews. This system reads multiple sources and writes a single, customized paragraph answering the user’s prompt, placing this text at the very top of the search results.
The new feature represents a massive shift in online information retrieval. Billions of people now encounter these machine-generated answers, often without actively choosing to use an artificial intelligence tool. Because Google holds a dominant position in the search market, the way its software selects sources and synthesizes facts shapes the information ecosystem on a global scale.
Haofei Xu, Umar Iqbal, and Jacob M. Montgomery, computer scientists at Washington University in St. Louis, conducted a massive evaluation of this system. They wanted to understand when the feature activates, where it gets its information, how accurately it summarizes those sources, and how it impacts the websites that originally published the information.
The AI Overview system relies on a large language model. This is a type of artificial intelligence trained on vast amounts of text to predict and generate human-like sentences. Unlike standalone chatbots that rely entirely on their internal training data, Google’s system is designed to anchor its answers in specific, retrieved web pages. This architecture is intended to reduce hallucinations, a common problem where artificial intelligence invents false information.
Over a 40-day period in early 2026, the research team used an automated web browsing program to run 55,393 trending Google queries. They distributed this program across cloud servers, programming it to mimic natural human typing and avoid bot detection. They spread these searches across 19 topical categories, such as health, sports, and politics. The team recorded whether an AI Overview appeared, what text it contained, and which external websites it cited.
The researchers found that AI Overviews activated on 13.7 percent of all queries. The structure of the user’s prompt heavily influenced this rate. When a search was phrased as a direct question, the system provided a generated summary 64.7 percent of the time, compared to just 9.5 percent for standard keyword searches.
The activation rate also varied widely by topic. The feature appeared often for hobbies and science queries, but rarely for politically sensitive subjects. The researchers noted that these variations suggest Google exercises undisclosed editorial discretion in determining when to suppress the artificial intelligence tool.
Next, the team evaluated the websites that the system used to build its answers. They compared the credibility of these cited sources against the standard web pages that appeared lower down on the exact same search page. To accomplish this, they used an established academic scoring metric that ranks the journalistic and informational quality of thousands of internet domains.
The domains cited by the AI Overviews were consistently more credible than the standard first-page search results. The generative system relied much less on unvetted user-generated content, such as social media posts and video platforms, than traditional Google search rankings do.
At the same time, almost 30 percent of the websites cited in the AI Overviews did not appear anywhere on the first page of the regular search results. This indicates that the generative system uses a distinct selection process to find its information. It elevates different websites than the traditional search algorithm surfaces for the same prompt.
To test whether the software accurately reported what its sources said, the researchers broke the overviews down into 98,020 individual factual claims. They built an automated verification pipeline, powered by a separate large language model, that read the text of the cited websites. The program then checked if each generated claim was actually supported by the source material.
The analysis showed that 11 percent of the generated claims were not supported by the cited pages. Most of these errors were omissions, meaning the summary stated a fact that simply did not exist in the source it linked to. In a smaller number of cases, the generated claim directly contradicted the text of its own cited source.
The researchers found that using higher-quality sources did not fix this error rate. The credibility of a website had no statistical relationship to whether the system would accurately summarize its contents. The unsupported claim rate also varied by topic, with health and political summaries showing higher accuracy than summaries about entertainment or technology.
Finally, the team looked at the economic reality for the publishers whose content fueled the generated answers. They analyzed the computer code of the cited websites to count how many of them relied on display advertising to generate revenue.
More than half of all the pages cited by the AI Overviews contained digital advertisements. When a synthesized summary answers a user’s question directly on the search page, the user is less likely to click through to the original website. The researchers pointed out that a suppressed click results in a loss of ad revenue for the publisher who produced the original information.
The study highlighted an asymmetry in this economic arrangement. While the original publishers lose the web traffic they rely on, Google continues to display its own sponsored search advertisements directly alongside the generated summaries. In some cases, Google places its own ads above the AI Overview.
The automated nature of the research presents certain limitations. When analyzing the accuracy of the generated claims, the researchers found that live-updating queries caused artificial errors in their data. For instance, if a summary reported the current temperature, the weather website it cited might have updated its data by the time the researchers’ automated program read the page. This time delay would falsely flag the claim as incorrect.
Additionally, the research team excluded websites that require user authentication or block automated scraping, such as major social media platforms, from their accuracy checks. Because they could not read the text of these specific pages, a portion of the unsupported claims might actually have been grounded in uncollected data.
Future research will need to measure how users actually interact with these search summaries. While this study outlines what the software generates and cites, it remains unknown if users notice the citation errors or if the summaries change their beliefs on high-stakes topics. Understanding how these features alter human behavior will require direct observation of internet users across different demographics.
The study, “Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact,” was authored by Haofei Xu, Umar Iqbal, and Jacob M. Montgomery.