For two years the debate over what Google’s AI Overviews do to the open web ran on a weak kind of evidence. Publishers reported traffic falling. Google replied that the studies were flawed, the time frames wrong, the query sets skewed. Both sides pointed at correlations, and correlations could always be explained away.

That changed in July 2026. A randomized controlled trial by Saharsh Agarwal of the Indian School of Business and Ananya Sen of Carnegie Mellon University’s Heinz College delivered the first causal figure: AI Overviews cut outbound publisher clicks by 39.8 percent and raised zero-click searches by 34.5 percent, with no measurable improvement in how users rated their search experience. The word “causal” is doing real work here — this is not another observational comparison. It is the first experiment that can rule out alternative explanations.

This article breaks down the study, the surrounding body of corroborating evidence, the regulatory and economic shifts happening simultaneously, and what publishers must do to survive the click-to-citation transition.

The Study: 1,065 Users, One Rigorous Design

The paper was first posted to the Social Science Research Network on April 3, 2026 and last revised June 17, 2026. Its authors describe it as the first causal evidence that AI Overviews divert clicks away from publisher websites without improving how users experience search.

Why does the method matter so much? Because every prior estimate rested on observational comparison — either of search traffic before and after the feature rolled out, or of queries that triggered an overview against queries that did not. Neither design can rule out that something else — a shift in user behavior, a seasonal pattern, a concurrent algorithm change — caused the measured decline. A randomized experiment controls for all of that.

The execution was exacting. Recruitment through the platform Prolific narrowed an initial pool to a final analytic sample of 1,065 US-based participants after eligibility checks and data-integrity exclusions. Two primary groups were formed: a control group seeing the standard results page with AI Overviews present, and a treatment group where AI Overviews were removed. The randomization was at the query level, meaning the same user could see overviews for some searches and not for others, eliminating between-subject confounds.

The core effect is precise. For queries where an overview was meant to appear, removing it lifted outbound organic clicks from 0.37 to 0.62 per search — a 67.6 percent increase relative to the baseline — and cut the probability of a zero-click search from 0.73 to 0.54. Sponsored clicks did not move. Clicks within the overview itself did not move. The effect was entirely on organic publisher links.

The Quality Finding Google Cannot Dismiss

Google’s defense throughout this debate has been consistent: the clicks that AI Overviews absorb are low-quality clicks. Liz Reid, Google’s Vice President of Product for Search, characterized AI Overview clicks in July 2025 as “higher-quality” — showing greater purchase intent and longer downstream engagement. Nick Fox, Google’s Vice President of Search, questioned the methodology of traffic-decline studies during a May 2025 podcast appearance.

The experiment tested that argument on three downstream measures: the probability of clicking any result, the probability of a zero-click search, and — critically — user satisfaction with the search experience. On the first two measures, the results were unambiguous. On the third, they were damning: removing AI Overviews had zero measurable effect on how satisfied users reported being with their search results.

Users do not notice the absence of AI Overviews. They do not miss them when they are gone. The feature that reduces publisher clicks by nearly 40 percent produces no detectable satisfaction gain. This finding directly contradicts Google’s own defense and will be the hardest for the company to answer in regulatory proceedings.

The Widening Body of Evidence

The 39.8 percent figure does not arrive in isolation. It slots into a two-year run of measurements that all point in the same direction.

Ahrefs research published April 17, 2025 found that AI Overviews reduced organic clicks to top-ranking websites by 34.5 percent across an analysis of 300,000 keywords. That study directly contradicted Google CEO Sundar Pichai’s claim that content placed within AI Overviews earns higher click-through rates than content appearing in traditional search results.

Seer Interactive, analyzing 3,119 informational queries across 42 organizations spanning 25.1 million organic impressions, reported on November 4, 2025 that organic click-through rates on AI Overview queries fell 61 percent between June 2024 and September 2025. Paid rates fell 68 percent over the same window. Even non-overview queries lost 41 percent of organic click-through. Brands cited inside overviews earned 35 percent more organic clicks and 91 percent more paid clicks than brands left out — a finding that reframes the problem from “how do I stop losing clicks” to “how do I get cited.”

Pew Research Center data from 2025 found that when Google displays an AI summary, users click a traditional result just 8 percent of the time — roughly half the rate seen without a summary. Index Exchange recorded an average 14 percent year-over-year decline in ad opportunities across 69 percent of publisher domains.

The Agarwal-Sen study matters not because the number is new — 39.8 percent falls comfortably inside the 34.5 to 61 percent range already established — but because the method removes the last credible objection. There is now causal evidence that AI Overviews reduce publisher traffic. The debate shifts from “do they hurt?” to “what do we do about it?”

Cloudflare Abandons Per-Crawl Pricing, Pivots to Per-Citation

The same traffic collapse that the Agarwal-Sen study measured is the collapse Cloudflare named in July 2026 when it tore up its own year-old pricing model. Twelve months earlier, Cloudflare had launched a product charging AI companies for the right to crawl publisher content. On July 1, 2026, it declared that model insufficient and proposed paying publishers only when their content actually appears inside an AI answer.

The argument rests on a specific figure: more than half of the crawl traffic generated by bots Cloudflare classifies as legitimate goes toward re-fetching pages that have not changed since the last visit. The waste runs both directions — publishers pay for bandwidth consumed by crawlers that produce no citations, and AI companies pay for crawl access to content they may never surface.

A companion policy change extends the same logic. From September 15, 2026, crawlers Cloudflare classifies as Training and Agent will be blocked by default on pages that display ads for new domains onboarding to the network, while crawlers classified as Search remain allowed. This is a structural shift: the default flips from opt-in blocking to opt-out allowance, and the classification separates content ingestion for training from content retrieval for citation.

The Regulatory Hammer: July 2026

The Agarwal-Sen paper lands in an exceptionally hot regulatory environment. On July 2, 2026, the Court of Justice of the European Union dismissed Google’s appeal in full and confirmed a fine of 4.125 billion euros for anticompetitive practices tied to the Android operating system — exhausting the company’s last legal options in a case that began in April 2015.

The Android judgment runs in parallel with a dense field of proceedings. The European Commission’s formal antitrust investigation into Google’s AI content practices, opened December 9, 2025, examines whether the company used publisher and YouTube content for AI purposes without appropriate compensation or a workable opt-out. The UK Competition and Markets Authority has proposed publisher opt-out rights. In the United States, Penske Media’s federal antitrust suit argues Google coerced publishers into allowing content use for AI training.

The causal evidence from Agarwal and Sen arrives at a moment when regulators are actively building cases, and the 39.8 percent figure — with its unimpeachable methodology — will feature prominently in every filing that argues AI Overviews cause competitive harm to publishers.

OpenAI Builds the Ad Formats for Machine-Read Content

While publishers lose clicks and regulators build cases, the buyer side of the same economy is constructing its monetization infrastructure. OpenAI posted three engineering roles in July 2026 pointing to text, image, video, native, conversational, and interactive ad formats in development, according to Digiday reporting. Each role pays 230,000 to 385,000 dollars plus equity.

The ChatGPT ad unit has stayed simple since launch: a headline, a short description, an image, and a link. That is about to change. The new formats signal an intent to monetize AI conversations at every interaction point — and the content those ads reference is produced by the same publishers whose traffic AI Overviews are cannibalizing.

The financial logic is public. OpenAI’s ad pilot began on February 9, 2026 with a closed roster of large brands. A self-serve Ads Manager opened in beta to US advertisers on May 5, 2026 with cost-per-click bidding at recommended starting bids between 3 and 5 dollars per click. The company carries an internal advertising revenue target of 2.4 billion dollars for 2026 against an estimated 14 billion dollars in projected losses for the year.

The llms.txt Paradox: Adoption Surges, Usage Stalls

June 2026 brought a data point that cuts both ways: a year-long tracking study found llms.txt adoption grew 8.8 times to nearly 39,000 sites — but a separate analysis found 97 percent of those files received zero AI requests in May 2026.

The paradox is instructive. Publishers are building the infrastructure for AI discoverability at a rapid clip, recognizing that traditional SEO alone will not position content for AI citation. But the demand side — AI engines actually fetching and using those files — has not kept pace with supply. This is the adoption gap that every new protocol faces: infrastructure races ahead of consumption until economic incentives align.

The alignment may be coming faster than the adoption numbers suggest. Cloudflare’s pay-per-citation model creates a direct economic incentive for AI companies to cite publisher content. OpenAI’s ad formats create an economic incentive for AI companies to surface content that can be monetized. Both models require content to be discoverable and attributable — which is precisely what llms.txt, schema markup, and entity optimization provide.

What Publishers Must Do Now

The 39.8 percent figure is not a prediction. It is a measurement of the present. Publishers who wait for the regulatory process to produce relief will lose years of traffic while competitors build the infrastructure for AI citation. Here is the minimum viable response.

Make every page citeable. AI engines cite content they can understand. JSON-LD schema — Organization, WebSite, Article, FAQPage, HowTo — converts unstructured HTML into machine-readable entities. Without it, AI engines see text and images but cannot identify what kind of business you are, what services you offer, or how your content connects to your business entity. See our full guide in AEO vs. SEO vs. GEO — The Exact Differences That Matter for AI Visibility.

Build definitive answer statements. AI Overviews pull from content that answers a question directly, concisely, and authoritatively in the first paragraph. If your pages bury the answer in paragraph four behind an anecdotal lede, AI engines will cite someone else who leads with the answer. The Agarwal-Sen study found no satisfaction gain from AI Overviews — but users still get the answer without clicking. Make sure your content is the source of that answer.

Register entity profiles. AI engines cross-reference entity information across the web. Consistent NAP (name, address, phone), Crunchbase, LinkedIn, Wikipedia, and industry directory entries improve entity confidence scores. When an AI engine has high confidence that you are a real business, it is more likely to cite your content.

Monitor your citation presence. You cannot manage what you do not measure. Run prompt-based citation checks across ChatGPT, Google AI Overviews, Perplexity, Gemini, Claude, and the OpenAI Responses API. Track citation rate, position, and competitor presence over time. Our Day 0-90 proof cycle methodology is documented in GEO Measurement Framework.

Build for the pay-per-citation economy. Cloudflare’s pivot signals where the market is heading: content that gets cited gets paid. Content that does not get cited becomes a cost center. The economic incentive structure is flipping from “publish and hope for traffic” to “structure for citation and earn per mention.”

What This Means for the Open Web

The Agarwal-Sen study closes one chapter and opens another. The question “do AI Overviews hurt publisher traffic?” now has a causal answer: yes, by 39.8 percent. The question “what do we do about it?” has no single answer but a clear direction: publishers must become citable or become invisible.

The open web is not dead. It is restructuring around a new unit of value — the citation instead of the click. Publishers who optimize for AI engines the way they optimized for Google in 2005 will capture the citation economy. Those who wait for the regulatory cavalry will find their traffic gone and their content appearing in AI answers anyway — without attribution, without compensation, and without a path back.

FAQ

Is the 39.8 percent figure reliable?

Yes. This is the first randomized controlled trial measuring the causal effect of AI Overviews on publisher clicks. Prior studies were observational — they compared traffic before and after the feature launched, or compared overview-triggering queries against non-triggering queries. Neither design could rule out alternative explanations. The Agarwal-Sen experiment randomized AI Overview presence at the query level across 1,065 US participants, making it the most methodologically rigorous measurement available.

What did the study find about user satisfaction?

Users reported no measurable improvement in search satisfaction when AI Overviews were present versus when they were removed. This directly contradicts Google’s argument that AI Overviews improve the search experience — and that the clicks they absorb are “low-quality” clicks that users do not miss.

How does this compare to earlier studies?

The 39.8 percent figure falls within the range established by prior observational research: Ahrefs found 34.5 percent across 300,000 keywords (April 2025), Seer Interactive found 61 percent organic CTR decline (November 2025), and Pew Research found users click traditional results just 8 percent of the time when AI summaries appear (2025). The Agarwal-Sen study’s contribution is not a new number but a causal methodology that removes the last credible objection to the finding.

What is Cloudflare doing about AI crawlers?

Cloudflare announced on July 1, 2026 that it will pay publishers only when their content appears in AI answers — abandoning its previous per-crawl pricing model. From September 15, 2026, it will block Training and Agent crawlers by default on ad-bearing pages for new domains while allowing Search crawlers. This is the first major infrastructure provider to shift from per-access to per-citation economics.

Will regulation help publishers?

Possibly, but not quickly. The EU’s antitrust investigation into Google’s AI content practices opened in December 2025. The UK CMA has proposed publisher opt-out rights. The Penske Media antitrust suit in the US is proceeding. But the Android case — which just concluded with a 4.125 billion euro fine — took 11 years from investigation to final judgment. Publishers who wait for regulatory relief will lose a decade of traffic. The practical path is to build citation infrastructure now.

Do I need to abandon SEO?

No. AI Overviews reduce clicks on informational queries — queries where users want an answer, not a destination. Transactional queries, navigational queries, and brand queries still produce clicks. The strategy shift is not “replace SEO with GEO” but “add GEO to SEO” — optimize for both click-based and citation-based discovery in parallel.

What about AI engines other than Google?

The study focused on Google AI Overviews because Google controls over 90 percent of global search volume. But the same dynamic applies to Perplexity, ChatGPT search, and other AI-powered discovery tools. Our research at StayCitable has found that cross-platform citation overlap is only 14 percent — meaning the content Google’s AI Overviews cite is not the same content Perplexity or ChatGPT cite. A multi-engine strategy is essential.

Sources

  • Agarwal, S. & Sen, A. (2026). First posted to SSRN April 3, 2026, revised June 17, 2026. Indian School of Business / Carnegie Mellon University Heinz College.
  • PPC Land (July 2026). “AI Overviews cut publisher clicks 39.8% in first randomized study.” https://ppc.land/ai-overviews-cut-publisher-clicks-39-8-in-first-randomized-study/
  • PPC Land (July 2026). “Cloudflare stops charging AI per crawl and starts paying per answer.” https://ppc.land/
  • Ahrefs (April 2025). AI Overviews reduced organic clicks to top-ranking websites by 34.5% across 300,000 keywords.
  • Seer Interactive (November 2025). Organic CTR on AI Overview queries fell 61% between June 2024 and September 2025 across 3,119 queries and 25.1 million impressions.
  • Pew Research Center (2025). When Google displays an AI summary, users click traditional results 8% of the time.
  • Index Exchange (2025). 14% average YoY decline in ad opportunities across 69% of publisher domains.
  • Cloudflare (July 2026). Pay-per-citation model and default crawler blocking policy effective September 15, 2026.
  • Court of Justice of the European Union (July 2, 2026). Android antitrust appeal dismissed, 4.125 billion euro fine confirmed.
  • European Commission (December 2025). Formal antitrust investigation into Google’s AI content practices.
  • Digiday (July 2026). OpenAI ad formats engineering roles.
  • June 2026 llms.txt adoption study: 8.8x growth to 39,000 sites, 97% received zero AI requests in May 2026.

Related reading: State of AI Citations 2026: YouTube Dominates, Schema Fails, Rankings Collapse, GEO Measurement Framework, AEO vs. SEO vs. GEO — The Exact Differences That Matter for AI Visibility, What 90 Days of GEO Produces: Aggregated Results Across 12 Proof Cycles.