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. This week that changed. A randomized field experiment, the first of its kind, replaced the correlation with a cause, and the number it produced is now the most defensible figure in the entire argument.

A study of 1,065 desktop Chrome users found that Google's AI Overviews cut outbound organic clicks by 39.8 percent and raised zero-click searches by 34.5 percent when the feature appeared, with no measurable improvement in how users rated their search experience. Around that finding, reported across the marketing trade press over the past day, sat a cluster of stories that share its logic. Cloudflare abandoned its year-old model of charging AI companies per crawl in favor of paying publishers per citation, naming the same traffic collapse as its reason. The European Union closed an eight-year antitrust case against Google. OpenAI began drafting the ad formats that will run on machine-read content. Each replaced an assumption the market had leaned on with a measured outcome, and each measurement carried a price.

A randomized design settles what observation could not

The paper was written by Saharsh Agarwal of the Indian School of Business and Ananya Sen of Carnegie Mellon University's Heinz College. According to the study, it 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, and the word causal is doing real work.

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 behaviour, a competing product, or one of Google's own algorithm changes, moved at the same time. This study used a custom-built browser extension to randomly assign real users either to see or not to see the summaries Google places atop many results pages. Random assignment meant the only systematic difference between the groups was whether an overview that would otherwise have appeared was shown or suppressed. That is the difference between noticing two things happen together and demonstrating that one causes the other.

The execution was exacting, and the numbers reward a close reading. 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, a control seeing the standard results page and a treatment group in which the extension hid AI Overviews, were each assigned with 36 percent probability; a third arm, redirected to Google's AI Mode conversational interface, took the remaining share. Across 396 control users and 374 hide-overview users, the researchers observed 68,089 unique searches, with AI Overviews triggered in roughly 41 percent of queries.

The core effect is precise. According to the paper, for queries where an overview was meant to appear, removing it lifted outbound organic clicks from 0.37 to 0.62 per search and cut the probability of a zero-click search from 0.73 to 0.54. Sponsored clicks did not move, and neither did clicks within Google's own properties, a pattern the authors read as AI Overviews substituting specifically for organic website visits rather than reshuffling attention across the page. A placebo check reinforced the design: for queries where no overview was triggered at all, clicks did not differ between the groups. The effect concentrates exactly where the feature sits. More than 87 percent of overview appearances placed the summary in the very top position, above every organic listing, and removing it from that slot produced an 88 percent relative increase in outbound clicks, against no measurable effect when the overview appeared lower on the page. Informational queries, 71 percent of the sample, drove the result with 0.26 additional clicks per search, while navigational and transactional queries produced noisier estimates that could not be distinguished from zero.

The quality findings are the part Google will find hardest to answer, because they test the company's own defence on its own terms. Google's Vice President of Product for Search, Liz Reid, characterised AI Overview clicks in July 2025 as higher-quality clicks showing greater purchase intent and longer downstream engagement, and Google's Vice President of Search, Nick Fox, had 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 navigating back to results via the browser's back button, a bounce measure defined as a session under ten seconds with no further interaction, and total time on page. None differed meaningfully between the groups. The additional clicks generated when overviews were hidden looked, by every one of those measures, just as engaged as the clicks that already existed. An endline survey completed by 90 percent of active participants found no difference in overall satisfaction, perceived information quality, or ease of finding information, results the authors describe as precisely estimated nulls. The AI Mode arm, treated as exploratory because of high attrition, pointed the same way more steeply, with satisfaction falling 1.14 points and compliant participants recording 0.36 external clicks per search against 0.53 in the control group.

The number lands inside a widening body of evidence

The headline figure does not arrive in isolation; it slots into a two-year run of measurements that PPC Land has tracked, and its value lies partly in where it falls within that range. 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, a figure that became one of the most cited data points in the whole publisher-traffic debate. That study directly contradicted Google chief executive Sundar Pichai's claim that content placed within AI Overviews earns higher click-through rates than content outside them. A follow-up Ahrefs analysis published February 4, 2026 put the reduction at 58 percent for top-ranking pages, nearly double the earlier figure, which suggested the effect intensified as the feature expanded.

Other measurements varied by method but pointed one way. Seer Interactive, analysing 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, while paid rates fell 68 percent over the same window, with even non-overview queries losing 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. The geographic picture is consistent: SISTRIX data from March 2026, cited in PPC Land's interview with Liz Reid, found AI Overviews cut the click-through rate at position one from 27 percent to 11 percent in Germany, costing that market 265 million organic clicks per month across 100 million keywords, and Index Exchange recorded an average 14 percent year-over-year decline in ad opportunities across 69 percent of publishers on its platform through 2025.

Reid's own position, set out in that interview, is that the click AI Overviews absorb is the shallow one. A reader who intended only to skim a headline before leaving now gets what they wanted from the summary, she argued, while a reader preparing to spend ten minutes with a long article will not treat an overview as a substitute. The randomized study complicates that framing rather than confirming it, since the additional clicks it surfaced when overviews were hidden showed no worse bounce, back-navigation, or dwell time than existing clicks. Independent data she was shown presents a more contested picture too, with Chartbeat figures indicating small publishers lost 60 percent of their search traffic in two years while AI chatbot referrals stayed below 1 percent of total page views.

The regulatory audience for all of this is not hypothetical. The paper lands while 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, and in the United States Penske Media's federal antitrust suit argues Google coerced publishers into supplying content for AI systems while cutting the traffic those publishers depend on, with a USA Today executive quoted rejecting Google's traffic-denial claims. Randomized evidence is a different standard than the correlational work that shaped the debate until now, and it surfaces precisely as several jurisdictions weigh whether opt-out or compensation should be mandatory rather than voluntary. That question of who pays, and for what unit, is where the next story begins.

Cloudflare moves the meter from the crawl to the citation

The same traffic collapse that the Agarwal and Sen study measured is the collapse Cloudflare named this week when it tore up its own year-old pricing model. Twelve months separate two announcements, and the distance between them tracks how fast the economics of machine-read content have moved. On July 1, 2025, Cloudflare told AI companies they would pay each time a crawler fetched a page. On July 1, 2026, the same company declared that model insufficient and proposed paying publishers only when their content actually appears inside an answer.

The argument rests on a specific figure. According to Cloudflare, more than half of the crawl traffic generated by bots it classifies as legitimate goes toward re-fetching pages that have not changed since the last visit, and that share is likely to grow as crawl volumes climb. The waste runs both directions: AI companies burn compute retrieving stale content, and publishers pay hosting costs to serve pages that yield no new information. Cloudflare, which says it sits behind more than 20 percent of websites on its network, casts that footprint as the reason it can attempt a fix, since no single publisher or AI company observes traffic at comparable scale. The commercial half of the announcement runs through two named partners operating two different mechanics. Ceramic.ai, led by founder Anna Patterson, runs a pay-per-query model under which opted-in publishers are paid when their content appears in Ceramic's search results rather than each time a crawler fetches a page, and participating publishers gain query-level reporting on the exact snippet shown and its average ranking position. The second partner, You.com, lets AI agents pay on demand for a specific piece of premium content at the moment it is needed. No pricing was disclosed for either.

A companion policy change extends the same logic to defaults. According to Cloudflare, from September 15, 2026, crawlers it 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 by default. An ad, the company reasoned, signals that a site owner meant a person to land on the page, so on those pages human attention is treated as the goal. The distinction between a search bot assembling a live response, an agent browsing on a user's behalf, and a training bot collecting data to fine-tune a model is the axis the whole scheme turns on.

The pressure behind both moves is the documented decline in referral traffic. According to Cloudflare, it cites 2025 Pew Research Center data finding 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, and click a link inside the summary itself only 1 percent of the time, figures Pew drew from browsing data across 900 US adults during March 2025 and Google disputed at the time as methodologically flawed. Cloudflare describes the entire effort as an experiment, states there is much to learn about how the approach holds at internet scale, and positions itself as an infrastructure layer providing the plumbing for payment rather than a party setting prices. Whether a citation is worth more or less than a crawl is, for now, a question the market has not priced. The randomized traffic study is precisely the evidence that makes the pivot legible, since it establishes that the crawl and the click have decoupled in a way a per-crawl fee cannot capture.

Luxembourg closes the Android case and hands regulators a template

While two of the week's stories measured the erosion of the click, a third settled an older argument about how Google reached the top of search in the first place. 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, in a proceeding that began when the European Commission opened its investigation in April 2015. Alphabet remains jointly and severally liable for 1.52 billion euros of the total, and the ruling exhausts the company's legal options.

The case rests on a Commission decision of July 18, 2018, which identified four sets of contractual restrictions and treated them together as a single and continuous infringement of Article 102 of the Treaty on the Functioning of the European Union, the provision prohibiting abuse of a dominant market position. Mobile Application Distribution Agreements required manufacturers seeking the Play Store to pre-install both Google Search and the Chrome browser, with the search tie running from January 1, 2011 and Chrome added from August 1, 2012. Anti-Fragmentation Agreements prevented manufacturers from selling devices running unapproved Android versions, and revenue share agreements paid manufacturers and operators a share of advertising revenue for not pre-installing any competing search service. According to the judgment, between 2011 and 2016 more than 50 percent of Android devices sold in the European Economic Area were covered by revenue share agreements requiring Google Search as the default, which meant the theoretical option of pre-installing a rival service was, in practice, unavailable on at least half of all devices.

The arithmetic of the fine has its own history. The Commission originally set 4.342 billion euros; the General Court in 2022 found procedural errors in the handling of the revenue-share element and cut the figure to 4.125 billion, leaving the search tie, the browser tie, and the anti-fragmentation obligations intact. The Court of Justice has now confirmed both that approach and that number. What matters for the marketing sector is the doctrine the ruling leaves standing. The Court held that no as-efficient-competitor test, the benchmark asking whether an equally efficient rival could have survived the conduct, is universally required under Article 102, and that in digital markets marked by high barriers to entry, network effects, and ecosystem lock-in, conduct falling short of competition on the merits can make entry by such a competitor practically impossible. That holding, together with the ruling that the combined effects of complementary agreements can be assessed jointly even where each agreement is not independently unlawful, hands enforcement agencies a template extending well beyond Android.

The template has somewhere to go, since the Android judgment runs in parallel with a dense field of proceedings PPC Land has tracked. According to that reporting, the Commission imposed a separate 2.95 billion euro fine in September 2025 for abuse in publisher ad servers and programmatic buying markets, a case still in its early stages, while the Court of Justice upheld a 2.4 billion euro fine in the Google Shopping case in September 2024, and in the United States Google filed an appeal in May 2026 seeking to reverse the August 2024 finding that it had illegally maintained a search monopoly. The pre-installation and default arrangements documented in the Android regime shaped where search demand landed for more than a decade, and the ruling removes any remaining ambiguity about their legality just as the same company faces newer scrutiny over how it treats content in the age of generated answers.

OpenAI drafts the ad formats for machine-read content

The buyer side of the same economy showed its own construction site this week, and the timing is pointed: the inventory being prepared for monetization is generated from content whose owners are, in the same news cycle, negotiating what a citation is worth. Three job listings on OpenAI's careers page point to text, image, video, native, conversational, and interactive ad formats in development, Digiday reported on July 1, 2026. The company is seeking an ad formats software engineer with at least seven years of experience for what the listing calls a foundational role across the full stack, sitting within the monetization team and responsible for the infrastructure defining how ads are structured, rendered, and delivered across surfaces and media types. Two further roles focus on iOS and Android experiences and require at least four years of experience. All three are based in San Francisco, and each carries a compensation package of 230,000 to 385,000 dollars plus equity.

What the listings reveal beyond intent is sequencing. Since launch, the ChatGPT ad unit has stayed simple: a headline, a short description, an image, and a link, with changes minor and iterative, including a recent tweak narrowing the unit from 480 to 440 pixels. Text ads with a label already run; conversational and interactive formats, in which the ad becomes part of the dialogue rather than an insert beside it, remain unbuilt. The people close to the effort are explicit about the tension. Rob Webster, chief executive of TAU Marketing Solutions, told Digiday the roles must tackle attribution, brand safety, and device modelling, and that setting this up will not be easy because no one yet knows the right way to run ads inside a conversational system. Andrew Frank, research vice president at Gartner, framed the deeper problem as one of dual alignment: whether the system optimizes for user trust or for advertiser value, objectives he described as often incompatible. Nate Elliott, principal analyst at eMarketer, argued it was overdue for OpenAI to test creative formats, since the company had, in his words, plowed ahead with a global launch of a single format and placement without knowing whether it works best for advertisers or users.

The financial logic is public. According to PPC Land, OpenAI's ad pilot began on February 9, 2026 with a closed roster of large brands, and a self-serve Ads Manager opened in beta to US advertisers on May 5, 2026 alongside cost-per-click bidding with 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. A market that cannot yet agree on the value of a machine-read page is building, at the same moment, the ad formats that will run on top of pages read exactly that way.

What the week measures

The stories share one structural feature: each replaces a proxy with a measured outcome, and the measurement carries a cost. The Agarwal and Sen experiment replaced observational correlation with randomized causation because the policy stakes had outgrown the older evidence, and it produced a number, 39.8 percent, that is harder to wave away than any figure before it. Cloudflare abandoned the crawl for the citation because a crawl no longer predicts value, naming that same decoupling as its reason. Luxembourg confirmed that default settings and exclusivity payments are abuse when a dominant firm deploys them, closing a proxy Google had defended for eight years. OpenAI began drafting a native taxonomy of formats rather than borrowing the assumption that AI ads look like search ads. The common thread is a click that used to be worth counting and now has to be measured against what replaced it. Markets reprice when their measurements change, and across a single week the marketing economy changed several of the numbers it lives by.

Also noted

  • July 1, 2026: Cloudflare confirmed that from September 15, 2026 it will block crawlers it classifies as Training and Agent by default on ad-bearing pages for new domains, while leaving Search crawlers allowed, according to PPC Land.
  • July 2, 2026: The Court of Justice of the European Union confirmed the 4.125 billion euro Android fine and rejected all six of Google's grounds of appeal, according to PPC Land.
  • May 27, 2026: Google extended its Preferred Sources feature into AI Overviews and AI Mode, reporting a doubling of click-through for labeled links and more than 345,000 unique sources selected by users, according to PPC Land.
  • July 1, 2026: OpenAI's monetization team posted three engineering roles pointing to image, video, native, and conversational ad formats in development, each paying 230,000 to 385,000 dollars plus equity, according to Digiday.
  • June 2026: A year-long tracking study found llms.txt adoption grew 8.8 times to nearly 39,000 sites, even as a separate analysis found 97 percent of the files received zero AI requests in May 2026, according to PPC Land.