
A note on scope: this article reviews the evidence I have been able to find as of November 21, 2026. When I say there is no evidence that the remaining visits are better, I mean none that I have found—not that none exists.
Two claims, unequal evidence
Do AI Overviews reduce publisher traffic—and are the visits that remain better? The available evidence gives unequal answers: a randomized experiment supports reduced outbound organic clicking within its tested setting, but its engagement comparison offers no measured support for better remaining publisher visits. [1, 2] An observational study agrees with the direction of the click finding, without establishing causation. [3]
The crucial distinction is that a traffic reduction and a quality offset are separate claims. Confidence in the experiment’s overall click result cannot simply be transferred to its comparison of people who went on to click, because that comparison selects users after the experimental treatment. [1, 2] Nor does the absence of a detected engagement difference prove that the remaining visits are equally valuable. [2]
What the randomized experiment establishes
A version of the paper The Impact of Google AI Overviews on Publisher Traffic and User Experience dated June 17, 2026, reports a randomized desktop Google Search experiment in which displaying AI Overviews reduced outbound organic clicks on overview-triggering queries by 39.8% relative to hiding them. [1] That is evidence of a causal traffic effect in the experimental setting—not merely a comparison between searches that happened to display an overview and searches that did not. [1]
The scope matters. Participants were desktop Chrome users recruited through Prolific, and the observation window lasted two weeks; the paper does not give calendar dates for that window. [1] These boundaries limit extrapolation across devices, populations and longer periods. [1] The reported percentage therefore should not be treated as a forecast of every publisher’s traffic loss. [1]
The intervention also changed the page layout: removing the overview moved ordinary results upward. [1] Consequently, the result measures the overall effect of displaying the overview in that layout, not an isolated effect of its text. [1]
There is also a reporting qualification worth retaining: Section 4.1 inconsistently describes the percentage change as an increase, while this article follows the introduction’s reduction framing. [1] This version may differ from a final published version. [1]
Finally, this experiment did not randomize citations or separately measure cited publishers or citation-link clicks. [1] It establishes neither the incremental traffic value of being cited nor a publisher-specific loss attributable to an overview. [1]
Observational agreement, not another causal test
The observational preprint Investigating Click Behaviors On Google Search Result Pages That Produce an AI Overview, published August 5, 2026, reports that search-result links were clicked on 8% of visits to U.S. Google pages with AI Overviews, compared with 15% of visits to pages without them. [3] Its observation dates are not given, and the search-result pages were reconstructed after the observation period. [3]
The direction agrees with the experimental finding: pages with overviews were associated with less clicking. [1, 3] But the two percentages compare page types, and query and user controls do not turn that association into proof that overviews caused lost publisher traffic. [3] They should not be presented as a second causal estimate of the experimental effect. [1, 3]
The preprint uses a different study design; it is not clear whether its authors or underlying data overlap with the experiment. [1, 3] Likewise, the experiment’s traffic and engagement findings come from the same June 17 paper, not two independent replications. [1, 2]
Are the remaining visits better?
The June 17 experiment reported no difference in downstream engagement, conditional on a click, between users shown AI Overviews and users with them hidden. [2] On the reported engagement metrics, that offers no measured support for the claim that the remaining publisher visits are better. [2]
However, the engagement comparison is not as causally clean as the overall click comparison. [1, 2] Clicking is itself an outcome that the overview affected, so restricting the analysis to people who clicked selects a post-treatment subset rather than preserving the original randomized comparison. [1, 2] The experiment’s strength on overall clicking therefore does not establish a causal answer about visit quality among those remaining visitors. [1, 2]
Statistical uncertainty also matters. The paper’s summary of this result reports no effect sizes or confidence intervals for the engagement outcomes, so it does not show the precision of that comparison or which differences the analysis could resolve. [2] “No detected difference” is not evidence of equivalence, and it does not rule out differences the reported analysis could not detect. [2]
The defensible conclusion is narrow: improved engagement was not demonstrated—not that better visits are impossible, and not that visit quality is proven unchanged. [2]
What the evidence does not show
Neither the experimental findings nor the observational comparison establishes a conversion, revenue, brand-value or citation-specific offset to reduced clicking. [1, 2, 3] The reported engagement comparison cannot stand in for those business outcomes. [2]
The evidence also leaves publisher-specific effects and broader generalizability unresolved. [1, 3] For publishers, business owners and marketers, this means these findings support concern about fewer outbound clicks, but not a measured business-value calculation or a promise that higher-quality visits compensate for the loss. [1, 2, 3]
What would resolve the remaining questions?
To evaluate the quality-offset claim, ask for a study designed to address selection into clicking, with clearly reported engagement effect sizes and confidence intervals. To evaluate a business offset, ask for measured conversions, revenue or other explicitly defined value outcomes—not an assumption that “better” traffic makes up the difference. To assess citation value, ask for evidence that separates the effect of a particular citation from the overall presence of an overview.
The bottom line remains bounded but useful: the randomized experiment supports fewer outbound organic clicks in its tested setting, observational evidence agrees in direction, and the reported engagement analysis does not show that the remaining visits are better. [1, 2, 3] The proposed quality offset still needs its own evidence.
Evidence I may have missed
This review is limited to the studies I have found so far. I could be wrong, and there may be research I haven’t come across—especially newer work, or studies that measure what happens after the click. If you know of evidence that supports or contradicts anything here, I’d like to see it. Send it to me through the link, and I’ll review it and update this article if it changes the conclusion.
Sources
- June 17, 2026. The Impact of Google AI Overviews on Publisher Traffic and User Experience. Randomized experiment — overall outbound click finding.
stackcyber.com/files/google-ai-overview-click-reduction.pdf (opens in a new tab) - June 17, 2026. The Impact of Google AI Overviews on Publisher Traffic and User Experience. Same paper as [1] — engagement-after-click finding.
stackcyber.com/files/google-ai-overview-click-reduction.pdf (opens in a new tab) - arXiv (August 5, 2026). Investigating Click Behaviors On Google Search Result Pages That Produce an AI Overview. Observational preprint.
arxiv.org/html/2608.04831v1 (opens in a new tab)