Dwi Wanna
The landscape of search engine optimization (SEO) and digital publishing has been in a state of flux since Google began the aggressive rollout of its AI Overviews (AIO). For publishers, the primary concern has been a sustained decline in organic referral traffic. While Google has consistently defended the technology by suggesting that AIOs filter out low-quality "bounce" traffic, new empirical evidence is challenging this narrative.
A revised working paper by researchers Saharsh Agarwal and Ananya Sen provides a rigorous, data-driven look at the impact of AI Overviews on search behavior. Their findings suggest that the traffic lost to AI summaries is not merely "low-value" noise, but rather high-intent traffic that would otherwise have engaged with publisher content.
Main Facts: The 40% Traffic Squeeze
The core of the research relies on a randomized field experiment designed to isolate the impact of AI Overviews on user behavior. By comparing search results with and without AIOs, the researchers identified a striking 39.8% decrease in organic clicks when the AI summaries are present.
This figure is an update to their previous analysis, which reported a 38% reduction in April. The consistency of these findings across different data sets underscores the significant "cannibalization" effect that generative AI is having on the traditional search funnel. Furthermore, the study confirms that when AI Overviews are shown, there is a commensurate rise in "zero-click" searches—queries where the user obtains their answer directly on the search engine results page (SERP) and never clicks through to an external website.
Given that AI Overviews are currently triggered on approximately 41% of all queries, the cumulative impact on the web ecosystem is substantial. The research suggests that as Google expands the reach of AIOs, the aggregate loss of outbound traffic to publishers will likely intensify.
Chronology of the Experiment
The researchers approached this study with a methodology designed to mitigate bias and confirm causality. The study was conducted in distinct phases:
Phase One: Initial Observation
During the initial two-week observation period, participants were divided into groups, with some receiving standard search results and others receiving results augmented with AI Overviews. The divergence in traffic patterns was immediate, with the AIO group experiencing a sharp decline in outbound clicks.
Phase Two: The Treatment Switch
To ensure the results were not due to inherent differences in the user groups, the researchers rotated the assignments. Participants who were originally served AI Overviews were moved to the non-AIO group, and vice versa. The results were consistent with the initial findings: the behavior reversed immediately. Those who lost access to the AI summaries began clicking through to websites at a higher rate, while those who were newly exposed to the summaries saw their outbound clicks drop. The zero-click rates mirrored this pattern, further strengthening the evidence that the AI summaries—not user intent—were the primary driver of the shift.
Deep Dive: Does Click Quality Actually Improve?
One of the most contentious points in the ongoing debate between Google and the publishing industry is the nature of the "lost" traffic. Google’s VP of Search, Liz Reid, has publicly stated that AI Overviews improve the quality of clicks by reducing "bounce clicks"—visits where a user clicks a link but immediately returns to the search page, suggesting the site did not meet their needs.
However, the data from Agarwal and Sen’s experiment contradicts this defense. To evaluate click quality, the researchers analyzed three key metrics:
- Back-button frequency: How often users returned to the search page after a click.
- Short-duration visits: The percentage of visits ending within 10 seconds without interaction.
- Time on site: The total duration of the visit.
Across all three metrics, the researchers found no statistically significant difference between the group that saw AI Overviews and the group that did not. Approximately 40% of same-tab clicks resulted in a "bounce" in both conditions, and roughly 18% of all visits lasted less than 10 seconds.
The researchers conclude that the traffic being "absorbed" by AI Overviews is not merely low-quality or irrelevant traffic. If the AI were successfully filtering out low-value users, the "no-AIO" group would logically show a higher rate of bounces or shorter time-on-site metrics. Instead, the behavior remained identical. This finding is directly at odds with the narrative that AIOs are merely a helpful tool for pruning poor search results.
Categorizing the Impact: Informational vs. Transactional
The study also provides a nuanced breakdown of how different types of queries are affected by AI Overviews.
The Informational Dominance
The most significant traffic losses are concentrated in "informational" queries—searches where users are looking for facts, definitions, or explanations. This is the primary domain of generative AI, which excels at synthesizing information. The data shows that AIOs are triggered or intended for 53% of these informational searches.
Navigational and Transactional Stability
In contrast, navigational queries (e.g., searching for a specific brand’s website) and transactional queries (e.g., looking to purchase a specific product) showed no measurable change in click behavior. These categories are less prone to AIO intervention, with triggering rates of just 15% and 6%, respectively.
The "Top-Three" Effect
The research also examined which publishers bear the brunt of the traffic loss. When an AI Overview is removed, the top three ranked results see the most significant gains in traffic. Specifically, the result in the number one position nearly doubled its click-through rate when the AI summary was removed. This suggests that AI Overviews are not just replacing the "answer" at the top of the page; they are actively suppressing the visibility of the most authoritative and high-ranking content.
Official Responses and the Information Gap
Google has been protective of its internal metrics regarding AI Overviews. While executives have touted the technology’s ability to streamline the user journey and provide high-quality information, the company has yet to release granular, independent data to support the claim that the lost traffic is primarily composed of "low-value" clicks.
The publishing industry has reacted with skepticism. Many SEO professionals and digital media executives argue that Google’s positioning of AIOs serves a dual purpose: enhancing the user experience while simultaneously keeping users within the Google ecosystem to increase ad inventory and retention. Because the Agarwal and Sen study uses a rigorous, randomized experimental design, it serves as a necessary counter-balance to the proprietary, often opaque data provided by the search giant.
Implications for the Future of the Web
The implications of this research are profound for both the future of search and the sustainability of the open web.
The Erosion of the Referral Model
If the "informational" web—which includes news, tutorials, and deep-dive journalism—is systematically stripped of its traffic, the economic incentive to produce high-quality content diminishes. If websites cannot monetize traffic through direct visits, the ecosystem risks shifting toward a walled-garden model where only the largest, most entrenched entities survive, or where content is created exclusively to feed AI training models rather than to serve human readers.
The Role of Peer Review
It is important to note that the paper by Agarwal and Sen is a working draft on the Social Science Research Network (SSRN) and has not yet completed the formal peer-review process. However, the methodology is transparent, and the results provide a much-needed baseline for further discussion. As the digital industry continues to grapple with the "AI shift," academic scrutiny will be vital in separating marketing rhetoric from the reality of user behavior.
A Call for Transparency
Ultimately, this experiment highlights an information asymmetry. Google holds the keys to the world’s search data, and until the company provides verifiable evidence to support its claims about "click quality," the findings of independent researchers will remain the most reliable metric for understanding the impact of AI on the web.
As we look toward the future, the primary challenge will be finding a balance between the convenience of AI-summarized answers and the survival of the content creators who make those answers possible. If the current trajectory continues, the web may evolve from a network of linked destinations into a static archive, where the discovery of new, authoritative information becomes increasingly rare.
Evan Lee Salim
Lina Irawan
