Google’s AI Overviews may be siphoning traffic away from websites, but a rigorous new field experiment indicates that the clicks they absorb are not of lower quality. This finding directly challenges a key defense Google has offered for its generative search feature, suggesting that the loss in organic traffic corresponds to a genuine loss of valuable visits, not just the filtering out of low-engagement, “bounce” clicks.
Researchers Saharsh Agarwal and Ananya Sen have published a revised working paper on SSRN that updates their earlier analysis of AI Overviews’ impact on web traffic. The new study, which builds on an initial finding that AI Overviews reduced organic clicks by 38%, now refines that figure to a 39.8% decrease in outbound clicks when the summaries are present. More critically, the paper provides the first experimental evidence on the quality of the clicks that are lost, concluding that the missing traffic is statistically indistinguishable from the traffic that remains when AI Overviews are shown.
How the Click Quality Test Worked
The updated study deployed a randomized field experiment where a control group of users did not see AI Overviews, while a treatment group did. To assess quality, the researchers focused on clicks for queries where an AI Overview was triggered or would have been triggered. They measured three specific engagement signals: the rate at which users hit the back button to return to the search results page, the proportion of visits that ended within 10 seconds with no interaction (a proxy for a bounce), and the total time spent on the destination site.
For all three metrics, the difference between the two groups was statistically insignificant. In both conditions, approximately 4 out of 10 same-tab clicks resulted in a return to the search results. Roughly 18% of visits ended within 10 seconds, and the average time spent on destination pages was effectively identical. The authors state plainly that this result is “at odds with the view that AIOs primarily eliminate low-engagement website visits.”
This evidence directly counters statements from Google’s VP of Search, Liz Reid, who has suggested that AI Overviews reduce so-called “bounce clicks” from the search results page. While that claim implies that the summaries are absorbing only low-value traffic that would have quickly left a website anyway, the experimental data shows that users in the control group—who clicked on regular links instead of an AI Overview—behaved in the same way as those who landed on a site after clicking through an overview.
The Meaning of “Bounce Clicks” in Google’s Defense
Google has long argued that its AI Overviews improve the user experience by providing direct answers that save users from clicking on unhelpful links. The company’s internal narrative has been that many of the clicks lost to AI Overviews were low-quality visits—clicks that would have resulted in a quick back-button press and a poor user experience. This experiment tested that assumption from the opposite angle: if AI Overviews were truly absorbing only low-value clicks, then the additional clicks that appear when the overviews are removed should be of noticeably worse quality. They are not.
The aggregate data from the experiment shows that when AI Overviews are removed, the total number of outbound clicks per search increases, and zero-click searches decline. The quality of those additional clicks, however, remains unchanged. The implication is that publishers are not losing a fringe of bad traffic; they are losing a significant volume of ordinary, valuable visits that users would have engaged with normally.
Behavior Reversed When Treatment Groups Switched
The robustness of the study was strengthened by a crossover design. After an initial two-week period, the researchers rotated the assignments: users who had been seeing AI Overviews no longer saw them, and vice versa. The results mirrored the first phase precisely. Participants who were served AI Overviews saw their external clicks per search decrease, while those who stopped receiving them saw their external clicks per search increase. The zero-click rates tracked in the opposite direction, confirming that the effect is causal and directly tied to the presence of the AI-generated summaries.
Where the Traffic Loss Is Concentrated
The researchers also broke down the data by query type, revealing a clear pattern. The traffic losses are overwhelmingly concentrated in informational queries—searches where users are seeking to learn something or find an answer. For these queries, AI Overviews were triggered or intended to trigger on 53% of cases. When the summaries were removed, clicks per search increased substantially for informational terms. Navigational and transactional queries, by contrast, showed no measurable change, though the sample sizes for those categories were smaller. AI Overviews were triggered on only 15% of navigational and 6% of transactional queries.
A further breakdown of click positions reveals that the top three organic results gained the most traffic when a top-of-page overview was removed. The first organic position nearly doubled its click share in the no-overview condition, indicating that AI Overviews are directly cannibalizing traffic from the highest-ranking, often most authoritative, sources.
The Aggregate Effect and Future Growth
Because AI Overviews were triggered on approximately 41% of all queries in the experiment, the 39.8% reduction in outbound clicks per search translates into a meaningful aggregate traffic decline across the entire search ecosystem. The authors note that this aggregate loss could grow significantly if AI Overviews are expanded to appear on a higher percentage of queries over time.
The paper remains a working draft on SSRN and has not yet completed formal peer review. However, its experimental design and the directness of its evidence give it considerable weight in the ongoing debate over how generative AI is reshaping the economics of web search and content publishing. The study also cites two Search Engine Journal articles in its discussion of Google’s public claims, providing a documented counterpoint to the company’s stated position.
The central implication for publishers, SEO professionals, and the broader digital media industry is that the traffic lost to AI Overviews is not a clean trade-off where low-quality visits are exchanged for a better user experience. The data suggests the opposite: valuable, engaged users are being kept on the search results page, and the websites that rely on their visits are bearing the full cost of the trade. As Google continues to refine and expand its AI-powered search features, the evidence from this experiment provides a critical, data-driven benchmark for measuring the real-world impact on the open web.
