Google's September spam update ends after 13 days as manual actions rise

Enforcement became visible this week, and nobody involved explained very much.

Google closed its September spam update on October 8 after thirteen days and sixteen hours, the longest such run in more than a year, without naming a single policy it had targeted. Two days earlier it had published a pipeline for ingesting forum and social posts whose payload specification stays private until a platform is accepted. On the same page where it tells publishers what makes content worth ranking, it added a sentence calling fabricated author profiles a form of deception. Three disclosures, all partial, all from the same company, inside four days.

The week's other stories turn on the same hinge. Spain began requiring employers to tell each worker, in writing and in concrete terms, when an automated system helps decide pay or dismissal. Florida sued a router maker over what its security pages claimed. Giorgia Meloni filed four seconds of her own voice as an EU trade mark. A Catalan regulator shipped free software for scoring what an AI system does to fundamental rights. And a small bulk-upload tool published the share of its monthly submissions that arrived from machines rather than people, which is a number no larger platform has volunteered.

Amazon, meanwhile, cut roles it would not count while raising capital spending it would.

Thirteen days, sixteen hours, and no policy named

The September 2026 spam update finished rolling out on October 8, and the arithmetic is the story. Google logged the opening note on the Search Status Dashboard at 09:15 Pacific on September 24, saying the update applied worldwide and to all languages and that the rollout might take up to two weeks. The completion note carries a timestamp of 01:37 Pacific on October 8. That is thirteen days, sixteen hours and twenty-two minutes, or roughly 98% of the stated allowance. The dashboard header gives 01:00 as the close, a thirty-seven-minute gap on a single page, which shortens the run to thirteen days, fifteen hours and forty-five minutes. Search Engine Roundtable logged the finish at 4:37 a.m. Eastern and the start at about noon Eastern, a few minutes off the dashboard's own conversion.

Set against the year's other runs, the length is conspicuous. March 2026 opened on March 24 and closed on March 25, about nineteen and a half hours. June ran June 24 to June 26. August ran August 18 to August 21. September took roughly seventeen times March's duration. Among the fifteen spam updates Search Engine Roundtable has documented since June 2021, only five ran longer: the July 2021 link spam update at twenty-nine days, August 2025 at twenty-seven, the December 2022 link spam update at nineteen, and October 2023 and March 2024 at fifteen each. September 2026 ranks sixth. At a talk this month Gary Illyes showed internal timing data putting spam rollouts at one to two days, which makes the September run close to seven times the top of Google's own range.

On LinkedIn, Google Search Central called it "a normal spam update" covering all languages and locations. Barry Schwartz, who wrote the Search Engine Roundtable report, said he did not know why that label fits a two-week rollout rather than a two-day one. Google declined to say what share of queries was affected. It named no targeted policy, said no policy changes accompanied the release, said link spam was not the target, and said periodic refreshes would follow. Recovery, it repeated, can take many months.

What was in force at release is a matter of record even if what was deployed is not. The live policies cover expired domain abuse, scaled content abuse, site reputation abuse, doorway abuse, back button hijacking, and fake or undisclosed incentivised reviews. SpamBrain, described in the documentation as an AI-based spam-prevention system, is named as a component refined over time; its architecture, training data and thresholds have never been published. Since May 15 2026 the spam policies have covered generative AI responses in Search, including AI Overviews and AI Mode, which makes September the third spam update under that extension. And since August 30 2026, manual actions for site reputation abuse no longer affect rankings for searchers inside the European Economic Area, so this was the first confirmed ranking update after a carve-out that touches manual enforcement only, leaving the automated systems untouched.

Algorithmic demotions arrive without notice. That is the structural difference from a manual action, which appears in Search Console and can be appealed. A site demoted by this update learns about it from traffic data or a third-party tracker, which is why the week's volatility readings matter more than usual. Schwartz recorded three windows of elevated movement, on September 25 to 27, on September 30, and on October 4 to 6, while noting that the dates assume Google made no other changes, something he could not verify. His aggregate view drew on fourteen tools: AccuRanker, Algoroo, AWR, CognitiveSEO, DataForSEO, Mangools, Mozcast, Semrush, Serpstat, SimilarWeb, Sistrix, Wincher, Wireboard and Zutrix. PPC Land logged elevated movement from about September 22, which Schwartz said was not the update.

Measuring any of it has got harder in the same period. Google confirmed on August 26 that it routes result links through google.com/goto addresses, and Nozzle measured 500 to 1,000 requests to resolve a single five-page ranking. The Search Console generative AI report, which reached every property on August 31, shows impressions but neither clicks nor queries. Length has not reliably predicted impact either: during the twenty-seven-day August 2025 cycle, SISTRIX measured changes below typical daily fluctuation.

Two commercial details sit underneath. The rollout crossed the end of the third quarter on September 30, so any quarterly organic session count, affiliate revenue line or lead total now contains part of it. And since December 23 2024 Google Ads has disapproved ads pointing to destinations removed from Search by manual action, a rule that algorithmic ranking drops do not trigger on their own. Google has not said why this one took a fortnight. On March 31 2026 John Mueller said several ranking components are updated step by step, and the previous edition of this update was reported at the two-week window rather than at its close. Marie Haynes said in an October 2 video that she expects a core update soon while allowing that it might not arrive. In March 2026 a core update began three days after the spam update started.

Manual actions reappear, and so does the word deception

While the automated run was finishing, the manual kind came back into view. Some SEOs believe Google has issued more manual action penalties in recent weeks than in the weeks before, a shift Schwartz dated to several weeks ago and described as growing more recently. His framing is historical: manual action complaints were common a decade or more ago, then thinned out as the algorithms took on the work.

The evidence is anecdotal and the sources say so. Haynes, who has done manual action consulting for what she called a couple of decades, wrote on X on October 7 that she has had "a big increase in requests for help with manual actions lately." The sites she sees are mostly hit for thin content, pure spam and scaled content, and she described most of them as "legitimate long standing businesses." She told Schwartz the requests went from essentially none to several a week, well below pre-Penguin volumes but notable because she no longer advertises the service. Daniel Foley Carter, posting the same day, attributed the increase to SEO shortcuts, scaled content abuse, and businesses cutting SEO budgets on the assumption that AI could do the work. He put the damage at 90% or more of traffic lost overnight in some cases, and listed roughly fourteen action types, among them back button hijacking, third-party spam, user-generated spam, spammy free hosts, structured data issues, unnatural links in both directions, thin content, cloaking and sneaky redirects, major spam problems, cloaked images, hidden text and keyword stuffing, AMP content mismatch, sneaky mobile redirects, and site reputation abuse. He drew a distinction worth keeping: manual actions come out of quality rater review, while automated spam systems, user spam reports and routine sweeps are separate triggers.

Schwartz went through the Google Webmaster Help Forums by hand and collected twelve recent threads. The texture is repetitive. One poster removed 4,000 URLs and still carries a thin content action. Another cleaned more than 6,000 URLs and has had three reconsideration requests rejected. Three separate threads report a major spam problems action scoped to a /blog/ directory. A transparency action against news.ks.ua has produced no outcome after repeated requests. Schwartz also asked ChatGPT and Gemini to read the forums; both reported more complaints without a very large spike, and he noted that the two charts they produced disagree with each other. Google has said nothing about any of it.

The documentation moved in the same direction. On October 2 Haynes published a twenty-seven-minute video reporting that Google had edited its help page on creating helpful, reliable, people-first content, and the edit added a definition of main content, four attributes for judging it, and a warning about fabricated creator profiles. The new sentence says that fabricating creator profiles, with AI headshots, invented names and false credentials, to make content appear written by human experts is a form of deception, and that this renders pages untrustworthy to readers and to automated quality systems. The wording is known only from Haynes reading it aloud; the captured version of the page shows the opening section alone.

Her reading of main content tracks the Search Quality Rater Guidelines, January 23 2025 edition, section 2.4.1, and covers primary text and media, interactive features such as calculators, user contributions such as comments, tabbed or expanded sections, and titles and headings. Google's text reportedly accepts content behind tabs without penalty, which did not stop one client moving product video reviews out from behind an additional information tab that visitors rarely opened. The four attributes are effort, originality, talent or skill, and accuracy. Effort is defined as the extent of human work, with original analysis and a custom interactive game map as high-effort examples, and feed-generated pages and large volumes of generative AI text without human oversight as the low end. Haynes counted the word effort 120 times in the rater guidelines and the word trust 191 times, and read the text as accepting AI assistance where human work accompanies it. Accuracy carries stricter standards for topics affecting health and wellbeing, which she tied back to the August 1 2018 Medic update.

None of this establishes a new penalty or a new detection method, and the material does not address whether Google can identify a fabricated profile at all. Bylines, Haynes noted, are not a ranking factor, though readers like them. What the three threads share is a direction of travel: an enforcement system that publishes less about what it did, paired with documentation that increasingly describes what a human contributed, assessed by raters whose judgments feed actions that do come with notice.

A pipeline for forum posts, with the specification withheld

On October 8 Google published documentation for the UGC Fresh Data Program, a route by which approved forums and social platforms can send user posts and engagement signals directly to Google, logged in the Search Central changelog and picked up the same day by Search Engine Roundtable. Google calls it "a specialized pipeline designed exclusively for trending UGC content," separate from the Indexing API and from ordinary crawling, and states plainly that the program "doesn't guarantee that content will appear in Search."

The timing problem it addresses is quantified, if secondhand. At Search Central Live Deep Dive Europe in Barcelona on October 2, Illyes showed figures that an attendee recap put at about twenty hours to discover a new URL and about thirty days to refresh a known one. Against a forum thread that matters for six hours, thirty days is not a crawl budget question so much as a different product. The existing Indexing API is no help, being limited to JobPosting pages and BroadcastEvent markup inside VideoObject, with a default quota of 200 for onboarding and testing.

Six eligibility criteria are published. The platform must primarily host user-generated content, social posts or forum discussions, with each post on a dedicated page at a stable URL rather than a profile or feed. It must have high volume and a significant user base, with no numeric threshold given. Content must sit on public pages reachable by Googlebot and by readers, with login-gated and paywalled material excluded, and every item attributable to a creator with a public profile. Technically it needs OAuth 2.0 API authentication and strictly validated JSON-LD payloads, plus on-page schema.org markup such as SocialMediaPosting or DiscussionForumPosting with interactionStatistic sub-fields. Safety requires no illegal content and active moderation including a user reporting mechanism. Freshness is the tightest condition: new content should arrive "ideally within minutes," and engagement counters must be updated within seventy-two hours of creation.

The payload specification is not public. Google says developer documents will be shared only after acceptance, which inverts the usual order in which a platform decides whether an integration is worth building. The application is a Google Form with seven mandatory fields out of eight: company name, service domain, ten sample URLs formatted as the platform intends to send them, monthly active users, launch date, a Search Console owner email, and a contact email. Comments are optional. Google says it will reply within six to eight weeks, so an application filed on October 9 would get an answer between roughly November 19 and December 3. The form calls a submission "merely an expression of interest," disclaims any obligation to use the content, and notes that ranking algorithms determine serving.

Four gaps are worth recording. The documents name neither the Search features nor the other surfaces involved. The program's own name differs between the page and changelog, which say UGC Fresh Data Program, and the form's introduction, which says Google User Generated Content (UGC) Data Program. No volume or audience thresholds are stated anywhere. And there is no reporting, no Search Console visibility for what was submitted, and no stated processing time after a platform sends something.

The criteria point at platform owners rather than brands or individual creators, and a site that mainly publishes editorial content would fail the first condition outright. The structured data groundwork has been laid over three years: DiscussionForumPosting and ProfilePage support arrived in November 2023, a March 24 2026 changelog entry added properties for Discussion Forum and QA Page markup, and a September 24 2026 entry added a creator property with updated interactionStatistic types for VideoObject. On July 29 2026 Google made Search Console platform properties for Instagram, TikTok, X and YouTube available to all account holders, a separate product that the changelog does not connect to this one.

Amazon cuts a number it will not give while raising one it will

Amazon eliminated roles on October 6, mostly office-based and primarily in its Stores business, and a spokesperson described it as "the difficult decision to eliminate a small number of roles." Business Insider put the total below 1,000, citing one person familiar with the matter, and Amazon gave no figure of its own. Affected groups include customer service, marketplace support and retail engineering, plus staff in India and the United Kingdom. Employees received emails saying their roles had ended and entered a ninety-day notification period, with many seeking other positions inside the company.

The internal record is a Slack channel with nearly 37,000 members. Questions raised there covered severance, access to internal job postings, whether more notices were coming, medical leave, retained access to Slack and Outlook, and whether cuts would reach Mexico or elsewhere in Europe. The report records no answers to any of them. Investing.com, summarising the same reporting, dated the confirmation to Wednesday October 7 and placed it during Prime Big Deal Days, an event Business Insider does not mention.

Cumulatively the numbers are larger than the week's. Amazon announced about 14,000 corporate cuts on October 28 2025, roughly 4% of a corporate workforce of about 350,000, and confirmed about 16,000 more on January 28 2026. Business Insider puts the combined total at roughly 30,000 jobs, which against 350,000 is closer to 8.6% than to the 10% the coverage sometimes implies. In 2024 up to 160 roles in the Amazon Ads organisation were reported as potentially affected.

Capital spending runs the other way and is disclosed precisely. Amazon flagged about $200bn of 2026 capital expenditure alongside its fourth-quarter results in early February 2026, when analysts had expected something closer to $146bn, and raised the forecast by $20bn to $220bn at the second quarter. AWS sales grew 37% to $42.2bn in that quarter, the fastest pace in eighteen quarters, and advertising revenue rose 26%. For comparison, Meta's projected 2026 capital spending runs $125bn to $145bn.

Three facts therefore sit side by side without a stated connection: a small reduction in retail functions, a record infrastructure commitment, and a 26% advertising growth rate. Amazon has not attributed the cuts to automation, and no advertising or AWS team is named as affected. The industry pattern around it is more suggestive than the filing itself. Meta eliminated about 8,000 roles from May 20 2026 while about 7,000 employees moved into AI-focused positions. Microsoft cut more than 5,000 jobs across July and September. Amazon is simultaneously contacting former employees, including some laid off earlier, about openings in AWS and its AI businesses.

What remains unknown is almost everything that would make the figure interpretable. The under-1,000 number rests on a single anonymous source. Only two countries are named, and Mexico and continental Europe appear in the record as employee speculation. No schedule for further rounds exists.

Florida sues a router maker over what its security pages said

The Florida Office of the Attorney General filed suit against TP-Link Systems Inc. on October 6, seeking civil penalties of $10,000 per willful violation, rising to $15,000 where the victim is a senior, a person with a disability, a servicemember, a servicemember's spouse or dependent child, or a veteran. The e-filing stamp reads 11:41 a.m., the case number is blank, and the venue is the Circuit Court of the Tenth Judicial Circuit in Polk County. Attorney General James Uthmeier brought the action; Senior Assistant Attorney General Henry Q. Johnson signed the complaint, with Culper Law PLLC of Phoenix listed as counsel pending admission. The statutory basis is the Florida Deceptive and Unfair Trade Practices Act and nothing else: paragraph 44 disclaims any federal claim, and the Chinese entity TP-Link Technologies Co., Ltd. is not a party.

The five counts are deceptive security representations, misrepresentations or omissions about corporate separation and supply chain, omissions in data practices and privacy policies, unfair practices including ending security support for models still in use, and unconscionable practices. Relief sought runs to a permanent injunction under Fla. Stat. section 501.207(1)(b), disgorgement, the penalties under section 501.2075, fees, costs and a jury trial. The complaint states that each advertisement, sale or offer is a separate violation, but gives no penalty total and no consumer count, saying only that thousands of devices were sold in Florida.

What makes it a marketing case rather than a security one is that the allegations attach to published claims. HomeShield, built into TP-Link's HomeShield routers and Deco mesh products, is described in the complaint as covering "all security scenarios," and an archived page of November 14 2025 called it a "100% safeguard." Model-specific claims are cited for the TL-WR940N, the Archer C7, whose video said it would future-proof the home, and the Archer AX21. Archived pages carry dates of October 6 2022, March 29 2023, November 14 2025 and March 31 2026. Eleven retailers are named, from BrandsMart USA and Office Depot to Costco, Walmart and Staples, alongside six Florida internet providers listed as customers rather than defendants.

Against those claims the complaint sets a vulnerability table. The TL-WR940N carries CVE-2023-50224, was used by the Quad7 botnet, has no automatic updates and was sold through 2024. Archer C7 v2 and v3 were compromised by Quad7 and also lack automatic updates. Archer AX21 v1 to v3 carry CVE-2023-1389, an unauthenticated command injection added to the Mirai arsenal from April 2023, with v1 and v1.20 declared end of life in May 2024. The Archer AX55 v4 carries CVE-2026-18167, a stack buffer overflow in EasyMesh rated 7.7, with an advisory of September 3 2026. The Archer BE800, BE3600 and AX75 carry CVE-2026-9254 at 8.7, a parental-controls command injection running with root privileges, disclosed in late August 2026. The Tapo C120 and C200 cameras carry an authentication bypass at 8.7 and a denial of service at 7.1, reported April 16 2026 and patched August 18 2026. The Omada ecosystem accounts for fifteen flaws including hard-coded keys and predictable serials, disclosed by Forescout in August 2026.

The patching record is the recurring complaint. The AX21 update notice appeared only as a red dot in the Tether app. In April 2026 TP-Link's affected-router list omitted the TL-WR940N; in May it acknowledged that the model was affected. Microsoft reported on October 31 2024 that CovertNetwork-1658, also called Quad7, was built mostly from TP-Link routers and used for password spraying against think tanks, governments, law firms and defence firms, with the China-linked Storm-0940 then using the harvested credentials. An FBI and NSA alert of April 7 2026 described GRU DNS manipulation enabling adversary-in-the-middle attacks when a user clicks past a certificate warning.

On corporate separation, TP-Link announced its restructuring completion on May 11 2024 and said in March 2025 that its operations were "entirely different" from TP-Link Technologies. Bloomberg reported in April 2025 about 11,000 employees in China against about 305 in the United States. The complaint counts at least four Chinese facilities with a fifth building in Chengdu against a reported $180m investment, and says only 0.5% of components by value are bought in Vietnam despite the company citing a Vietnamese factory as supply-chain assurance. On June 8 2026 the Defense Department identified TP-Link Technologies as a Chinese military company under section 1260H of the FY2021 NDAA, with a Federal Register notice on June 10.

The privacy count concerns the Tether, Deco, Tapo and Kasa apps, which collect emails, location and phone identifiers, with geofencing enabling precise location collection. The complaint's argument is that the privacy policies mention affiliate sharing and legal compliance but not exposure to China's 2017 National Intelligence Law, whose Article 7 requires cooperation with intelligence work, nor to the 2021 Regulations on Network Product Security Vulnerabilities, which require reporting flaws to the MIIT within two days. A disclosure obligation to one government, on this reading, is incompatible with a marketing claim of covering all security scenarios.

Market share is openly contested. The complaint cites Rob Joyce's March 5 2025 testimony putting TP-Link at 60% or more of the US retail market for Wi-Fi systems and small-office routers; TP-Link cites Dell'Oro research putting its North American residential share under 10%. The filing does not reconcile the two. Everything here is an allegation. No court has ruled, TP-Link denies Chinese government or Communist Party ownership or control, and no response to the Florida filing has been reported. Texas sued the company on February 17 2026 under its own deceptive trade practices statute, and Netgear filed counterclaims on June 11 2026 alleging false advertising over the same separation claims. Florida had subpoenaed TP-Link Systems on December 2 2025 and created a Consumer Harm from International Nefarious Actors unit on February 5 2026.

Spain makes employers explain the algorithm, and Catalonia ships the scoring tool

Royal Decree 723/2026, dated September 9 2026, took effect on October 5, and it requires employers to tell each worker in writing when an automated system helps decide pay, hours, tasks, promotion, place of work or dismissal. The decree implements Article 8.5 of the Workers' Statute and partially transposes Directive (EU) 2019/1152. The framing in the briefing is precise: what is new is not an authorisation to use algorithms but a duty to disclose that they are in use.

The mechanics are specific enough to be auditable. Chapter II covers employment relationships lasting more than four weeks, including special relationships and public-sector employees subject to their own rules. Article 3.2.k requires written notice that such systems exist, and where they affect working hours, task allocation, pay, career progression, place of work or termination, the guidelines, criteria and operating rules must also be explained. New hires must receive the information before the relationship begins, on paper or electronically, in a form that is accessible, storable and printable, with the employer keeping proof of transmission or receipt. Changes must be communicated no later than the day they take effect. Existing staff could begin requesting the information from October 5, and the employer has thirty working days from receipt. Article 3.3 allows some items to be satisfied by reference to law or collective agreements, but not the algorithmic information, which must be given in concrete terms. The State Public Employment Service is to supply a template; its absence does not make the duty unenforceable, and it had not been published when the briefing was written.

Three disclosure tracks now run in parallel and do not substitute for one another. The individual duty sits in Articles 3.2.k, 6 and 7 of the decree. The collective duty sits in Article 64.4.d of the Workers' Statute and covers the parameters, rules and instructions behind algorithms affecting working conditions, access to employment and retention, and profiling. The data protection duty sits in GDPR Articles 12 to 14, 15 and, where applicable, 22, with Organic Law 3/2018 alongside. A notice to employee representatives satisfies neither of the other two.

Running beside the decree is AEPD warning AI-00009-2026, issued under Article 58.2.a of the GDPR and therefore neither a finding of infringement nor a penalty. It concerns an unnamed company's screening tool, not yet deployed in Spain, which would analyse CVs, score candidates for fit and prioritise applications in recruitment and internal mobility. The AEPD received information on January 6 2026; in December 2025 the company had told employee representatives that a person would make the final screening or rejection decision, that sensitive attributes would not be analysed, and that periodic audits would occur. The warning is unimpressed by the first of those assurances. The formal presence of a person does not exclude Article 22, it says, because a reviewer must be able to assess a score critically and depart from it. It cites the CJEU in SCHUFA, C-634/21 of December 7 2023, where an automated score can itself constitute an automated decision if the human decision-maker relies on it heavily, and in Dun & Bradstreet Austria, C-203/22 of February 27 2025, where explanations must cover the procedures and principles actually applied, source code need not be disclosed, and generic explanations are insufficient.

The required measures read as a specification: a risk assessment and, where high risk is likely, a data protection impact assessment under Article 35; data protection by design and by default from selection through deployment; security scaled to risk; clear transparency and meaningful human intervention. Audit reports should cover purpose, scope, system version, variables tested, error rates, differentiated impacts, oversight effectiveness, incidents and corrective actions, and reviewers should record disagreements, since a systematic absence of disagreements may indicate practical dependence on the score. A vendor's statement that no sensitive attributes are used is not sufficient on its own, and a group-developed tool does not relieve the Spanish entity using it of controller obligations. Scores, rankings, predictions and inferences linked to an identifiable person can be personal data in their own right, which triggers accuracy duties and checks for proxy variables.

Two days later, from Barcelona, the Catalan authority shipped a tool for the measurement side. On October 7 APDCAT published version 1.0.0 of free desktop software guiding organisations through a fundamental rights impact assessment of an AI system, with Windows, macOS and Linux builds, each carrying a SHA-256 checksum, and a twelve-page manual. It runs in Catalan, Spanish and English, switching languages without data loss, and moves through six steps: project data, six questions on objectives and affected rights holders including vulnerable groups, context questions with a selectable rights catalogue, two questions on existing impact assessments, eight questions on due diligence and supplier audits, and then a per-right assessment. Each selected right is rated on four indicators, each at four levels from low to very high: probability of harm, exposure as the share of the affected population, gravity as the intensity of harm, and effort as the difficulty of overcoming it. These feed three composite values, likelihood, severity and global impact, the last of which is the main decision indicator. The combination rules are not published; the manual says the thresholds follow the APDCAT model.

Two discrepancies survive into version 1.0.0. The download page says the application makes no remote connections, while section 7.2 of the manual says the first PDF export downloads typographic resources, taking several minutes and requiring internet access. The English page calls the third indicator severity, while the Catalan manual calls it gravity and reserves severity for a computed value. The default rights catalogue, drawn from the EU Charter, the Spanish Constitution and Catalonia's Statute of Autonomy, is described as indicative only, and custom catalogues can be imported as JSON.

APDCAT presented the underlying model to the Parliament of Catalonia on January 28 2025 citing four assessed real cases, with Alessandro Mantelero of Politecnico di Torino leading the working group. At the CIDAI 2025 congress, data protection officer Joana Mari Cardona said Croatia, Brazil and the Basque Country already use the model as a reference. The legal hook is Article 27 of the AI Act, which attaches to deployers that are public bodies or private providers of public services, and to deployers of credit-scoring and life and health insurance systems under Annex III points 5(b) and (c). The Digital Omnibus on AI, Regulation (EU) 2026/1744, was published on July 24 2026, entered into force on July 27, and moved the main Annex III obligations to December 2 2027 after Parliament committees backed fixed dates in March 2026 by 101 votes to 9. Because Article 27 binds deployers rather than vendors, agencies and platforms serving banks, insurers or public bodies can expect indirect demands for documentation they do not currently produce.

The OAAA builds an automation body as physical screens take the money

Digital out-of-home accounted for 38.4% of out-of-home spend in the second quarter of 2026 and grew 18.5% year on year, figures the Out of Home Advertising Association of America published and which sit behind the trade body's newest institution. Roughly two weeks before speaking to AdExchanger, the OAAA launched a Programmatic and Automation Center of Excellence, and chief operating officer Patrick Dolan set out its four pillars: standardisation and optimisation, workflow automation and agentic AI, reporting and attribution requirements, and omnichannel and retail media. Its ambition reaches past programmatic growth to automating manual out-of-home processes and the sale of static inventory, which is still mostly direct-sold.

Dolan's background explains the shape of it. He joined the OAAA in February 2025 after co-founding the IAB Tech Lab, serving fourteen years as IAB chief operating officer and three as its president from 2017 to 2020, and before that building data products at DoubleClick in the late 1990s and early 2000s. The friction he describes is the kind that standards bodies exist to remove. Metadata is inconsistent across platforms and content management systems; transit in particular needs more specific categories for digital buyers. Creative standardisation covers file size and file type across screens built in different eras with different technical capabilities, and the body is examining how to reduce the changes needed to fit one creative across many, including through responsive design. The recently released venue taxonomy is making the channel more granular, with work on a second version under way, and Dolan tied metadata, taxonomy and data structure to media mix modelling rather than to trading alone.

On attribution the Center defers: Geopath and other measurement groups lead, and the OAAA's role is fostering technology. IAB Tech Lab chief executive Tony Katsur raised the Tech Lab SDK at a listening session that brought buyers, sellers and technology companies together, and the OAAA may use it to ease integration. Embedded verification and measurement of the kind familiar in digital media is arriving in out-of-home and is a Center focus. On agentic standards Dolan was deliberately uncommitted, describing competing efforts from the IAB Tech Lab, AgenticAdvertising.org and others, saying the body is still evaluating them and wants to work with whichever standard others adopt. He described AI as a translation layer between disconnected datasets and between seller and buyer agents. The myth he wanted retired is that out-of-home is hard to buy.

The money behind the institution-building is unusually clean. US out-of-home revenue topped $3bn for the first time in a single quarter, reaching $3.16bn at 10.7% growth, with digital formats approaching 40% of the category. Mark Stenberg's October 7 column made the causal argument that the trade body carefully did not: as AI erodes online trust and web traffic, physical space becomes more valuable to marketers, and the tradeoff is that consumers have fewer ways to avoid the result. He cited a 2024 Adobe report in which 87% of US consumers said generative AI has made it harder to tell fact from fiction online, noted that answer engines resolve questions without sending readers to publisher sites, and pointed out that agents complete tasks without viewing the advertising that funds websites. Mariano Jeger left his post as executive creative director at Droga5 for Outfront Media citing that erosion. Ezra Klein, discussing New York's subway and bus advertising on an October 2 podcast, floated a modest fare increase in exchange for an ad-free subway; the MTA takes about 1% of its operating revenue from advertising.

The week's donated inventory makes the physical-space argument concrete in a way a revenue table does not. Clear Channel Outdoor and the Afterschool Alliance put a student-designed creative on nearly 2,000 digital billboards across 25 US markets for the 27th annual Lights On Afterschool. The design is by Evelyn Daniel, a tenth-grader at Cass Technical High School in Detroit, made through Usher's New Look programme at the Michigan Central Boys and Girls Club Spark Lab, with communications and design professionals guiding her and Clear Channel advising on billboard formatting. She said she is an artist at heart, and described learning to plan ahead and follow a plan. October 22 is the focal date, with two jumbo billboards in Times Square, Boston's South Station illuminated, and about 8,000 events hosted by schools, clubs, cities, parks and museums. Jodi Grant, the Alliance's executive director, framed the project as evidence of the learning and mentorship afterschool programmes offer; Dan Levi, Clear Channel's executive vice-president and chief marketing officer, framed it as student creative reaching billboards nationwide. The Alliance's America After 3PM survey puts demand at parents of 29.6 million US children against 7 million enrolled, roughly 4.2 times, though a 2020 release citing a survey of over 30,000 families put participation at 10.2 million, a figure the new number does not reconcile with.

What the materials omit is the commercial substance: the value of the donated space, impression or audience counts, measurement method, the names of the 25 markets, the billboards per market, and what illuminating South Station means. They also pass over the company's pending ownership change. Clear Channel agreed on February 9 2026 to a $6.2bn take-private deal with Mubadala Capital and TWG Global at $2.43 a share, a 71% premium to the $1.42 close of October 16 2025, approved by stockholders on May 12 2026, with an October 1 filing expecting closure early in the fourth quarter. During Advertising Week New York, which ran October 5 to 8, Land O'Lakes and Getty Images ran rural imagery in Times Square with Clear Channel without stating whether the space was bought, donated or exchanged.

A German broadcaster opens addressable TV to firms that cannot afford a campaign

Seven.One Media started the ProSiebenSat.1 AdManager on October 7, announced on LinkedIn, and the platform lets smaller mid-sized German companies plan, book and start addressable TV campaigns in the SwitchIn format across all group channels in what the company describes as a few steps. Audiences are defined by targeting criteria, campaigns can run regionally, and optimisation can be changed in real time. Nicola Lussana, Seven.One Media's chief executive, called the German Mittelstand traditionally particularly strong and a real growth engine, and said the AdManager should lower the entry barrier for advertising on the big screen decisively. The quotes are translated from German.

The historical figure is what gives the launch its weight. ProSiebenSat.1 became the first private German broadcasting group to run live HbbTV operations in 2010 and the first German broadcaster to offer addressable TV in 2015, combining SwitchIn with an interactive function on ProSieben, SAT.1 and kabel eins for viewers aged 30 to 59. A 2017 Serviceplan post, using that year's audience data, put the average gross budget of an addressable TV campaign at EUR 146,000. That is the number a self-serve tool is built to undercut. Seven.One Media's material also lists SwitchIn XXL, SwitchIn Zoom and SwitchIn Masthead, and which variants the AdManager offers has not been published.

Nor has anything else that would let a buyer evaluate it: prices, minimum budgets, fee structure, start dates, the number of targeting criteria, selectable regions, creative formats and file specifications, measurement and reporting, or who may open an account. The platform was built by MFE Advertising and was already running in Italy and Spain, where Publitalia '80 launched Mediaset AdManager on September 30 2024, growing out of a booking tool it had adopted in 2020. Paola Colombo, MFE Advertising's chief digital officer, cited about 16 million connected devices and 10 million households watching Mediaset programmes daily; chief executive Stefano Sala said the project was built to scale across countries. Italian access is limited to agencies and advertisers enabled for the service, and whether Germany follows that model is unstated. MFE secured 75.61% of ProSiebenSat.1 through its takeover offer, with the result published on September 4 2025.

The commercial pressure is visible in the group's own accounts. First-quarter 2026 revenue fell 9% to EUR 775m, with entertainment advertising revenue down 10% while digital and smart advertising revenue rose 10%. At the half year, segment advertising revenue fell 9%, digital and smart advertising rose 6%, and advertising-funded streaming revenue at Joyn rose 8%. JOM Group's December 2025 forecast expected German advertising to grow about 0.5% in 2026 to just over EUR 28bn. A shrinking linear business opening a self-serve door is a strategy with a clear precedent: on June 23 2026 ITV, Channel 4 and Sky started Universal Ads in the United Kingdom for small and mid-sized businesses using Comcast technology, with results tracked through a pixel, and on the same day Walmart agreed to buy Vibe.co, a self-serve CTV platform with more than 10,000 advertisers. Amazon DSP's managed tier typically requires a $50,000 threshold, which is the gap all of this is aimed at. MNTN's second quarter showed 4,225 active customers, up 40%, with average trailing spend per customer down 14% to about $74,200, which is roughly what democratising a channel does to an average.

Supply at the other end of the same market also moved. At Advertising Week New York on October 7, VIZIO unveiled new WatchFree+ programming and said viewing time on the free service rose 58% year on year, citing internal Q2 FY27 data, alongside a claim of 32 million yearly active devices, up 20%. The library now runs past 45,000 movies and shows and more than 400 live and local channels, with expanded deals covering MGM, NBCUniversal, Disney, A and E, and A24. A GEICO-sponsored original film, Jingle Bells Wedding Bells, starring Rachel Bilson and Scott Foley, premieres exclusively on WatchFree+ this holiday season with no date given; GEICO chief marketing officer Arianna Orpello said where people spend their time watching entertainment matters. More than forty creator-led channels sit in the free ad-supported tier. None of VIZIO's figures is independently verified, neither yearly active device nor viewing hours is defined, and earlier audience measures do not reconcile: SmartCast was cited at more than 18 million active accounts in February 2024, while executives cited 92 million households at the March 2026 NewFronts. The release gives no pricing, formats, sponsorship tiers or buying routes, and does not say who sells the inventory.

Google hands incrementality testing to advertisers and quietly loosens a view

Two measurement changes landed within a day of each other, pulling in opposite directions. Google Ads now allows advertisers to set up Conversion Lift measurement themselves, with the help documentation no longer requiring a Google Ads representative. Hana Kobzova spotted the change and wrote it up on PPC News Feed, noting that Search and Performance Max previously needed a representative. Schwartz compared the rewritten page against an archived version of April 22 2026 and found it largely rebuilt.

The thresholds are the operative detail. A study requires at least 1,000 observed conversions, with conversions that rely on supplementary data excluded from the minimum, and a minimum campaign budget of $5,000. The account must track at least one compatible conversion action. Supported campaign types now run Display, Search, Video, Demand Gen, App Campaigns and Performance Max, with App campaigns targeting iOS and Travel Ads excluded. A campaign can belong to only one study at a time across Brand Lift, Search Lift and Conversion Lift, and must be removed from an existing study before joining a new one.

Those numbers have moved before. On November 11 2025 Google lowered the minimum spend for incrementality experiments from levels approaching $100,000 to $5,000, claiming up to 50% more conclusive results from updated models and faster data availability. That announcement rated studies by feasibility, with a high rating carrying a 60% to 90% probability of a conclusive result and a low rating 0% to 30%, allowed runs as short as seven days while recommending at least fourteen, and reported that shorter studies saw absolute lift drop by up to 17% for businesses with longer conversion lags. It also set out the split that has now closed: Conversion Lift for Video, Discovery and Demand Gen could already be configured without an account representative, while Display, Search, Shopping and Performance Max required one. Most advertisers, Google said then, run about one to two incrementality studies a year. A holdout study that a marketer can start without a sales call is a different instrument from one that requires permission.

The second change is smaller and harder to defend. Google altered how it counts view-through conversions for Display ads in Demand Gen campaigns, replacing the Active View standard with a definition it calls a rendered ad impression. Under the new rule a view counts once at least one pixel of the ad appears on screen, with no minimum duration. The change applies to Demand Gen only, rolls out automatically, and was surfaced by Arpan Banerjee on LinkedIn with no dates attached. Active View exists precisely to impose a pixel-and-duration threshold; a single pixel for no stated time is the loosest definition available, and every view-through conversion counted under it will attach to a larger population of qualifying views than the same campaign produced a month ago. One platform therefore gave advertisers a sharper instrument for proving incremental effect in the same week it widened the aperture on a metric nobody can audit.

One in six ad batches now arrives from a machine

Ads Uploader, a bulk ad-submission tool for Meta founded in 2025, published its September figures on October 7, and AI agents connected directly to the service submitted 17.5% of the 36,289 batches it processed that month. That works out to roughly 6,350 batches, about 212 a day, though the company did not publish the derived figures. Total volume was 286,137 Meta ads in the month, about 9,500 a day, averaging around 7.9 ads a batch, with no split by source. Founder Chris Pollard put it as about one in six batch launches coming from an agent.

The measurement rule is the most valuable thing in the release, because it is stated. The window ran September 1 to October 1 in UTC. A batch counts as agent-made if it was submitted through the platform's MCP server or command-line interface; web-app and ad-duplicator batches count as human. Test-mode and internal accounts are excluded. The label therefore reflects the submission channel and nothing about whether a person reviewed the batch, which Pollard acknowledged in describing a gradual pattern: agents first draft and people approve, then teams let agents run unattended once the output proves reliable. The repetitive work he names is building ad sets and writing ad variations, with submission running through the same machinery a person would use and a preview before spend.

Two other figures come with weaker footing. The company says 6.6% of ads carried Meta's AI-generated content label, with no denominator stated, which would be about 18,900 ads if it refers to the month's total. And 84% of batches went into existing campaigns rather than new ones, applied in the release to people and agents together while the accompanying pitch email reads as if it covers agent batches alone.

What the terms of service say matters more than usual for a tool that lets software spend money. Last updated June 18 2026, they do not mention agents, MCP or the command-line interface at all. Section 8 makes the account holder fully responsible for ads, budgets and targeting, and warns that bulk configuration errors multiply. Section 11 excludes liability for spend, targeting and budget errors and provides the service as-is. Section 10 allows immediate suspension or termination without notice. Section 13 sets Wyoming law, arbitration and a class-action waiver. Pricing is flat regardless of volume: Solo Unlimited at $44 a month or $528 annually with a 25% discount, Team Unlimited at $54 per user per month, with teams from two to twenty users and volume discounts. Both tiers include agent-ready CLI and MCP access, which is to say the cheapest way to let an agent spend on Meta costs $44 a month.

The platform groundwork arrived fast enough to date precisely. Amazon introduced Ads Agent and Campaign Manager at unBoxed on November 11 2025, opened a closed beta for agent integration two days later, and moved its MCP server to open beta on February 2 2026. Meta opened AI connectors for Claude and ChatGPT on April 29 2026 with write access from the start, and opened its ads MCP to any developer on July 16 2026. X released an Ads MCP on August 24 2026 with twenty-three tools, ten of them write-enabled, creating campaigns in a paused state.

The caveats are substantial and the article states them. The data is vendor-supplied, unaudited, covers one month and one platform, and has no earlier baseline, so no trend can be drawn from it. Human review is not measured. Budget and spend data are absent. Ads Uploader handles roughly 300,000 Meta ads a month across more than fifty countries, a small share of Meta's volume. Agentic adoption claims have circulated for a year without a denominator attached to any of them; this is one denominator, from one small vendor, for one month. Kargo announced general availability of Karlo on October 1 with a single Ad Council campaign as evidence and a vendor-reported 21% lift in ad recall. Set beside that, a published measurement rule and a stated exclusion list look like a higher standard than the category has been holding itself to.

A prime minister files her own voice, and finds out what it would protect

Giorgia Meloni filed an application with the European Union Intellectual Property Office on October 5 to register a four-second recording of her own voice as an EU trade mark. The file is number 019431219, it is under examination, and the register notes that the application fulfilled Fast Track conditions. Reuters reports that her office said the aim is protection against AI-generated deepfakes; the register states no purpose, and nothing has been approved.

The entry is sparse and specific. Mark type is recorded as a trade mark without text, type Sound, nature Individual. The filing date and date of receipt are both October 5 2026, the filing language is Italian with English second, and the applicant reference is TM31699EU00. Acquired distinctiveness: no. Description: no data. The embedded player length is four seconds. Publications, oppositions, decisions, appeals and renewals are all recorded as no data. The owner is Meloni as a natural person in Italy, represented by Praxi Intellectual Property S.p.A. of Civitanova Marche with correspondence to a Rome law firm. Reuters says the clip has her saying in Italian that she is Giorgia Meloni, twice, a phrase connected to her 2021 autobiography and a 2019 rally declaration.

The classes reveal what was actually claimed. Fifty-one terms span three classes: six in class 9 covering downloadable and recorded media content, multimedia files, electronic publications and multi-media recordings; forty-three in class 41 covering cultural activities, conventions, seminars, teaching, training, live entertainment and publishing across books, magazines, newspapers, journals, reviews and electronic publications; and two in class 45, organisation of political meetings and legal research in the field of public policy. Under the fee schedule in force since 2016, that is EUR 850 for the first class, EUR 50 for the second and EUR 150 for each further class, with one 2026 guide putting the online total for three classes at EUR 1,050. The extract shows no payment details.

Fast Track, set up in 2014, admits word, figurative, three-dimensional and sound marks where applicants pick terms from the harmonised database, pay at filing, and file without a description or disclaimer, and an application leaves the route if an examiner raises an issue. Third-party estimates put examination at about a month and the opposition period at three months after publication, roughly 50% faster than the standard route, with a certificate about six months after publication if unopposed. Applied to October 5, that suggests examination closing around November 5 2026 and an opposition window closing in early February 2027, though those are practitioner figures rather than office dates.

What the mark would actually do is where the story turns. Francesco Venturoli of Bugnion, writing on September 23 2026, said a registration gives no exclusive right over the voice as such: opposing an AI copy would require showing sign similarity, similarity of goods and services, and a likelihood of confusion, and a deepfake may avoid trademark use altogether. Bugnion lists personality rights, data protection and copyright as complementary tools and doubts a trade mark would suffice against faithful replicas. Il Sole 24 Ore reported that registration could add EU-wide action, takedown of unlawful material, damages and criminal proceedings for counterfeiting under the Italian criminal code. Italian media quoted elsewhere made the narrower point that approval would not by itself stop deepfakes but could add legal obstacles for their producers. Meloni urged users to verify content and said she can defend herself while many others cannot.

The precedents do not settle the question of whether a spoken name is registrable. In Ardagh, T-668/19 of July 2021, the General Court's first ruling on an audio-file sound mark upheld a refusal because the sound of a can opening is functional or inherent to the goods. In Globo, T-408/15 of September 13 2016, refusal was upheld because the sound resembled a banal ringtone. ABG-IP counted 313 registered EU sound marks up to May 2021; Inventa counted 532 cases involving EU sound marks including international designations. No cited authority addresses a spoken name, and the office has not said how it will treat one. Two other Italian voices are already in the register or queue: Luca Ward filed in February 2026 and is now registered, and Giusy Ferreri filed on April 23 2026 as EUTM 019353063 in classes 9 and 41, still pending.

The surrounding law has been moving faster than the trademark file. Italy's Law 132/2025 took effect on October 10 2025, adding Article 612-quater to the Criminal Code to cover non-consensual dissemination of misleading AI-falsified content with one to five years' imprisonment, and Italian firms now face advertising bans over AI-generated content offences. A Berlin court held on August 20 2025, in case 2 O 202/24, that AI voice cloning violated personality rights, awarding EUR 4,000 plus EUR 1,155.80 in costs against a YouTuber with about 190,000 subscribers. Italy's data protection authority, by order 577 adopted July 23 2026 and published August 7, banned further processing of Enrico Mentana's data in AI-altered clips broadcast by Striscia la Notizia, with thirty days to report compliance and no fine imposed. AI Act Article 50 transparency obligations applied from August 2 2026. And on March 10 2026 YouTube extended likeness detection to a pilot for politicians and journalists, with audio detection planned later in the year and four dispute categories covering face or voice added on August 19 2026. No cited source describes any platform policy treating an EU sound mark as a takedown trigger.

Meloni's own litigation sits behind the filing. She complained to Rome postal police in June 2020 about AI-made fake videos on a US erotic website; two Sassari residents were reported for defamation in March 2021 after international cooperation identified the uploader; she sought EUR 100,000 for the Interior Ministry's fund for women victims of violence; and she testified by video link on October 9 2024, calling it a form of violence against women. Reuters describes the matter as still pending.

Women's sports sponsorship runs into the measurement question

The Golden State Valkyries are two years old, and their chief revenue argument is not reach. Jess Smith, speaking to Digiday on October 9, said that in the room with brands the team pitches a place where mission and capital can coexist, combining conventional inventory, signage, radio, digital advertising and on-site fan activations, with a community component. She declined to give dollar figures but said partnership revenue grew 30% year on year between regular seasons. The team has more than forty partners including CarMax, Rakuten and Kaiser Permanente, reportedly became the first women's sports team to reach $100m in annual revenue, has doubled its marketing headcount, and holds a data partnership with the talent agency The Team, formerly Wasserman.

Exclusivity is the commercial pressure point. Smith called it super important, for teams as well as brands, and noted that categories are broad while non-traditional ones are increasingly interested: Sephora is a partner, as is Iren, an Australian AI storage company that the team is working with on international efforts and on local STEM programming. Owning a category costs money, and the team balances that against filling the roster. A recent Athleta deal adds a Gap brand to the list.

Measurement is where the conversation gets harder, and the asks are specific: event attendance, social fellowships and views, and traditional media value, across radio, local television and national television coverage. Larger brands run their own studies and want consistent metrics so that sports spending can be compared with non-sports spending. The Valkyries share an eight-person partnership insights team with the NBA's Golden State Warriors, and brought in The Team to measure community programmes such as court renovations and youth access. Next season adds six games, a 15% increase in the schedule, which Smith translated directly into signage exposure, and a new league broadcast agreement should lift ratings. Her argument to a sponsor is about the forward curve rather than the current audience.

Which is precisely the argument a brand without league rights has to make differently. JPMorganChase has no WNBA sponsorship, so it went around the boundary: Tara Bakhle, executive director of sports and entertainment marketing strategy, told an Adweek panel that the bank cannot activate traditionally within the boundaries and used Togethxr instead. Togethxr, founded by Alex Morgan, Sue Bird, Chloe Kim and Simone Manuel, opened a house at the Women's Final Four in Phoenix this spring without being an official partner, and Chase ran a pop-up shop there with Valkyries guard Tiffany Hayes and Round21, accessible to cardholders. The pair worked together again at WNBA All-Star Weekend in Chicago, where Togethxr recorded its Everyone Watches Women's Sports podcast live at the Chicago Theatre while Chase offered cardholders presale tickets and hosted a VIP experience for wealth management clients.

The reason Chase wants that audience is a single number from its own card data: comparing WNBA and NBA transactions, the overlap is only 5% to 6%. Bakhle said WNBA fans skew younger and more affluent, matching the customers the bank wants to acquire and keep, which is to say its existing sports portfolio largely misses them. Deloitte predicts the women's elite global sports market will reach at least $3bn in 2026, up 340% since 2022, and Nielsen measured women's sports at 46 billion minutes of viewing in 2025.

Both sides named the same infrastructure problem. Bakhle said rights are split across ESPN, NBC and others, and that even accessing the programming or finding the game on television is still actually pretty hard, while investment still does tend to go into the men's side. Bailey Williamson, Togethxr's senior vice-president of partnerships, said fans are harder to impress because tickets are extremely expensive, and that fans paying those prices want more from the day than the game. Togethxr has sold more than 200,000 Everyone Watches Women's Sports shirts, helped by a Nike partnership that took sales global, and extended the slogan to Stanley tumblers and hats, which is a merchandising business built on a sentence rather than on broadcast rights.

Also noted

• October 7 Tunnl put 131 survey-based audience segments into DeepIntent's healthcare demand-side platform, among them COVID vaccine intent, flu shot intent, birth control funding support and confidence in FDA medication approval, built by modelling large-sample survey findings across the whole US population so that most people in a segment were never surveyed, with activation priced on a data CPM that has not been published alongside no disclosed sample size, refresh rate, score validation or identity graph coverage. PPC Land
• October 8 Google's Merchant Center documentation now warns that products whose price or availability changes more than once a day may be disapproved rather than updated when a mismatch is detected, and that updates may stop altogether if detected mismatches become too numerous, a page Hana Kobzova dated to a September 27 revision. Search Engine Roundtable
• October 7 Google restructured its alcohol advertising help page and will allow 0% alcohol beverages to be advertised in India, Indonesia and Egypt subject to a new certification application, alongside permission for drinks below 0.5% ABV and alcohol alternatives, plus fresh prohibitions on misleading strength claims and on presenting alcohol as a way to remove social or sexual inhibitions. PPC News Feed
• October 7 LG Ad Solutions made its television Home Screen inventory available through Best Buy's advertising network, extending a placement format whose economics are still mostly undisclosed across the category. PPC Land
• October 7 USA TODAY Sports and PlayVS staged an esports contest for eight high schools, a publisher-run competitive event rather than a sponsorship of someone else's. PPC Land

By the numbers

• 13 days, 16 hours Time Google's September spam update spent rolling out, against one to two days in its own internal figures. Source
• $220bn Amazon's 2026 capital spending plan, raised by $20bn mid-year while retail roles were being cut. Source
• 6 to 8 weeks Reply time Google states for a UGC Fresh Data Program application, before any specification is shared. Source
• 30,000 Roles Amazon has eliminated since late 2025, about 8.6% of a corporate workforce near 350,000. Source
• Nearly 2,000 Digital billboards across 25 markets carrying a Detroit tenth-grader's design, donated rather than sold. Source