Ten items, and the week has a shape: two of the largest consumer settlements in American history closed with nobody admitting anything, a measurement currency is changing its own math five days before football, and three companies spent the week describing a business their own filings describe differently.

The agent announcements were the noise. The filings were the signal.

Connecticut AG William Tong calls it the largest settlement in American history with a single defendant. Meta signed a week into trial and conceded nothing. $12.19 billion is guaranteed, paid to 51 attorneys general over ten years. Another $4.91 billion is contingent on Snapchat, TikTok and YouTube each accepting comparable terms, and if they do, the remedies tighten: the daily cap on under-18 use of Instagram and Facebook drops from two hours to one, and the overnight blackout widens from midnight-to-6am out to 10pm-to-7am. School Mode kills push notifications 8am to 3pm on weekdays. Judge Yvonne Gonzalez Rogers called it a good step forward and has not entered the consent judgment.

The trades filed this as child safety. BMO's Brian Pitz filed it as ad load, and he's the one reading it correctly. Usage caps, notification blackouts and overnight shutoffs are inventory reductions with a safety label on them, and they are the first ad-load remedies any American regulator has ever extracted from Meta. New Mexico and Florida did not settle.

The Richard Foster whistleblower suit stopped being an HR dispute this week. Per Foster's amended complaint, a Sony Pictures internal investigation found that on Chinese media spend in 2024, WPP generated roughly $460 million in rebates, returned about $110 million to clients, and retained approximately $350 million. Foster's broader estimate runs to $1.5 to $2 billion improperly retained out of $3 to $4 billion in rebate-driven deals over five years. He is seeking at least $100 million.

WPP calls the claims baseless and without merit, called Foster a disgruntled employee angling for an exorbitant payout, and suggested he obtained the Sony material by improper means. That last argument is the interesting one, because it is a complaint about how somebody got hold of a document WPP is simultaneously insisting says nothing. Then WPP asked a New York court to seal it. You do not fight that hard to hide a document that exonerates you.

Three separate stories, one company, one week. TTD shipped Kokai Zuma, which quietly deprioritizes the periodic-table interface that was the centerpiece of the 2023 Kokai launch. Adweek got a leaked roadmap deck showing Ask Koa, a conversational front door routing buyers to six backend agents, moving from alpha to closed beta. Lift studies now launch in one click and return in 48 hours instead of five to seven days. There's a NielsenIQ tie-up, a new GeoLift function, and a promise to open the AI layer to third-party tools. Also this week: Aravind Chandrasekharan, SVP of Engineering, twelve-plus years in, last day August 28, leaving for Branch as chief product and technology officer. Also this week: the stock closed $13.03, down 24% in a month and roughly 65% year to date, against a 52-week high of $56.39.

The number to watch is the one that stopped appearing. The "85% of clients on Kokai" figure being recycled in this week's coverage is from Q3 2025 earnings. TTD has not disclosed a 2026 adoption percentage. It reported JBP counts instead. When a company swaps out its adoption metric mid-rollout, the swap is the disclosure.

Seven Big Data + Panel currency changes land August 31, five days before the fall season: co-viewing enhancement using wrist-worn wearables, integrated weighting, latency-adjusted DASH universe estimates, Hispanic methodology, an ACR monitored-tuning adjustment, and two more. Paul Ballew, NFL SVP of data and analytics, put his name on the objection, saying the changes will muddle measurement for the coming season and that what the league has seen so far is negative for football and for sports generally.

Underneath the fight is a gap nobody is pricing. Nielsen's Q2 Gauge puts ad-supported TV at 71.5% of total viewing, down 1.3 points from Q1. Streaming is 48.2% of ad-supported viewing on a total-audience basis but 44.4% among Persons 18+, and Nielsen is shifting emphasis to the 18+ cut. That is a 3.8-point hole in every media plan built on the other number. Meanwhile the currency is spending $2.15 billion to buy DoubleVerify, closing by Q1 2027. The company that counts is acquiring the company that verifies, in the same quarter the biggest rights holder in America says the counting is broken.

The best original reporting anyone filed this week. AdExchanger's Allison Schiff got co-founder Ashish Chordia on the record about the endgame for Alphonso, LG Electronics' CTV ad business. A court-mandated tender offer at $118 per share closes September 11. A non-binding Koch Equity Development LOI at roughly $200 per share, about $1 billion all in, was submitted August 5. There is a confidential S-1 filed in September 2025 and a possible Comcast joint venture through Atairos. Chordia has a $4.5 billion damages suit pending in California and a Delaware case going to trial in June with potential direct LGE liability. Revenue projections split along party lines: $850 million per LG management, roughly $1 billion per Chordia, on flat headcount of about 100 people.

Read this one with tongs. Every figure above traces to a single interview with a party adverse to LGE. No filing, no second outlet, no corroboration, and the man supplying the numbers is suing for $4.5 billion. It is still the most interesting story of the week. It is still one man's account, and this desk is working the LG thread separately.

ChatGPT Ads went live across 31 European markets on August 24, six months after the US pilot and the largest expansion yet. Germany, France, Spain, Italy, the Nordics, Netherlands, Austria. Buying runs through holdco partners, not self-serve. Two days later Adweek reported OpenAI is testing exclusion targeting with a select group of advertisers, no launch timeline attached.

The reason for the exclusion test is buried in the same story: buyers say they can't describe audiences to the system, can't control placement, and can't see where their ads land. Research from SE Ranking, Emberos and Seer Interactive found ChatGPT ads carry minimal relevance to the surrounding conversation. Microsoft is separately eliminating the Max CPC field from new Max Conversions and Max Clicks campaigns on October 1, and OpenAI preselects "Maximize results" while stating outright that it does not guarantee delivery against CPA, CPC or ROAS targets. Unmeasurable placement, no negative targeting, no bid control, 31 new countries. Ad tech has run this configuration before. It was called the ad network era, and it is the business I was in when it ended.

X shipped an Ads Model Context Protocol server letting any third-party LLM build and manage campaigns. Twenty-three X Ads tools exposed, ten of them carrying write access to live ad accounts. Meta, Pinterest, TikTok and Snapchat already offer MCP. X is the one that put write access on the table with the fewest words about it.

Nobody in the trade press asked the question that matters, so: who pays when an agent with write access spends the budget wrong. Not the vendor, per every MSA currently in market. The answer is the advertiser, every time, and the paused-by-default rail protects exactly one moment in the lifecycle of a campaign that then runs unattended.

DOJ announced August 21. $300 million immediately, $100 million more on entry of an order vacating the 2019 Musical.ly consent decree. It resolves the 2024 COPPA complaint alleging under-13 users routinely walked around Kids Mode. TikTok and ByteDance admitted no wrongdoing, and DOJ said the deal resolves allegations only, with no determination of liability.

Read the second tranche again. TikTok is paying an extra hundred million dollars for the removal of a six-year-old order, which means TikTok has priced what that order costs it and decided the number is north of nine figures. The vacatur is the asset here. The fine is the shipping cost.

California AG Rob Bonta cancelled a scheduled settlement meeting in the 12-state challenge to Paramount Skydance's $110 billion Warner Bros. Discovery acquisition, accusing the company of misrepresenting the discussions and demonstrating a lack of good faith. Bonta wants structural remedies, meaning divestitures, and publicly rejected David Ellison's pledge of 30 theatrical releases a year as unenforceable. The complaint's arithmetic: the merged company takes 27% of wide-release theatrical distribution, leaves three distributors controlling 75% and four controlling 86%, and takes 27% of basic cable channel licensing across 45-plus primary channels.

The clock is the leverage. Trial is set for March 2027, the merger must close by June 4, 2027, and Paramount starts owing a $650 million per quarter ticking fee on October 1. Paramount's response is to demand a $1.8 billion bond from the attorneys general, which is the corporate equivalent of asking the referee to post collateral before he throws the flag.

The quietest story of the week and probably the largest. People Inc. reported session-based revenue down from 61% to 57% of digital revenue year over year, core sessions down 22%, Google search traffic down 40% year over year, and Google now 21% of traffic against 25% the prior quarter. Non-session revenue is up 16%. JWX CEO John Nardone, who placed the first paid internet ad in 1994, put the trend line more bluntly: search referrals to publishers have gone from roughly 70% of visits three years ago to under 20% today, and he calls it structural rather than cyclical. Brainlabs client data across 54 advertisers from January 2025 to April 2026 shows organic traffic down 10.5%, from 140.1 million to 125.4 million sessions, while AI referrals rose 163% to about 200,000 monthly sessions.

Everyone reported the 163%. Nobody reported the ratio. The replacement channel is running at roughly one-seventieth the volume of the loss. And the phrase every publisher reached for this week, the pivot to non-session revenue, deserves translation: non-session revenue is licensing, licensing is a one-buyer market where the buyer sets the price, and the seller's traffic is falling 40% a year. USA Today Co. is already reformatting its content for machines, metadata and markdown and templates invisible to humans, while blocking roughly 99% of self-identified AI bots and whitelisting the ones that pay. Time and The Economist got there first. That is not a pivot. That is a liquidation with better branding.

Prebid named OpenX CTO Joel Meyer chairman after its president, chairman and product director all left in May, and now has to stay neutral between AdCP, AAMP and MCP. IAB Tech Lab says the rival agentic protocols overlap on 13 functions.

Political spend hit $6.1 billion through August 24 against an $11.6 billion full-cycle projection, the largest midterm ever. Broadcast $3.12B, CTV and streaming $1.14B, digital $937M, cable $770M.

One in four CTV viewers is a ghost. A report this week put roughly 25% of CTV impressions in the potentially-invalid bucket, in the same seven days CTV took $1.14 billion of political money. Single-vendor report, unaudited, and worth someone's log files.

The great Meta migration. Ex-Meta ad staff are now roughly 10% of OpenAI's workforce, with about 700 at TikTok. Reuters separately reported Project OT, Meta's plan to cut headcount and offload service to agentic tech, which is why agencies spending hundreds of millions currently reach a chatbot.

Maryland's digital ad tax was struck down as unconstitutional on Google, Peacock and Apple's challenge, which puts the Illinois, Utah and Washington proposals in jeopardy.

Butler/Till ran an autonomous CTV campaign with PubMatic for Geloso and claims 80% lower supply-chain and tech costs, 40% more impressions on the same budget, 98% video completion and under 1% waste. Four unaudited numbers, one agency, one campaign. Somebody should ask for the log files.

How this was reported: ten items selected from roughly 180 stories across five trades for the week of August 21 to 28, with every figure traced to a primary source where one exists, meaning court filings, AG releases, DOJ announcements and company disclosures rather than trade summaries. Where a number comes from a single interested party, it is flagged in the item and not carried in the headline. The Alphonso figures in item five all originate with one adverse party and none are independently confirmed. Companies named here were not given advance review of this issue. ADOTAT takes no payment from any company named above and runs no paid placement in editorial; I ran a competing ad network in the 2000s and have known executives at several of these companies for two decades. Corrections get a dated update note at the bottom of this issue and an entry on the corrections page.