Two weeks ago the outcome of the largest advertising antitrust case in the United States was reduced to a single sentence: Google would not have to sell AdX. The reasoning behind that sentence, and the machinery the court intended to build in place of a divestiture, sat under seal. On Tuesday evening the seal came off, and a 106-page opinion arrived unredacted. What it contains is not a rebuke softened into a warning. It is a set of engineering obligations with a six-year supervisor attached.

That document landed on the same day that publishers gathered in Miami to describe, under a promise of anonymity, artificial intelligence licensing agreements they are contractually forbidden from discussing. A television trade body published a study of how many advertiser accounts Google suspends and how little the company says about the ones it misses. The European Commission prepared a draft law that would require the largest platforms to submit new features for approval before launch. And three separate television companies announced arrangements that move viewing data further inside the buyers who use it.

The connecting thread across all of it is disclosure, and specifically who gets to choose when it happens. A federal judge decided that Google will publish how its ad server picks a winner. Publishers signing licensing deals discovered that they may not publish anything at all. A platform publishes enforcement totals in the billions and declines to publish the denominator. A regulator proposes that compliance plans be filed in advance rather than explained afterwards. Some of these disclosures are compelled, some are volunteered, and the difference between the two categories is turning into the more interesting story.

What the sealed opinion actually said

Judge Leonie Brinkema filed her remedies decision under seal on September 2, 2026, accompanied by a short public order confirming that AdX would stay in Google's hands. The full text became public on September 16, unredacted and exactly as written, fourteen days later, in a filing reported by Allison Schiff. The Department of Justice sued in January 2023. The liability trial was held in 2024. What has changed since the case was last examined here is not the verdict but the specification, and the specification is where the consequences live.

Start with interoperability, because that is the centre of the ruling. Google is required to build API integrations connecting both AdX and DFP to Prebid, the open-source header bidding framework that publishers have used for the better part of a decade to route demand around Google's own pipes. Those integrations must be functionally equivalent, a phrase that does a great deal of work: it forecloses the familiar pattern in which a rival connection technically exists and technically underperforms. AdX is further required to submit real-time bids into competing publisher ad servers on the same terms that DFP receives them. The exchange, in other words, can no longer treat its sibling ad server as a privileged destination.

Then comes data. Google must share bid data covering both wins and losses with publishers, and must publish technical documentation explaining how DFP selects a winning bid. Publishers have spent years reconstructing that logic from observed outcomes, building analytics practices around inference because the mechanism was proprietary. The court has now made the mechanism a publication requirement rather than a research project.

The buy-side restrictions are narrower and more surgical than the DOJ had sought. AdWords may no longer bid directly into DFP. It is prohibited from favouring Google's own ad tech tools, and it is restricted from using first-party data in ways that advantage Google's own infrastructure. Display and Video 360 emerged untouched: the court found an insufficient connection between that platform and the anticompetitive conduct at issue, and imposed no restrictions on it at all. For a demand-side platform that competes directly with independent DSPs, that is a meaningful asymmetry inside a ruling otherwise concerned with levelling access.

Compliance will be supervised by a court-appointed technical monitor serving a six-year term, with full access to Google's employees, systems and source code. Six years is an unusually long horizon for behavioural oversight of this kind, and source code access is an unusually deep one. The obligations apply globally rather than only in the United States, and take effect sixty days from judgment.

The refusal to order divestiture rested on two stated concerns: uncertainty about how the remedy would survive appeal, and the harm a forced sale could inflict on small businesses. Both are defensible. Both also produce the outcome Google preferred, and the compensating mechanism is a set of requirements that Google itself will implement, document and operate, under observation, for the next six years. Whether functional equivalence holds in practice is a question that only the monitor will be positioned to answer, and monitors report to courts rather than to the market.

One detail deserves particular attention from anyone modelling what happens next. The remedies were all argued publicly during the remedies phase, so the shape of the order was not a surprise. The sealing covered the reasoning and the precise drafting, which is exactly the material an engineering team needs to estimate the cost of compliance and exactly the material a competitor needs to judge whether compliance has occurred. Fourteen days is not a long delay by the standards of federal litigation. It is a long delay by the standards of a market that reprices on quarterly earnings.

Publishers describe deals they are not allowed to describe

Roughly fifteen hours after the opinion was unsealed, Digiday published an account of what publishing executives said in Miami this week under Chatham House rules at its Publishing Summit. The anonymity is the point. Almost every substantive remark about AI licensing came with an explanation of why it could not be made on the record.

"Everyone is coming with a different deal, and none of us are allowed to talk about these deals publicly," one executive said. "It's super freaking frustrating." Another described the resulting information vacuum in operational terms: "You have no idea what the metrics mean, and so there are so many different models out there. They're completely opaque." A third was blunter still: "We can't even talk about the deals that we sign." The assessments of the terms themselves ranged from "some of the contract terms can be really, really ugly" to "this is a new fresh hell" to, from one attendee, "I'm going to cry myself to sleep tonight."

Underneath the exasperation sits a structural argument worth separating from the mood. Several executives observed that the money is largely concentrated among the very biggest publishers, and one offered a theory for why no reference price has emerged. "I think it's because those who have big checkbooks to go acquire data are avoiding the marketplace because they do not want a value established there," the executive said. The comparison drawn was with the platforms publishers already deal with: "With the Metas and the Googles, you kind of know what's the standard contract." Without a standard contract, collective negotiation becomes difficult in a way one attendee described precisely: "I think that is scary because we can't put a concentrated effort."

That is a market failing to form. Prices exist, transactions occur, and no participant can observe any other participant's terms. Compare it against the ruling described above, in which a court ordered a platform to publish the logic of its auction so that counterparties could evaluate their own outcomes. The publishers in Miami are counterparties without that instrument, and nothing obliges anyone to give them one.

On traffic, the tone was resigned rather than alarmed. "The idea of Google Zero is, we're not going to hit zero," one executive said. "I think what people mean is search zero. It will decline." Several expected Discover to grow, though as one put it, "not to the degree that we saw search." The commercial conclusion followed: "I think the days of making money on mass traffic from Google referrals are generally down and over," alongside an observation that "the days of high yield programmatic revenue is just coming to an end." Esther Cohen, director of audience and subscriptions at The Verge, put the planning consequence plainly and on the record: "I think of Google traffic now as extra. I don't like to plan or optimize for Google."

Two measurements surfaced in the discussion that are worth recording because so few public figures exist. Executives referenced the AI performance report available to every publisher in Google Search Console, with data reaching back to May on AI Overviews, and noted the limitation: impressions are reported, but the comparison that would matter is missing. "If they show clicks and impressions, you'd see how much traffic you're losing because of CTR drops," one said. The working estimate offered in the room put click-through rate from the AI Overviews section at about 1 per cent. Separately, one publisher reported a 40 per cent increase in customer acquisition and traffic after concentrating on optimising articles for AI systems, and another described writing about specific products inside an onsite AI search chatbot on a licensing fee.

The content strategy that follows was described without sentiment. "We wrote a lot of listicles. We did a lot of stuff that, frankly, was commoditized. That's gone away." The replacement was framed as a filter rather than a plan: "Focus on the things that actually provide value, the things that AI slop can't replace." Whether that filter survives contact with a declining page-view base is an open question, and one attendee gave the honest version of the forecast: "Programmatic is going to go down consistent with what we see in our page view decrease."

A platform reports 24.9 million suspensions and not the number that got through

The Video Advertising Bureau published a study on September 15 titled Bad Actors: Examining How Much Objectionable Advertising Google Takes Action Against, and the headline comparison is deliberately provocative. Drawing on Google's 2025 Global and US Ads Safety Reports, both published in April 2026, and on Nielsen Ad Intel data covering January 1 to December 31, 2025, the VAB set Google's 3.3 million suspended United States advertiser accounts against the 86,800 brand advertisers that ran television commercials in the same year. The ratio is roughly thirty-eight to one.

The global totals are larger still. Google reported blocking or removing 8.3 billion advertisements in 2025, suspending 24.9 million advertiser accounts, blocking or restricting more than 480 million publisher pages, and taking action against more than 245,000 publisher sites. Scam-related enforcement accounted for 602 million removed advertisements, about 7 per cent of the total, and more than 4 million suspended accounts, about 16 per cent. In the United States specifically, 1.7 billion advertisements were blocked or removed, a fifth of the global figure, against 13 per cent of global account suspensions.

The category breakdown reads like a map of what automated systems catch most easily. Abusing the ad network led at more than 1.29 billion advertisements, followed by personalisation violations at more than 755 million, legal requirements at 646.7 million, misrepresentation at 421.5 million and trademark at 372.7 million. Dating and companionship accounted for 354.2 million, financial services 327.8 million, sexual content 321 million, gambling and games 270.7 million, copyright 229.4 million. Further down the list, counterfeit goods produced just 513,000 actions. Sexual content, which is restricted rather than banned, made up 409 million of the 480 million publisher pages actioned, about 85 per cent of that entire figure.

Two arithmetic observations complicate the presentation. The named categories account for 5.27 billion of the 8.3 billion advertisements, roughly 63 per cent, leaving more than three billion actions uncategorised in public reporting. And the suspension total fell sharply year over year, from 39.2 million accounts in 2024 to 24.9 million in 2025, a decline of 36 per cent. Google said in November 2025 that it had cut incorrect advertiser suspensions by 80 per cent, which offers one explanation for the drop, though a reduction in false positives and a reduction in enforcement are indistinguishable from outside the system.

The gap the VAB identifies is the one that matters commercially. Google states that its Gemini-powered systems stopped more than 99 per cent of policy-violating advertisements in 2025 before they were served. The company provides limited visibility into the remainder: how long those advertisements stayed active, how many people saw them, or where they appeared. Applied to 8.3 billion, the residual implies a ceiling of up to 83 million advertisements that could have served before removal. The actual figure is undisclosed, and the ceiling is a calculation rather than a disclosure.

The VAB is a television trade body, which is the relevant context for a comparison constructed between suspended digital accounts and legitimate television advertisers. Those are not equivalent units, and the report's own takeaway slide sits awkwardly against its ranking data, while 4.8 billion restricted advertisements in 2025 do not appear in the deck at all. The critique is self-interested. It is also, on the specific question of the unserved denominator, correct.

A smaller disclosure moved in the opposite direction the following day. Google Ads has updated its spend benchmarks report so that it now compares an account's spending against similar businesses, showing whether spend sits above, below or level with competitors alongside the clicks each produced. Screenshots shared on LinkedIn by Thomas Eccel showed an account at 284 euros against a competitor figure of 268 euros, and 912 clicks against 765. The comparison is built on industry and advertising location, and sits under the account Overview tab. Eccel's caution was that businesses within a single industry operate with completely different margins, conversion rates, average order values and strategies, and that profitability rather than a platform recommendation is what should govern a spending decision. A platform that declines to say how many violating advertisements reached users will now say how much a competitor spent.

Brussels proposes approval before launch

The European Commission was scheduled to present an EU approach to online child safety on September 17, a Communication to Parliament and Council carrying the reference COM(2026) 680/2. The draft circulated ahead of the presentation, and its central structural proposal is a four-bracket age framework for social media and comparable services. Children below three would have no access to social media or risky services. Between three and thirteen, access would run through parent-controlled accounts on child-friendly services. Between thirteen and fifteen, users would hold introductory accounts with strictly limited features, capped peer contacts and daily screen time limits, none of which the draft yet quantifies. Autonomous accounts would begin at fifteen, in environments the draft describes as safe by design.

The design requirements are specific enough to name mechanisms. The draft mandates no addictive design features, listing infinite scrolling, artificial notifications and certain reward mechanisms. Recommender systems would be required to go beyond engagement-based signals. Accounts would be private by default, unwanted contact from unknown users blocked, reporting and support routes made easy to reach, and parental control tools made age-appropriate. Scope is defined through a category the draft calls social media plus, covering social platforms and video-sharing services with design risks, online games, and AI chatbots and companions described as virtual tools capable of giving mental health and personal development advice. Services designed for education, public authority services, industrial AI and office products fall outside it.

The enforcement design is the part with the widest implications. Very large online platforms and very large online search engines would submit compliance plans to the Commission, and new features could not launch without a positive opinion issued within thirty days. That is prior approval, applied to product release, for a defined class of companies. Age verification would become mandatory at account creation using the EU age verification tool or an equivalent public authority solution, with member states targeted to deploy privacy-preserving verification applications by December 31, 2026. App stores would carry consistent age ratings and verify age at download for games and restricted applications. A supervisory fee would fund Commission enforcement, at an amount the draft does not specify.

The evidentiary base is drawn largely from a June 2026 Eurobarometer. Children average 4.5 hours online on school days and 6.1 hours at weekends, rising to 7.5 weekend hours among those who started using social media before the age of ten and falling to 5.7 among those who started after fourteen. One in three teenagers report stress, social exclusion or concentration problems. One in four encounters harmful content, and one in four experiences online sexual solicitation before turning eighteen. One in ten reports being pressured to share intimate images.

The Commission's framing is unusually direct for a legislative communication. Platform business models "designed to maximise the time spent online, result in a massive scale of online engagement and may exploit the vulnerabilities of children," the draft states. Two lines carry the political argument: "Parents, not algorithms, should be raising Europe's children," and "the Union's approach needs to limit the access of tech companies to our children, and not the other way around." On the gap being filled, the draft notes that existing EU law does not specify a minimum age for platforms with risky features and contains no outright ban on endless scrolling, excess notifications or harmful recommender systems.

The proposal arrives on a crowded field. Seventeen member states are preparing or negotiating national legislation, among them Austria, Denmark, France, Greece and Spain. Australia's under-sixteen ban covering nine named platforms, with no parental exemption, took effect on December 10, 2025. The United Kingdom announced on June 15, 2026 that it is targeting spring 2027 for an under-sixteen ban. In the United States, a federal consent judgment of August 26, 2026 imposed a two-hour daily limit on teenage Instagram and Facebook use. Enforcement under the Digital Services Act is already running in parallel: preliminary findings against TikTok on February 6, 2026 addressed infinite scroll, autoplay, notifications and recommender systems, and findings against Meta on July 10, 2026 addressed addictive design, each carrying exposure of up to 6 per cent of global turnover.

France supplies the cautionary precedent. A national under-fifteen ban was approved on July 21, 2026 and struck down by the Constitutional Council on August 14, eighteen days before implementation, on proportionality grounds. A reworked proposal was submitted on September 14. Whether a Union instrument survives the same test is listed among the draft's own outstanding questions, alongside how platforms will be classified as risky, what the contact and screen-time caps for introductory accounts will be, whether the thirty-day approval covers all features or only those reachable by minors, and what the supervisory fee and penalty levels will be. Reporting from POLITICO and Euractiv described the under-thirteen treatment differently from the draft text, which is a reasonable indicator that the provision is still moving. For advertisers, the operative change is not the age thresholds but the data minimisation consequence of verification at account creation and at download, applied across an entire category of inventory.

Television buyers move the data indoors

Three separate arrangements in three days point the same direction: the parties that buy television advertising are pulling measurement and content signals inside their own systems rather than requesting them.

Horizon Media has integrated Roku's television data directly into blu., its proprietary data and intelligence platform, in an arrangement described on September 16. The mechanical change is the removal of the request cycle. Where campaign information previously arrived through manual requests, Roku's device and audience datasets now flow continuously, carrying viewing behaviour such as genre preference and viewing habits, along with conversion rate metrics tied to those patterns. Samantha Rose, executive vice president and head of integrated investment and programmatic at Horizon Media, framed the value as timing: brands can act on the insights while a campaign is still in flight. Sal Candela, vice president of global agency partnerships at Roku, and Elizabeth Luciano, executive vice president of marketing at A+E Global Media, represented the supply and advertiser sides. The substantive shift is from probabilistic panel-based estimation to deterministic device data, and the latency reduction is what an agency is actually buying.

Wurl approached the same problem from the inventory end. On September 15 the AppLovin subsidiary launched a content intelligence platform that attaches contextual descriptions to connected television advertising requests through direct OpenRTB integration, before demand-side platforms receive them. The scale claimed is 140 billion signal-rich impressions a month across roughly 70 per cent of the ad-supported streaming ecosystem, up from 95 billion across more than 55 streamers reported in August 2025. The problem being addressed is documented: Peer39 found in March 2026 that only 40 per cent of connected television bid requests carry usable program-level content signals, an IAB study in July 2026 found 43 per cent of CTV buyers unsure where their advertisements ran, and 79 per cent of respondents to an AdExchanger and Wurl survey agreed that data fragmentation limits effectiveness. KERV supplies object-level recognition and ad-break targeting, Peer39 supplies pre-bid context, suitability, fraud and quality signals, and IRIS.TV, a Viant subsidiary since November 2024, supplies content classification. Yahoo DSP, StackAdapt and Viant are integrating. Greg Joseph, vice president of inventory development at StackAdapt, described buyers as seeking transparency on placement and precise brand-content alignment. Pricing, revenue splits, geography, the participating streamer list and performance data were not disclosed.

The third arrangement concerns a format rather than a dataset. Nexxen has opened DIRECTV Advertising and Philo pause ad supply to Simpli.fi buyers, transacted through the Nexxen supply-side platform and reachable through its own demand-side platform and approved third parties, of which Simpli.fi is the only one named. A pause ad occupies the screen when a viewer stops playback, and the behavioural case is specific: Philo reports that 81 per cent of its viewers pause to avoid missing content, and that 54 per cent of pause sessions last between one and five minutes. Attention benchmarks cited from WunderKIND put pause units at roughly twice the attention of a standard sixty-second connected television spot, with an automotive creative recording 34.2 seconds against 12.2 seconds for comparable video. Kara Puccinelli, chief commercial officer at Nexxen, offered the format's thesis: "When a viewer pauses content, they're not tuning out, they're leaning in." Aulden Kaye Yi of Philo called pause advertising one of streaming's most impactful formats. Michael Schoen, chief product and technology officer at Simpli.fi, emphasised locally relevant messaging and connecting digital intent to real-world action. Deal structure, pricing, minimum spends, measurement partners and creative specifications were all absent from the announcement.

Set against that instrumentation, a survey published on September 14 measures what the instruments are competing with. Omdia found that phone use while watching television has risen fourteen points among Americans aged 55 to 64 since 2023, from 42 per cent to 56 per cent. The 45 to 54 bracket moved from 62 to 73 per cent, and the 35 to 44 bracket from 69 to 76 per cent. Nearly three in four American television viewers now split attention with a handset. María Rúa Aguete, global head of media and entertainment at Omdia, reduced it to a sentence: "We don't have a content problem. We have an attention problem." Her second line was the strategic one: "The future isn't TV versus mobile. It's TV and mobile." The release discloses no sample size, no fieldwork dates and no survey method, and does not establish whether the 2023 and 2026 readings come from identical panel designs or identical question wording. A fourteen-point movement measured by undisclosed means is exactly the kind of figure that will be quoted in planning decks for the next year.

Which is the pattern underneath the week. A court compelled a platform to document its auction and accepted a six-year monitor as the price of leaving its exchange intact. Publishers signed licensing agreements whose terms they cannot compare with anyone else's. A platform reported enforcement in the billions and left the residual unstated. A regulator proposed to see product plans before launch rather than after harm. And a television measurement claim circulated with no methodology attached at all. The compelled disclosures are getting more detailed. The voluntary ones are not.

Also noted

  • September 16: Google cut the follower threshold for Search Profiles from 35,000 to 10,000 across YouTube, Instagram, X and TikTok, redesigned article views with larger headlines and optimised thumbnails, and now allows up to ten profiles per eligible brand managed from a single login (Search Engine Roundtable).
  • September 16: Google documented its multi-source conversions beta, which links an existing website conversion action to a backend source such as a CRM or order database through Data Manager or the Data Manager API, recovering conversions lost to browser restrictions and ad blockers and allowing real-time value adjustment (Search Engine Roundtable).
  • September 17: livestreams generated 94 per cent of United States luxury resale revenue on TikTok Shop year to date, against a 400 per cent year-over-year rise in category gross merchandise value, with Fashionphile attributing a quarter of its United Kingdom business to the platform (Digiday).
  • September 15: a 31-page VAB report projects podcast advertising at $4.69 billion by 2030, or 48 per cent of United States digital audio spending, up from $3.40 billion and 42 per cent in 2026, while noting that streamed podcasts inside closed platforms remain unmeasured until an IAB guideline planned for 2027 (PPC Land).
  • September 15: Mediaocean launched a venture arm offering AI startups capital plus integration into systems carrying $200 billion in annualised advertising spend and more than 100,000 users, with no fund size, cheque size or portfolio company disclosed (PPC Land).