arXiv 2605.16428 Measures AI Search's Drag on Publisher Traffic Using Paired Google and Reddit Data

The first quasi-experimental evidence on how Google’s AI Overviews redistribute publisher traffic is now on arXiv. arXiv

.16428¹, revised May 20, exploits a natural experiment in Reddit’s content policy to isolate the causal effect of AI-generated search summaries on engagement. The finding runs against the prevailing narrative: AI Overviews increased daily comments in surfaced Reddit communities by 12.0% and commenting users by 12.3%.¹ But that gain is conditional on content type, and it evaporates under Google’s newer AI Mode interface.

The SFW/NSFW natural experiment

The paper’s identification strategy is clever. Google excludes NSFW Reddit communities from AI Overviews. SFW communities appear routinely. This creates a treatment group (SFW subreddits surfaced in AIOs) and a control group (NSFW subreddits excluded for policy reasons unrelated to content quality), observed over the same time window. A difference-in-differences design then attributes engagement changes to AIO exposure rather than to confounding trends in search behavior or Reddit’s own growth.

The DiD framework addresses the core methodological problem in every SEO-vendor study published to date: you cannot compare traffic “before and after AI Overviews” without controlling for the dozen other things Google changed on the same SERP in the same period. The SFW/NSFW split provides a credible counterfactual because the only systematic difference between the groups, in the paper’s argument, is whether Google’s AI layer surfaces them.

Experience-based content benefits, factual content does not

The 12%¹ engagement increase is not uniform. It concentrates in communities hosting experience-based content: personal advice, opinions, recommendations, first-person accounts. Communities organized around factual information saw no statistically significant engagement change from AIO exposure.

This is the paper’s most consequential finding for publishers. AI Overviews appear to be acting as a discovery layer for subjective, experiential content, directing users who read a summary toward the original discussion for nuance and personal testimony. Factual queries get answered in the SERP; experiential queries get a teaser that drives clicks.

AI Mode erases the experience-based advantage

The paper then examines Google AI Mode, the conversational follow-up interface that lets users refine their query within the AI summary rather than visiting the source. When AI Mode is active, the experience-based engagement gains disappear.

This is the mechanism that should worry publishers most. AI Overviews in their current form may drive some traffic to experiential content, but the conversational layer removes the user’s reason to leave Google at all. The user asks a follow-up question; the AI synthesizes an answer from the same sources without requiring a click. The content type that initially benefited from AIO exposure is the same content type AI Mode is designed to absorb.

The implication: any traffic gain from AI Overviews is temporary if Google’s interface trajectory continues toward conversational follow-ups. Publishers optimizing for AIO citation are investing in a distribution channel whose next upgrade reduces their value to zero.

The unsupported-claims problem

A companion measurement study, arXiv

.14021², audited 55,393 trending queries across March and April 2026 and found that 11.0%² of atomic claims in AI Overviews are unsupported by the pages Google itself cites as sources. This is not a hallucination rate in the conventional sense; it measures synthesis drift, where the AI generates a claim that is plausibly related to the cited source but not actually present in it.

The same study found nearly 30%² of domains cited in AI Overviews do not appear on the first page of organic results for the same query. Google’s source selection for AIOs is not simply “surface the top-ranked results.” It operates an independent selection mechanism, and that mechanism is opaque to publishers trying to optimize for citation.

Over half of the pages cited in AIOs carry display advertising, per the same study. Publishers lose ad revenue when AIOs answer the query without a click-through, even as Google continues to run its own sponsored ads on the same SERP. The asymmetry is structural: Google monetizes the page; the cited publisher does not.

Winners and losers: a two-tier SERP

Seer Interactive reported organic CTR on AIO-triggered queries dropped from 1.76% to 0.61%, a 61% reduction across 3,119 terms.³ The split is between queries where the user is looking for a specific brand or entity (which AI Overviews reinforce) and queries where the user wants an answer (which AI Overviews satisfy without a click). Mid-tier publishers producing generic explainers, tutorials, and how-to content occupy the worst position: their content is exactly the type AI Overviews synthesize directly.

For publishers cited inside AI Overviews, the picture is not uniformly negative. Seer Interactive found that cited publishers earn 35% more organic clicks and 91% more paid clicks⁴ than uncited competitors on the same SERP. Semrush’s 2026 data⁴ shows AI search traffic converting at 4.4x the rate of standard organic search. Getting cited is valuable. Not getting cited, while your competitor is, is catastrophic.

Reddit’s citation dominance

Reddit content appears in 49.4% of AI Overviews, according to Ziptie’s analysis.⁵ YouTube accounts for roughly 23.3% of citations and Wikipedia for 18.4%, per Surfer SEO’s audit of 46 million citations.³ AIOs now trigger on approximately 48% of tracked queries (BrightEdge, February 2026), up from roughly 31% a year earlier.³

Reddit’s dominance in AIO citations is not accidental. The experiential content that AI Overviews struggle to synthesize, personal opinions, advice, first-hand accounts, is Reddit’s primary output. Google’s AI layer effectively treats Reddit as a structured corpus of human experience, and the Zhang et al. paper confirms this is the content type that currently benefits from AIO exposure.

For publishers running owned sites, this is an uncomfortable signal. The content Google’s AI prefers to cite is community-generated discussion, not the single-author explainer format that mid-tier publishers have optimized for over the past decade.

What the data actually supports

The brief suggests covering “what publishers should actually do.” The honest answer from the data available: not much is well-supported. The research identifies the mechanism and measures the effect. Prescriptive SEO recommendations based on these findings would be speculation.

Two observations that follow directly from the evidence:

First, citation matters more than ranking. Nearly 30%² of AIO-cited domains are not on the first organic page. Traditional SEO position is no longer the primary gate for AIO exposure, though the independent selection criteria are unknown.

Second, the content-type asymmetry is the structural finding. Experience-based content retains some click-through value under current AIOs; factual explainers do not. AI Mode removes even that residual value. The economics of producing mid-tier factual content for search traffic have deteriorated, and no schema markup tweak or structured data hack is going to fix a problem caused by the AI layer answering the query directly.

Frequently Asked Questions

Do branded searches see the same CTR decline as generic queries?

No. Amsive’s analysis found branded searches with AI Overviews see an 18% CTR increase, while generic informational queries drop 34–46%. This creates a two-tier SERP favoring established brands — a split the Zhang paper’s Reddit-focused methodology captures indirectly at best.

How does the 61% CTR drop compare to independent measurements?

Pew Research tracked 68,000 real-world queries and found a 46.7% relative CTR reduction; DMG Media reported up to 89% drops on specific search types. Google has publicly disputed the Pew methodology, making the Zhang paper’s experimental design a more defensible baseline despite its narrower Reddit-only scope.

Are question-format queries affected differently?

The companion audit (2605.14021) found 64.7% AIO activation for question-formatted queries versus 13.7% across all query types. Questions are the format most likely to trigger an AI Overview, meaning FAQ-style and how-to content faces disproportionately high exposure to click suppression.

Is the CTR decline accelerating or stabilizing?

Industry tracking shows organic CTR on AIO queries partially recovered from ~1.3% to ~2.4% between December 2025 and February 2026. However, this coincides with Google expanding AIO coverage from ~31% to ~48% of tracked queries, so the rebound may reflect changes in AIO design rather than reduced AI presence.

Does the engagement finding generalize beyond Reddit?

Only partially. The SFW/NSFW split exploits Reddit’s unique content-policy structure, which has no equivalent on most publisher sites. The 12% engagement increase also reflects Reddit’s threaded discussion format — a static article page has no equivalent comment signal, so the magnitude likely differs significantly for traditional publishers.

Footnotes

1. The Impact of AI Search on the Online Content Ecosystem ↩ ↩² ↩³ ↩⁴

2. Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact ↩ ↩² ↩³ ↩⁴ ↩⁵

3. AI Overviews Statistics 2026: Google Search Impact Data ↩ ↩² ↩³

4. AI Search Is Eating Publisher Traffic ↩ ↩²

5. AI Search Citations 2026: Why Reddit Dominates ChatGPT, Perplexity, and Google AI Overviews ↩