Television spent a decade being described as the last analogue holdout in media buying. The week ending September 12 suggests the description is expiring in two directions at once. On one side, the mechanics of buying local broadcast inventory are being handed to software agents that negotiate with each other. On the other, the free ad-supported streaming channels that grew fastest over the past three years are watching their prices collapse because nobody outside the platforms can see who is watching them.

Both things happened within forty-eight hours. Neither is a story about audience growth. Both are stories about who controls the numbers.

Agents arrive at the local station

Magnite and ITN announced on September 10 that ITN's seller agent will connect into Magnite's Orchestration layer, with full integration scheduled for the fourth quarter of 2026. The plain description of what changes: an AI agent can generate impression-based forecasts and media plans across local linear and connected television in a single workflow, and a buyer can issue one request for proposal covering both.

The supply behind it is not marginal. ITN's network reaches 75 broadcast ownership groups and more than 1,100 stations. The US local television advertising market was worth roughly 21 billion dollars in 2025, a pool that has resisted programmatic plumbing largely because the paperwork around it never went away. Avails requests, phone calls, revised spreadsheets, make-goods negotiated by email.

Matt McLeggon, SVP of Advanced Solutions at Magnite, framed the problem as friction rather than demand. "Local linear TV remains a powerful medium for reaching audiences but the operational friction of buying local linear TV inventory has historically limited its growth," he said. Craig Sulema, Chief Investment Officer at ITN, put the same point in operational terms: "we are eliminating weeks of back-and-forth operational drag."

Weeks to hours is the claim. What supports it technically is a stack Magnite has been rebuilding for a year. ClearLine, the curation and activation platform that carries execution here, was rebuilt in October 2025 on SpringServe ad server infrastructure and technology from the streamr.ai acquisition. Orchestration, the neutral coordination layer the ITN agent plugs into, opened in beta in June 2026. Execution routes through SpringServe. Agent-to-agent communication runs on the Advertising Context Protocol, the open standard that has spent most of this year being argued over in working groups rather than used in production.

Joe Cerone, founder of Cerone Advisory Group, described the addition as "another step closer to aligning local linear with the technology, workflows and expectations that define modern media buying." Dano Ehler, Chief Revenue Officer and co-founder of Digital Marketing Group, is named as an early partner.

The numbers that do not appear in the announcement are the interesting ones. No performance metrics for the ITN integration itself. No pricing. No forecast-accuracy validation method. No comparison of CPMs achieved by agent-planned buys against conventionally planned ones - which matters, because a DataBeat measurement in June 2026 found conventional buyers still holding a 13.4 percent CPM advantage over AI agents. Magnite's own chief executive has put a ceiling of 700 million dollars on agentic spend flowing through the company by 2027, and an EMEA agentic test produced a 70 percent reduction in connected TV setup time. Speed is the demonstrated gain. Price is not.

Supply expanded on the same day from a different direction. Teads opened Whale TV's home screen inventory globally through Teads Ad Manager, extending a partnership that had run only in Europe. Whale TV powers 47.5 million monthly active televisions across more than 400 TV brands; Teads counts 500 million addressable TV devices across its combined manufacturer partnerships, now reaching more than 100 countries. Since 2023 the partnership has delivered over 6,000 HomeScreen campaigns, up from 5,000 in June 2026 - roughly a thousand campaigns in a quarter.

The formats are masthead placements, native display and CTV in-stream units that expand to full-screen video on interaction. Simon Klein, Global SVP Commercial Strategy CTV at Teads, described the demand as structural rather than creative: "Advertisers are looking for global CTV scale without the complexity of managing separate platforms and regional workflows." Chris Hock, VP and Head of Whale TV Ads, said the arrangement "has grown from a strong regional collaboration into a global opportunity for advertisers."

Financial terms were not disclosed. Neither were CPM rates, regional device distribution, nor expected campaign volume. The pattern repeats across every television announcement of the week: the supply is specified, the money is not.

The channels that will not show their homework

Which leads to the sharpest piece of the week, and the one carrying the least new data. Writing in AdExchanger's On TV And Video column on September 11, Brian Cullinane, Chief Commercial Officer at VideoElephant, argued that FAST channels are being discounted because their operators refuse to hand over the audience data that would justify the price.

The growth figures are not in dispute. Comscore recorded a 43 percent year-over-year increase in US FAST viewing hours through August 2025, reaching 1.8 billion hours. Amagi measured 55 percent global growth through mid-2026. Wurl logged a 25 percent increase in session duration. Sports programming on FAST grew 30 percent in the first quarter of 2026, and local news viewing over OTT rose 69 percent year over year. eMarketer projects 131 million US FAST users in 2026, which would be 54 percent of all connected TV users.

Against that, the economics moved the other way. Revenue per viewing hour fell from 0.18 dollars to 0.08 dollars. CPMs that once sat around 30 dollars now trade in single digits. Viewing doubled; the price per hour of it halved.

Cullinane's diagnosis is that platforms are "sitting on audience data that would validate their inventory, and they are choosing not to share it." Nielsen, he notes, cannot measure smaller FAST channels because the operators do not provide access. Device-level confirmation of what content and which ads actually rendered remains unavailable to third parties, even though the technology to produce it exists. Agencies that cannot verify a viewing claim discount the inventory holding it, and the discount compounds.

It is a self-interested argument - a supply-side executive asking platforms to open their books - and it should be read as one. It is also the same structural complaint that has followed walled-garden reporting for the past five years. The difference in FAST is that the price collapse is measurable and recent.

One measurement layer did expand this week, though not the one the column asks for. Nielsen added France as the fifth market for connected TV coverage in Ad Intel, announced September 10, tracking a forecast 1.5 billion euros of annual French CTV advertising investment across Amazon Prime Video, Netflix, HBO Max, Disney+, YouTube, DAZN, Canal+, TF1, M6+ and FranceTV. The rollout began with Germany in August 2025, added the United Kingdom in September 2025 after a July 14 announcement, took in Australia from an August 13, 2025 announcement, and covers Italy.

Inam Mahmood, General Manager EMEA at Nielsen, said the product lets clients stop planning "in silos" and see "where rivals are placing their bets." That is competitive spend intelligence - declared advertising investment, by platform, by advertiser. It is not audience measurement, not attribution, and not device-level verification. Ad Intel can say how much money a competitor pushed into Canal+ last month. It cannot say whether anyone was in the room.

So the television picture at the end of this week has three layers moving at different speeds. Transaction mechanics are automating quickly. Supply is expanding quickly. Independent verification of who saw what is expanding slowly, in specific markets, and only for the categories platforms agree to expose.

Sponsorship stops being a proxy

The sports announcements of the week attack the same gap from the sponsorship side, where the measurement problem has always been worse.

GumGum launched Sports Media on September 9, a three-product suite built on Relo Metrics computer vision and VideoAmp identity resolution. Relo detects branded exposures during broadcasts - logo prominence, share of screen, duration, placement. VideoAmp's VXA resolves that exposure to household-level viewership. What comes out is an audience segment defined by having actually seen a logo on screen, which can then be targeted across web, connected television and online video.

Three products sit on top. Replay re-engages verified viewers of a brand's own placements. Counterplay targets households that saw a competitor's logo. Sports Audiences reaches verified sports fans through VXA. Coverage spans the NFL, NBA, WNBA, MLB, NHL, MLS, NWSL, F1 and NCAA Football.

Danielle Zazula, SVP Revenue at VideoAmp, was blunt about what the category has been doing until now: sponsors "spend billions on sports sponsorships and have largely had to defend that spend with proxies instead of proof." Hailey Denenberg, SVP and Head of Strategy and Business Operations at GumGum, described the mechanism as matching "what appeared on screen to the households that were actually watching, and turn that into an audience they can reach." Dylan Cox, VP of Data Strategy and Operations at Relo Metrics, said the purpose is to reach verified fans "as they continue their digital journeys beyond the sporting event itself."

No performance figures accompanied the launch. No pricing. No addressable audience size per league or per broadcast, which is the figure that determines whether Counterplay is a usable tactic or a demonstration. A household that saw a competitor's board during a Thursday night NFL game is a precise definition; how many such households can be reached at what frequency is unstated.

A day later, Genius Sports, Equativ and StackAdapt connected live game data to creative decisioning inside StackAdapt's demand-side platform. The chain runs in three stages: Genius Sports classifies live game events through its Moment Engine, Equativ resolves creative dynamically at insertion, StackAdapt activates the campaign. Buyers get two routes - the Genius Deal Library, which overlays live scores onto existing assets, and Custom High-Impact, which adapts the message itself to what is happening in the game.

Genius Sports covers more than 400 leagues across over 150 countries. The supporting research claims 84 percent of sports fans use a companion screen during broadcasts and 91 percent are more likely to pay attention to advertising carrying relevant live scores, though the methodology behind that second figure was not published. A precedent exists: an earlier Magnite integration with Genius data added 5,800 live sports advertisers and a 56 percent increase in spend.

Josh Linforth, Chief Revenue Officer at Genius Sports, described the value as timing rather than targeting - "live sports give marketers another powerful signal: when that consumer is most engaged." Frank Maguire, SVP Product Marketing at Equativ, called live sports "the most powerful emotional environment in media." Greg Joseph, VP Inventory Development at StackAdapt, acknowledged the operational history: "Live sports creates moments of extraordinary attention, but historically it hasn't been easy."

Pricing, participating leagues, latency requirements and performance metrics from the integration itself all went undisclosed. That is four sports and television announcements in three days, all of them specifying mechanism and withholding outcome.

Numbers the buyer cannot check

Google introduced a metric on September 10 that puts a figure on something advertisers have been asked to accept as a general principle for four years. The Data Strength Uplift Metric quantifies the additional conversions recovered by an advertiser's first-party data setup, displayed inside the Google Ads interface alongside campaigns. Nipoon Malhotra, VP of Ads Analytics, Insights and Measurement, announced it.

Converting a category argument into an account-specific number is a meaningful product change. Enhanced conversions, consent mode and Data Manager have all been sold on the premise that better-matched data recovers measurement lost to browser restrictions and consent gates. The premise was plausible and unquantified at the account level. Now it carries a figure.

The methodology behind the figure was not published. Google cited 14 percent, 11 percent and 26 percent uplift results, all drawn from internal testing rather than independent research. The formula, the counterfactual construction, the separation of causation from correlation, the confidence intervals, the inputs beyond the phrase first-party data setup - none disclosed. The metric quantifies recovered conversions using the same platform that gains when advertisers supply more data, and no external party can reproduce it.

That structure - a number produced and audited by the party it benefits - is the thread running through the FAST discount, the withheld CPM data in the television announcements, and the sponsorship proxies GumGum is trying to replace. It also shapes the week's most concrete study of AI-driven shopping behaviour.

Similarweb published its State of Ecommerce 2026 report on September 10, with forecasting from Statista and analysis led by Daniel Reid, Principal Insight Analyst for Consumer Goods and Retail. The headline behaviour is measured on US desktop data from July to December 2025, looking at seven-day windows after an AI assistant named a brand.

When Capital One was recommended, 14.2 percent of those shoppers visited Capital One within seven days, against 3.8 percent who went to American Express. When American Express was recommended, 7.2 percent visited AmEx and 3.1 percent went to Capital One. Kayak recommended produced 12.0 percent to Kayak and 3.4 percent to Skyscanner; Skyscanner recommended produced 9.5 percent to Skyscanner and 7.6 percent to Kayak. In beauty, Sephora recommended sent 7.9 percent to Sephora and 3.3 percent to Ulta, while Ulta recommended sent 7.6 percent to Ulta and 4.6 percent to Sephora.

The asymmetry is the finding. Being named does not merely lift the named brand; it suppresses the alternative by a factor that varies sharply by category and by brand. Capital One converts a recommendation almost twice as efficiently as AmEx does. Skyscanner recommended still leaks 7.6 percent of attention to Kayak, while Kayak recommended leaks only 3.4 percent back.

Context for the volume matters. AI referrals to ecommerce sites grew 203 percent year over year while direct traffic grew 1.2 percent, and AI-assisted journeys account for 11.4 percent of sessions. Sessions combining AI and search convert at 23.0 percent, the best-performing pattern in the dataset. Reid's summary: "Consumers are not switching tools - they are stacking them." The absolute referral volume remains small, and the report says so plainly.

The attribution problem is structural rather than technical sloppiness. A recommendation generates a visit without generating a click, a session or a referral record; the shopper receives the suggestion in one channel and completes the purchase in another with no shared identifier. Google Analytics only added a dedicated AI assistant channel grouping in May 2026. Similarweb and NIQ have a joint measurement product scheduled for the fourth quarter of 2026 covering ChatGPT, Gemini, Google AI Mode, Perplexity and Claude, which is an admission of how little the current stack captures.

The comparative record cuts both ways, and the report includes it. Ahrefs found on February 4, 2026 that ChatGPT sends roughly 190 times less traffic than Google while handling around 12 percent of query volume. An October 24, 2025 study found ChatGPT referrals underperforming traditional channels on conversion rate and revenue per session. Criteo measured the opposite in May 2026, with ChatGPT conversion approaching twice traditional search in consumer electronics, lifestyle and home. SISTRIX put the cost of AI Overviews in Germany at 265 million organic clicks a month, with position-one click-through falling from 27 percent to 11 percent where a summary appears.

Elsewhere in the same dataset, global B2C ecommerce revenue is forecast above 4.9 trillion dollars by 2030, a 27 percent increase from 2026. The United States accounts for 1.5 trillion dollars in 2026 on 23.9 percent growth, the United Kingdom 159 billion dollars on 16.6 percent, Germany 133 billion dollars on 21.2 percent. Ecommerce app sessions are growing 1.3 times faster than web visits, and 86.5 percent of US consumers shop primarily on a smartphone or tablet.

Germany's oldest profiling machine goes to court

The week's regulatory story concerns a database older than digital advertising and governed by the same statute.

noyb confirmed it will seek a court injunction against SCHUFA after the credit bureau refused to sign four cease-and-desist declarations. The sequence is documented. NDR and Süddeutsche Zeitung published investigations of SCHUFA's historical archive on July 15, 2026. On August 26 SCHUFA said it would defend its storage practices in court, and noyb sent its cease-and-desist letter while opening a class action interest list. The signing deadline expired September 9. On September 10 SCHUFA published a written reply declining all four declarations.

The scale is the whole point. SCHUFA holds records on 69 million people, effectively the adult population of Germany. noyb estimates immaterial damages at roughly 500 euros per affected person, and Max Schrems, chair of noyb, said interest in the class action is "going extremely well." He called SCHUFA's arguments "utterly grotesque" and said the company behaves "as if it stands above European law." Roughly 1.6 million people a year are reported to receive incomplete responses to access requests; SCHUFA issued 2 million consumer disclosures in 2025.

noyb's demands rest on Article 15 and Article 14 of the GDPR - access and transparency. SCHUFA's defence rests on Article 40, which contemplates sector-specific codes of conduct, and Article 6(1)(f) legitimate interests, alongside Section 31 of the Federal Data Protection Act covering the scientific basis of scores.

Its four arguments are worth stating precisely, because they are the arguments every large data holder will reach for. First, the archive is not secret but necessary for regulatory oversight, audit documentation and verification by data protection authorities. Second, credit forecasting models are built from historical analysis, and without it the quality checks banking law requires become impossible. Third, retention follows the Code of Conduct on examination and storage periods agreed with federal and regional data protection authorities. Fourth, SCHUFA already provides processing information beyond the statutory minimum, including twelve months of score history with explanation, and will exceed the new requirements when Section 37a of the Federal Data Protection Act takes effect on November 20, 2026, obliging historical score data to appear in data copies.

Three adjacent rulings frame where this lands. A Wiesbaden court in January 2026 ordered individual explanation of the factors behind SCHUFA's risk scores, in a case involving an 85.96 percent figure, pushing toward specific rather than generic disclosure. A Federal Court decision in November 2025 permitted telecommunications companies to transmit positive contract data to SCHUFA for fraud prevention under Article 6(1)(f), which widens the profile rather than narrowing it. A Berlin court in August 2026 and the Austrian data protection authority in September 2025 both questioned when automated scoring triggers automatic consequences.

Credit scoring is the oldest continuously operating consumer profiling system in Europe, and it runs under the same regulation that governs ad targeting, frequency capping and measurement. Whatever the Federal Court eventually says about how long a profile may be retained and how much of it must be shown to its subject will be read closely by everyone holding a large archive of behavioural records, advertising included.

Also noted