What changed since this newsletter last wrote about Apple's Safari blocklist is that Apple started talking. On October 5 an Apple engineer asked the vendor whose delivery domain had been blocked to go and test a new build, in public, on a bug tracker anyone can read. Nothing was conceded and nothing was explained. The exchange lasted fifteen minutes and twenty-two seconds and named no domain, no setting and no behaviour. It was still the first two-way conversation between Apple and the ad industry on the subject.

Around that silence, the week's other stories arranged themselves into a pattern about who gets to set the terms. A Finnish supervisory agency told a Google subsidiary to stop clearing forest at two data centre sites before the environmental assessment is finished. The Television Academy moved the Emmys off four broadcast networks and into a single company that also sells the advertising. Apple told developers that from April 2027 an app without screenshots for a device that does not yet ship cannot be submitted at all. Microsoft Advertising raised the ceiling on LinkedIn company targeting from a filing cabinet to a spreadsheet of ten thousand firms. Sky put its streaming supply inside Amazon's buying platform. Pixalate began labelling apps that buyers have asked for money back on. And a measurement vendor started selling education brands the number 9.3%.

Underneath all of it sat a question an opinion column in AdExchanger put more precisely than anyone else: when an autonomous agent commits money faster than evidence arrives, who pays for the gap?

Fifteen minutes on a bug tracker, and a server the publishers had ignored

The record is WebKit bug 324771, filed on September 21 at 19:13:36 by Ian Meyers, The Trade Desk's senior director of engineering. For roughly two weeks it produced nothing but holding replies. On October 5 at 12:29:29 Pacific time, John Wilander of Apple said a new iOS 27.2 beta had shipped that day and asked Meyers to test it, thanking him for the technical documentation he had supplied. At 12:44:51, fifteen minutes and twenty-two seconds later, Meyers answered that he could see changes in build 24B5099f and that his company would test and report back.

The bug's formal state has not moved. Status NEW, resolution blank, priority P1, severity Major, version Safari 27, hardware iPhone and iPad, assignee Nobody. Seven users now sit on the CC list, one more than at the end of September. The single attachment, numbered 481466, is a 51.16 KB screenshot of blocked requests on a yahoo.com article in ordinary, non-private browsing. A Radar importer linked rdar://problem/188653580 on September 28, which places the report inside Apple's internal tracker without saying what happened to it there.

The mechanism was never mysterious once somebody read the commit. WebKit pull request 58670, titled Unconditionally block requests going to certain domains, added eleven lines to a single file in February: Source/WebKit/Platform/cocoa/WebPrivacyHelpers.mm, eleven additions and no deletions, committed from Charlie Wolfe's branch. A macro named IS_REQUEST_UNCONDITIONALLY_BLOCKABLE returns false in public builds, and Apple's internal additions file can override it. Inside isRequestBlockable, at lines 773 to 775, the check runs against the request's registrable domain before the existing tracker lookup. A listed domain is reported as blockable before the normal branch has a chance to run. Nothing in the published diff shows whether any user-facing setting gates it.

Meyers named nine domains: tainted.example, uidapi.com, adsrvr.org, id5-sync.com, eu-1-id5-sync.com, rlcdn.com, pippio.com, permutive.com and ad.gt. Eight of those map to the post-cookie identity business, covering Unified ID 2.0, ID5, LiveRamp's two domains, Permutive and Audigent. The ninth, adsrvr.org, is The Trade Desk's core ad request and delivery domain, which is why Meyers argued that the list looked aimed at identity while catching delivery as collateral. Safari has blocked third-party cookies by default since March 24 2020, so the matching subdomain was serving a legacy cookie nobody much depends on. ID5's chief executive has already said regulators should look at the list. What build 24B5099f actually does, whether the other eight entries moved, whether the change reaches the production release of iOS 27.2, and whether Mac behaviour is affected, are all still open. The bug names only iPhone and iPad.

The more consequential development arrived on October 8, in a four-byline Digiday piece reporting that the blocklist has sent publishers back to a product the industry had been politely ignoring for eighteen months. IAB Tech Lab unveiled Trusted Server in March 2025 and pitched it on control and page speed. The pitch did not land. Benefits were hard to price, costs were visible, the software sat close to prototype, and running it meant handing servers to engineers in an organisation where ad operations had never owned infrastructure. Earlier server-side auctions had also earned less, because buyers who lost cookie access bid lower.

The architecture is the reason it matters now. Today a reader's device contacts dozens of ad tech companies directly, which is exactly what Safari can see and therefore block. Trusted Server moves that process onto servers the publisher controls, so the device talks only to the news site. Apple's remaining option would be blocking the site itself, which would break it for every Safari reader. Rowena Lam, senior director of privacy and data at IAB Tech Lab and the product lead for Trusted Server, said there has been "a big influx of publishers reaching out to get a better understanding of Trusted Server" and declined to put a number on it. Interest had been building for about a year, she said, and rose sharply in the week after the blocks appeared. Tech Lab is preparing to go live with a first publisher. The body has previously had to set out how Trusted Server and Prebid divide the work, since the two are easy to confuse.

One publisher is already running the experiment. Paradium, which owns Men's Journal and Parade and rebranded from Arena Media Group in August, uses Permutive to group readers for advertisers. Permutive is on the list, so that stopped working in Safari. Paradium rerouted the Permutive code through Trusted Server and is testing whether audience data flows again. Stephanie Mazzamaro, head of revenue and data, expects the gains to be uneven: identifier longevity, ad-block resilience and clean auction data should improve, while "viewability, verification and getting buyers to accept Paradium's numbers are where most of the work remains." Her objection is procedural rather than commercial. Safari blocks companies against a list Apple has not published, with no notice and no visible appeal, and the industry learned of it from broken ad delivery rather than from Apple.

Aditude, which helps publishers sell inventory, offers its own Prebid Server and connects through APIs to ID5, UID2 and LiveRamp. Justin Wohl, its vice-president, made the structural point: the disruption came from an operating system change rather than a browser change, and "those in control of the operating system have great control." He noted the same could happen on Android. A second publisher executive said the company knew about Trusted Server but had not studied it, and asked not to be named to avoid antagonising Apple, which is itself a data point about market power.

That power has a revenue line attached. Apple renamed Search Ads to Apple Ads last year, added placements to App Store search results this spring, and began selling advertising in Maps with a 15% credit on spend. Its Services division, which contains the advertising business, reported record revenue of $30.7bn for the June quarter. The same iOS 27.2 cycle also carries the alternative App Tracking Transparency prompt that becomes mandatory in Germany, France, Italy, Poland and Romania. Whether build 24B5099f relents on anything, nobody outside Cupertino yet knows. What is observable is that the industry's answer to an unpublished blocklist is to stop putting deterministic signals on the device at all.

Finland tells Google to stop cutting trees until the assessment is done

Finland's Permit and Supervision Agency, the Lupa- ja valvontavirasto, published two notices at 14:17 on October 6 addressed to Tuike Finland Oy, a Google subsidiary. Each one asks the company to halt land-altering preparatory work at a planned data centre site until the environmental impact assessment is complete. The Muhos notice carries diary number LVV-U/37128/2026 and the Kajaani-Otanmaki notice LVV-U/22159/2026. Both rest on section 32(2) of the Finnish EIA Act 252/2017, which implements EU Directive 2011/92/EU.

The scope of what must stop is specific: tree removal, topsoil stripping, excavation, blasting, crushing, earth movement, ditching, and the building of site roads and storage areas. Planning, measurement and soil investigation may continue. The suspension is to begin immediately and no later than October 23, and to run until the liaison authority's reasoned conclusion on environmental effects exists. A written reply from Tuike was due on October 14.

Terminology matters here, and the coverage has blurred it. The documents are a call to act rather than a binding order, and the agency said it first seeks voluntary correction. If work continues, it may issue an order under section 35 of the EIA Act or under special legislation, backed by a conditional fine under Act 1113/1990. The BBC and Helsinki Times both described the step as an order.

The sites are large. Muhos covers roughly 571 hectares and is planned for nine data centres plus support facilities, cooled from the Oulujoki river, with backup power above 300 MW. Kajaani-Otanmaki covers roughly 405 hectares and is planned for eight buildings, a substation and a 400 kV overhead line, with cooling from Lake Oulujarvi still under study and backup power again above 300 MW. Roughly 330 hectares at Muhos, about 58% of the project area, has already been felled or prepared. At Kajaani-Otanmaki the figure is just under 200 hectares, just under half. Two Natura 2000 areas, Otannevan and Talaskankaan, were screened by Sweco in May 2026, and Tuike reported no effect on either.

Tuike's defence is that the work began lawfully under section 109 of the Building Act 751/2023, which permits preparatory work before construction, and under a forest use notification filed under the Forest Act. The company says it protected the highest-value zones, limited clearing in the lower ones, avoided the April to July bird nesting season and tried not to cross streams. It also conceded something awkward: the no-build baseline for ecology and watercourses, the VE0 scenario against which everything else is measured, now differs from the original starting point. Tuike argued VE0 remains appropriate anyway.

The agency's reasoning goes the other way, and it is worth reading closely because it is the part that generalises. Preparatory work cannot be severed from the project. Large-scale topsoil removal and terrain change are comparable to starting the project. The changes are neither temporary nor easily reversed. And critically, the work narrows the alternatives an assessment is supposed to keep open, which is the point of doing an assessment first. The notices cite CJEU case C-215/06 on not splitting projects in ways that undermine assessment, and state that building control's view does not limit the agency's supervisory powers. Tommi Muilu, who heads the agency's environmental department, said the measures change the environment and affect what the assessment is meant to assess.

The sequence shows how the gap opened. Both assessment procedures started in summer 2025, with reports due at the end of 2026. The ELY Centres issued statements on Kajaani-Otanmaki on September 19 2025 and on Muhos on November 28 2025, and were replaced as liaison authority by the new agency on January 1 2026. Muhos filed a landscape work permit application on March 31 2026 and supplemented it on May 4. The agency advised the municipality on May 18 that no permit should precede the assessment report. On June 5 the municipality told Tuike that section 109 preparatory work was possible, which the notice characterises as neither a permit nor a permit decision. No building permits have been granted at either site. The agency learned of the Kajaani work at a site meeting on August 27 and through Yle's reporting on September 17, requested clarifications on September 18, and received Tuike's reply on September 25.

The money makes the timing conspicuous. On November 28 2024 the Council of State authorised Metsahallitus to sell four parcels covering about 1,400 hectares in Muhos and Kajaani to Google for more than EUR 27m, raw land with zoning still required. Antti Jarvinen, Google's country manager for Finland, called the sale a first step. On September 9 2026 Google said it would invest at least EUR 13bn in Finland across 2027 and 2028, covering Hamina, Kajaani, Muhos and Vaala, projecting more than 37,000 construction jobs, EUR 3.6bn added to GDP annually and 7,000 operational jobs a year, with EUR 31m over four years for community initiatives. The energy commitments run alongside: a 22-year power purchase agreement with Fortum supporting the Loviisa nuclear plant, a memorandum of understanding on new capacity, wind agreements bringing supported new capacity to 629 MW, and a 94 MW battery near Kajaani expected by late 2027. Google has been in Hamina since 2009 and opened its Finland cloud region in July 2018; its compute ambitions now extend to a space-based prototype targeted for 2027.

Google told the BBC it had "fallen short of our own high standards in this instance" and would follow the agency's guidance. Hanna Halmeenpaa, chair of the Finnish Association for Nature Conservation, said protected natural sites had been felled; the association put the Leppiniemi clearance in Muhos above 300 hectares in September. The agency has not yet assessed whether a Water Act permit is needed, is still processing ditching notifications, and will ask for a report on the volume, quality and placement of excavated soil. Nine data centres do not get built without the electricity and the water, and this is the first time a European regulator has told the company to stop moving earth while the paperwork catches up.

The Emmys leave four networks for the company that sells the ads

Amazon and the Television Academy disclosed on October 6 a six-year agreement making Prime Video the exclusive worldwide home of the Emmy Awards from the 2027 ceremony. The 2027 show airs live in September. No subscription is required, and Amazon's materials say the stream reaches 240-plus countries and territories, though chair Cris Abrego's quoted remarks say more than 200, one of two internal inconsistencies in the announcement.

The arrangement it replaces was the Wheel Deal, a rotation among ABC, CBS, FOX and NBC that ended with the 78th ceremony on September 14 2026. A 2024 Variety column put the old licence fee at roughly $8m a year split among the four networks. Variety reported advanced talks in mid-September 2026, and described an Academy that had considered a rotation including new outlets before one network, believed to be FOX, dropped out; negotiations reportedly turned on the fee and on the appeal of a single outlet holding rights for several years. The Emmys began in 1949, which makes Prime Video the first non-broadcast home in the award's history. Adweek covered the move the same day.

Abrego said "television has never been more global" and called the broadcast partners tremendous stewards of the Emmys. Mike Hopkins, who runs Prime Video and Amazon MGM Studios, framed a free stream as another way of delivering the best in entertainment. Neither statement answers the commercial question, and the release does not either: nothing says whether the stream carries advertising or sponsorship, who would sell it, or how the audience would be measured. The separate NATAS ceremonies are also unaddressed, as is what Amazon means by new experiences and entry points for fans across the year.

The reason those omissions matter is that a flagship live event has just moved inside one company's advertising stack, which changes who controls the inventory, the data and the reporting. Prime Video began carrying advertising in January 2024 with an ad-supported audience of 200 million and reached 315 million viewers globally by the end of 2025. At unBoxed on November 11 2025 Amazon put average monthly ad-supported US viewers above 130 million, with advertising live in sixteen countries including the US, UK, Germany, France, Japan, Brazil and India. Thursday Night Football entered its fifth season at the May 11 2026 upfront with viewership up 60% since 2022, and a January 2026 Wild Card game averaged 31.61 million viewers. NBA coverage started on October 24 2025 with programmatic guaranteed, contextual and run-of-service inventory. Amazon's advertising revenue reached $19.8bn in the second quarter of 2026, up 26%, with sports inventory sold out.

Three practical constraints sit against the announcement. A free global stream and an ad-supported global stream are different products, and the release does not say which this is. Upfront negotiations cluster in May, so 2027 is the earliest plausible window for an Emmys advertising package. And 315 million is a monthly average across a catalogue rather than a single-event audience, which is a different kind of number from the 8 to 10 million viewers AV Club attributes to Emmy broadcasts over the past decade. Netflix said in November 2025 that no universal streaming reach measure exists, and nothing in the Emmys release names a measurement source.

The pattern is now unmistakable. Netflix agreed in January 2023 to carry the SAG Awards from 2024. Prime Video took the Academy of Country Music Awards after they left CBS in 2021. On December 17 2025 the film academy said YouTube will host the Oscars from the 2029 ceremony through 2033, free to more than 2 billion viewers, with ABC keeping the event through 2028. Two Emmy ceremonies, in September 2027 and September 2028, will run before that handover. One further detail is worth recording without inference: Amazon MGM Studios earned 68 Emmy nominations across sixteen titles in 2026, and voting is conducted by the Academy's membership.

Education brands bidding on rivals see one invalid click in eleven

Lunio published an education and e-learning invalid traffic report on October 6 covering more than five million clicks across Google, Meta and other native and social platforms between October 2025 and June 2026. The headline split is the useful part. Competitor-targeting campaigns carried 9.3% click-weighted invalid traffic against 4.7% for generic, non-branded campaigns, which is roughly one invalid click in eleven against one in twenty-one. Sector-wide invalid trafficrose 33% over the window, from 4.08% in the fourth quarter of 2025 to 5.42% in the second quarter of 2026. No first-quarter sector figure was published, and the release's subheading attributes 4.7% to the sector overall while the body attributes it to generic campaigns.

Two unnamed campaign examples carry the argument further than the averages do. An online course platform's competitor-brand campaign ran below 8% in every quarter before the second quarter of 2026, then reached 26.1%, with no explanation offered. A second competitor-targeting search campaign registered 35.6% and then 33.6% across the first two quarters of the window. Nick Morley, Lunio's chief executive, called education "one of digital advertising's most aggressive brand-bidding arenas" and said brands are "losing both money and market share." That is a correlation presented as a mechanism; no test in the release separates targeting choice from audience, platform or bid effects.

The platform breakdown contains the detail most likely to change a media plan. Meta averaged 8.34%, peaking at 9.20% in the first quarter of 2026 and falling to 7.31% in the second, about a fifth below peak. Google averaged 3.70% but moved the wrong way, from 3.16% to 4.61%, a 46% increase. Google Search was the worst Google channel by the end of the window, rising 137% from 2.32% to 5.50%. Google Demand Gen had the highest Google average at 4.16% and was the only channel to improve every quarter, from 5.12% down to 3.43%. No education figures were given for Performance Max, Display or Shopping.

The dollar framing is a modelled estimate rather than an observation. For a brand spending $5m a year at a $6.23 average cost per click, about 802,600 clicks, Lunio puts wasted spend at $232,762 and lost revenue near $698,000 at a 3:1 return. The arithmetic implies a rate of about 4.66%, close to the generic figure; at the second-quarter sector rate of 5.42% the same model would produce roughly $271,000 and $813,000. The release does not say which rate it used. It also omits the number of advertisers, accounts or campaigns, whether accounts were protected or monitored only, and any LinkedIn or Bing data, which is a conspicuous gap given that LinkedIn carried the highest rates in Lunio's own banking study at 36.08% and its IT study at 15.34%. Lunio sells detection tools.

Set against the vendor's other verticals, education looks less like an outlier than a reshuffle. IT and security ran 4.29% across 64 million clicks; banking, lending and credit 5.92% across 86 million; retail 5% across 414 million, rising to 5.54% in the second quarter of 2026. Meta averaged 5.99% in retail against 8.34% in education, while Google averaged 4.75% in retail against 3.70% in education, so the platform ranking inverts by sector. Lunio also found retail AI Max campaigns at 5.28% against 3.07% for standard search. The education window closes in June 2026, before Google's September 1 conversion of some campaigns to AI Max, so the two findings do not overlap. Lunio's own survey of 131 senior marketing leaders, fielded in May 2026, found 75.6% estimating losses above 5% of monthly performance budget while 5.3% use a dedicated prevention tool and 39.7% believe the major platforms already handle it.

Pixalate starts publishing which apps buyers asked for money back on

The same day, from London, Pixalate added a Clawback Warning label to its Media Ratings Terminal. It appears on the page of any connected TV app, mobile app or website that buyers have disputed as invalid traffic within a rolling twelve-month window, refreshed monthly, alongside the risk ratings the company already publishes. The terminal evaluates more than 17.5 million properties across seven app stores: Apple's App Store, Google Play, Roku, Amazon Fire TV, Samsung TV, Apple TV and LG TV. Its platform table lists 745 global ad platforms. The label is public at ratings.pixalate.com; the Clawback Tool that feeds it is for Pixalate Analytics clients. The company's own comparison is Amazon's Frequently Returned Item flag.

The argument behind it is about the shape of the supply chain rather than about fraud detection. Low invalid traffic on one exchange does not make a publisher clean, because a property can route better traffic through some supply-side platforms and worse traffic through others, particularly in international markets, and each exchange sees only its own disputes. Pixalate names four uses: waterfall detection, cross-exchange screening, onboarding due diligence and continuous monitoring. The intended reader is an exchange deciding whether to take a publisher on.

The worked example is more informative than the description. A mobile game with a blurred name shows nine clawbacks in the last twelve months, distributed as one each in February, April, May, June and September 2026 and four in August 2026, which yields six affected months of twelve. The developer country is Cyprus and the download band is 100 million to 500 million. A footnote states that counts are clawback reports per month, that the reporting party is not shown, and that the counts include both Pixalate clients and external submitters. The release's own text says four of twelve while its screenshot says six, a discrepancy left unexplained.

Three limits are worth stating plainly. Eligibility depends on exceeding undisclosed clawback thresholds observed across multiple seller paths, and the release defines neither clawback nor whether a dispute must end in a granted refund. Coverage depends on who files, so a property trading mainly with buyers that do not file would look cleaner than its traffic warrants, and Pixalate publishes no monthly report volume, no count of labelled properties and no detail on how third-party submissions are verified beyond a review against supporting evidence. And there is no appeal: the label is applied automatically, cannot be added or removed by hand, and clears only after twelve months with no reported clawback. Pixalate's accreditation from the Media Rating Council covers sophisticated invalid traffic detection and filtration, and nothing states whether this label falls inside it. The company's own disclaimer calls its opinions neither facts nor guarantees and describes fraud as a custom measurement term rather than a legal one.

The refund machinery it documents has existed since September 2017, when Google said AppNexus, Index Exchange, OpenX, PubMatic, SpotX, Teads, Telaria and DoubleClick Ad Exchange had committed to refund invalid traffic detected within thirty days of monthly billing, covering an estimated 90% of programmatic buying. What has been missing for nine years is any record of who kept asking. Pixalate's adjacent work this year points the same direction: a February 2024 analysis found 25% of programmatic traffic carrying supply chain object data failed validation, with invalid traffic 64% higher in the failing set, and the company shipped OpenEPG Index 1.0 in June followed by OpenEPG DB 1.0 and a pre-bid API in private beta on September 21.

Microsoft Advertising raises LinkedIn company targeting to 10,000 firms

Microsoft Advertising's October product newsletter, the third monthly edition since the format began on LinkedIn in August, carries six items, and the one with the clearest operational consequence is bulk company targeting. Advertisers can now add up to 10,000 companies to a targeting or exclusion list, either during campaign creation or in the Audience library. Lists take one to two days to process and match. Company lists behave as or logic alongside other LinkedIn targeting and as and logic alongside device, location or content targeting.

The history explains why ten thousand is a change of kind rather than degree. LinkedIn Profile targeting arrived in Bing Ads as a US-only beta on October 30 2018 and expanded to search and audience campaigns in September 2020 covering more than a hundred industries and 80,000 companies, with job seniority added later. A 2024 review noted no bulk option, and a Microsoft support answer dated June 10 2026 said importing company lists from files was unsupported. PPC News Feed has reported a ceiling of 1,000 companies per ad group; the newsletter states no such limit, and the two have not been reconciled.

The HubSpot integration in the same edition is aimed at a different problem: what the bidding system is allowed to optimise toward. Advertisers can sync performance data including conversions into HubSpot, and send custom HubSpot lifecycle-stage events such as marketing-qualified and sales-qualified leads back the other way, so bidding can optimise toward higher-quality conversions rather than raw form fills. CRM contacts can be re-engaged across Microsoft's properties with verified LinkedIn profile data attached. Setup requires Publisher access in HubSpot and Super Admin rights in the Microsoft account, and the newsletter calls it a more seamless, two-way connection than third-party tools such as Zapier. HubSpot's documentation adds the caveats: only stage changes created after the event exists are counted, with no backfill, Microsoft may calculate conversions differently, and a setting governs whether all contacts or only ad-interacting contacts are shared. HubSpot's landing page cites 2026 Microsoft internal data of roughly 42% lower cost per click, roughly 64% higher conversion rate for job title, company and industry targeting, and 15% to 25% more qualified leads when Google and Microsoft run together, with no sample size or method. Requests for the integration date back to 2017 on HubSpot's own community forum, and a Microsoft support thread of October 2 2025 confirmed none existed.

Impression-based remarketing fills in the third piece. Audiences persist for one to thirty days with a seven-day default, one list can combine impressions from up to twenty campaigns or ad groups, and at least one source must be an Audience campaign for cross-type use. Connected TV can act as a remarketing source but not as a targeting destination, which means a CTV-exposed audience can be addressed on Search. Microsoft suggests always-on Audience campaigns at $5 to $10 a day to reach previously exposed users, and cites a 2024 blog drawing on July 2023 internal data showing three times site visitation and seven times conversion rate from cross-format exposure against Search alone, figures now more than three years old.

Two further items are policy rather than product. On AI-generated creative, Microsoft states that consumers should be able to tell when content is AI-created or altered, leaves advertisers responsible for legal compliance, rights, disclosures and provenance including watermarks and metadata, and reserves the right to reject or remove ads that mislead, impersonate or omit required disclosures. No jurisdictions are named, which distinguishes it from Google's attestation field in DV360 and its labelling in the EU, India and New York state. On product catalogues, the four-part checklist covers uploading the full sellable catalogue, accurate titles, descriptions, GTIN or MPN, price, availability and category, daily feed refreshes, and Universal Commerce Protocol context including return policy, support settings, merchant item identifiers and consumer notice fields. Microsoft's own scaling guidance is the most testable thing in the edition: when share of voice lost to budget exceeds share lost to rank, adding budget is likely to pay, while high rank loss points elsewhere. Budgets should be set two weeks before peak events and changed in 15% increments. Microsoft's search advertising revenue excluding traffic acquisition costs grew 10% in the quarter ended June 30 2026.

Sky puts its streaming inventory inside Amazon's buying platform

Sky Media and Amazon Ads announced on October 6 that Sky's premium video supply and audiences are buyable in the UK and Ireland through Amazon Ads Agent. Programmatic guaranteed deals are open immediately; invitation-only private marketplace deals are to follow, with no date given. One connection reaches five Sky products, Sky Go, NOW, Sky Glass, Sky Q and Sky Stream, spanning entertainment, drama, live sport and news, plus partner content from HBO Max, Discovery+ and TNT Sports. The sport list runs Premier League football, European and PGA Tour golf, world darts and Formula 1. No audience figures accompany any of it.

Pippa Scaife, Sky Media's director of digital advertising, said the future will be defined by "reducing complexity without compromising on quality," and framed the Amazon route as additional choice rather than replacement. Phil Christer, Amazon Ads' managing director for the UK, described it as a new way to reach those audiences. Sky Media's justification rests on a Barb figure: 70% of UK households have access to a subscription video-on-demand service, from the Barb Establishment Survey for the first quarter of 2026 published on May 21, with a footnote pointing at 20.8 million UK homes. Viewing spread across services and devices, the argument runs, makes buying video at scale complex enough that a single programmatic connection is worth paying for.

What is absent is the commercial substance. No prices, fees or minimum commitments. No impression volumes, formats, ad load or frequency controls. No private marketplace date. No statement of measurement currency, even though Barb launched its Ads Hub on January 27 2026 with a Prime Video integration under way. No confirmation that Ireland's supply and signals match the UK's, and no word on whether Sky's other buying routes change. Sky also says nothing about ITV supply, which is a live question given that Sky agreed in July to buy ITV's Media and Entertainment business for up to GBP 1.6bn, including GBP 1.2bn in cash, the transfer of Love Productions and an earn-out of up to GBP 0.2bn covering ITV's linear channels and ITVX. The Competition and Markets Authority issued its invitation to comment on July 23 with submissions closing August 6. Separately, Comcast said on June 29 it would spin off NBCUniversal and Skyfrom its broadband and wireless business in a tax-free deal expected to take about twelve months, without saying which company keeps FreeWheel and Universal Ads.

The pricing context explains why a broadcaster would do this at all. Guideline analyst Sean Wright estimated that about 82% of global streaming video supply was sold direct about two years ago. Ian Whittaker of Liberty Sky Advisors, drawing on Guideline data, placed UK linear and pay-TV equivalent rates at four to five dollars per thousand impressions against roughly 22 to 25 dollars for broadcaster streaming, and concluded that money leaving the broadcaster ecosystem accrues mostly to Google, Meta and Amazon. Neither set of figures relates directly to these deals. On Amazon's side the plumbing has been renamed twice in a year: Ads Agent arrived on November 11 2025 as an AI campaign management tool, Amazon DSP became DVA at unBoxed on September 29 2026 and was folded into Ads Agent as a conversational interface, and a four-step DVA+ campaign builder entered US closed beta on the same day, with open beta expected in late October and general availability in 35 markets targeted for the first quarter of 2027. The UK sits outside that closed beta, and nothing in the Sky release connects DVA+ to Sky supply. The sequence of UK video deals is now long enough to read as a strategy: Netflix supply in Amazon DSP across eleven markets including the UK from the fourth quarter of 2025, Amazon shopping and streaming signals live on Netflix buys in EMEA from May 18 2026, and Channel 4 opening its video-on-demand supply to five buying platforms on June 22 2026.

Heineken buys a threshold rather than a duration

Heineken is testing Attentive Lift, a method that overlays brand lift results with attention metrics, on display advertising for Birra Moretti and Old Mout, with GumGum. The purpose is unusually candid. Sabrina Court, digital media specialist at The Heineken Group, said the point is to tell a story to the brand team and unlock more budget, and that the main goal was to optimise toward attention so the team understands quality of exposure rather than volume alone. Heineken's marketing and sales investments ran at 10.1% of EUR 14.84bn net revenue, about $16.61bn, in its half-year results. The company is now adopting the method across its UK display activity, including Old Mout cider campaigns.

The Birra Moretti numbers are narrow and specific. Among respondents who looked at the advertisements for more than 2.2 seconds, unaided awareness rose 8% and familiarity with the brand rose 4%. The units were mobile and desktop skins placed next to martial arts, boxing, tennis and gaming content, and desktop skins typically held attention for 4.5 seconds, three times the IAB baseline. GumGum built the threshold by matching brand lift uplifts from On Device against attention time from a test panel of thousands of consumers, producing a baseline score. Each Heineken brand required its own analysis, because each carries different brand values. Matt Goodwin, GumGum's group digital director, described the object as an optimum attention threshold per brand, measured against awareness, familiarity and recall, which an in-house team can then use to guide buying. The tool is currently used for post-campaign reporting, and Goodwin said it could inform mid-flight changes. GumGum is running comparable brand lift work with Cint and Disqo for other clients, and has previously argued that contextual placement doubles ad recall.

The most useful objection came from a competitor. Marc Guldimann, founder of Adelaide, warned that optimising for the longest attention duration creates bad creative incentives, because the most attention-grabbing creative is not necessarily the best vehicle for a brand message. Pairing lift with attention reduces that risk, which is the structural case for a threshold rather than a maximum. Adelaide's own figures show high-attention online video placements outperforming low-attention ones on conversions by 62% for one higher education client, and Google put Adelaide attention scores into DV360 bidding from mid-July.

The budget pressure behind all of this is European and measurable. Advertisers there are increasing spend on social video, retail media and paid search while growth in inventory outside the walled gardens slows, and IAB figures put European social video spending up 25.7% in 2025. At Advertising Week New York, Mattress Firm's chief marketing officer Tom Murray described reorganising advertising around attention, message recall and awareness together, using creative effectiveness data from iSpot and targeting sleepers whose partners snore. He joked that the campaign would never win a creative award, and said the attention score was earned rather than taken as a shortcut.

Apple makes screenshots for an unreleased device mandatory from April 2027

On October 5 Apple posted a developer notice on preparing and submitting apps for the iPhone Duo, stating that apps and games submitted to the App Store from April 2027 must include screenshots made for the device. Apps already optimised for it can be submitted now. The notice sets out three tasks: build or recompile, prepare App Store assets, and submit.

The calendar is tight at the front and vague at the back. The iPhone Duo was unveiled on September 9 2026 at a starting price of $1,999, with a 7.6-inch inner display and a 5.4-inch outer one. Pre-orders open on October 16 and the device reaches customers on October 23, eighteen days after the notice, in roughly seventy countries. The screenshot requirement starts in April 2027 with no day named. Existing apps can be recompiled with Xcode 27.1, which adds development support, though the overview page lists Xcode 27.1 Release Candidate, leaving it unclear whether a final build exists. Device Hub shows app behaviour across poses and orientations without the pages listing what those poses are.

The asset rules are where the advertising consequence sits, because the same images feed paid placements. Apple's screenshot and app preview specifications now cover the device, with downloadable templates for product page headers, search results, app previews, screenshots and in-app events in Figma, Photoshop, Pixelmator and Sketch, and an App Store Connect preview tool for checking how metadata renders. A product page carries up to ten screenshots and up to three app previews, with up to three of each eligible to appear in search results depending on orientation. Prohibited content covers specific prices, discounts, website addresses, copyright symbols, unverifiable claims, references to other platforms, and Apple recognitions such as Editor's Choice or Apple Design Award. Everything displayed must meet a 4+ age rating, so games with violence may use action imagery but not gore or weapons aimed at a person or the audience. Video autoplays and loops, is usually muted, and needs a standalone poster frame. In-app events require a 16:9 event card and a 9:16 details page plus one of seven badges: Challenge, Competition, Live Event, Major Update, New Season, Premiere or Special Event.

Apple's own figure is that nearly 65% of App Store downloads follow a search, which is what makes a screenshot an advertising asset rather than a listing detail. Multiple ad positions began appearing in UK search results on March 3 2026, and custom product pages, tied to Apple Search Ads in January 2022 with a 35-page limit at the time, extend the asset pool into ad variations. Localisation multiplies the work, since text overlays must match every supported language. The notice leaves four things unanswered: whether existing iPhone images will be accepted for the new device, whether updates to live apps must carry the new set, how App Review will treat omissions, and which day in April 2027 applies.

Pinterest publishes 120 search phrases and no denominators

Pinterest released its first Beauty Trend Report on October 6, organised into eleven themes and 120 search queries, each carrying a growth percentage. The top entry is polka dot cat eye nails at plus 1,585%. The page gives no search volumes, no time window and no market, which is the single fact that governs how the rest can be read: a rise from 200 to 3,370 searches and a rise from 200,000 to 3,370,000 both print as plus 1,585%.

The distribution is heavily skewed. Median growth across the 120 queries is 145.5% and the mean is 277.3%, pulled up by outliers. Seventy queries sit at 100% or more, twenty-three at 500% or more, and three above 1,000%, while twenty-five fall below 50%. The lowest figure is 21%, for scalp detox. Nails dominate the top end: the twenty-two nail queries have a median of 713%, fourteen of them clear 500%, and ten of the eleven chrome queries do. The three above 1,000% are polka dot cat eye nails at plus 1,585%, jelly pink chrome nails at plus 1,469% and milky white chrome nails almond at plus 1,402%. At the other end, Skinvested has a median of 39% and Arctic Frost 45%.

The themes are looser than their names imply, which matters for anyone mapping them onto a product range. Pink-a-Boo's top query is a bedazzled claw clip rather than a hair colour. Copper Waves, nominally a shade theme, includes curly wolf cut man at plus 514%. Farm to Face is led by bone broth benefits for women at plus 782%. Pinterest's own framing is that beauty searches are becoming more specific, with a textured French bob rather than a bob and a dark-red cat-eye manicure rather than a red manicure, and that the new beauty flex is specificity.

Two survey figures accompany the report: 85% of weekly Pinterest users say they discovered new beauty trends or ideas on Pinterest before seeing them elsewhere, and 64% say Pinterest is better for discovering beauty trends that feel up to date. Both carry footnote markers whose text was not visible, so sample size, fieldwork dates and the definition of a weekly user cannot be checked. Pinterest's December 2025 Predicts report used normalised English-language queries from September 2023 to August 2025; the beauty report does not say whether the same method applies. The page reads as English (UK) with British spellings, and the market behind the search data is unnamed.

Alongside the report sits a product. Beauty Guides open from a hair or nail Pin in Closeup through a Get the Guide tap, and return a breakdown of the look including what to ask for at a salon, similar inspiration and shoppable products. Availability is US iOS only, with wider release promised over coming weeks, and the page does not say whether the shoppable products are paid placements or organic links.

The commercial logic is plain enough from the release calendar. The beauty report follows the men's report of September 9 by twenty-seven days and Visual Search Ads by nineteen. Pinterest reported first-quarter 2026 revenue of $1.008bn, up 18%, with 631 million monthly active users and Performance+ campaigns accounting for 30% of lower-funnel revenue, after launching Top of Search ads in beta on September 25 2025 with a reported 29% higher click-through rate in testing. A dated list of search phrases from the platform that sells advertising against them is a media product as much as an editorial one.

A B2B data vendor publishes a price, and it starts at $22,000

HG Insights released a bundled Contextual Intelligence Platform on October 6 from Santa Barbara, and the detail worth noting is that the company published a number: the first Copilot tier starts at $22,000. Pricing otherwise runs on consumption-based Intelligence Credits, and the release does not say whether $22,000 is annual or per user, how many credits it includes, or how usage is metered. In a category that almost never prints a figure, one number is still one number.

The components went live the same day. Fabric is the data foundation, licensable separately through an API or a direct feed into a customer's warehouse, CRM or models. HG Copilots covers Market Analyzer, Data Studio and Sales Copilot. HG Agents automate go-to-market tasks and are reachable from Slack, Microsoft Teams, the HG interface or a customer's own applications, with an HGSuperagent routing requests to specialists; the company claims a research brief returns more than thirty data points within ninety seconds. An MCP server lets outside AI systems query market, account and buyer data, TrustRadius reviews, SEC filings, federal contract data and live web research, with Amazon Quick, Anthropic, OpenAI and Microsoft named as reachable ecosystems. Customer Voice, powered by TrustRadius, draws on a community of more than 12 million technology buyers. Fabric, the Copilots and the MCP server are listed on AWS Marketplace; the Agents and Customer Voice are not, with no reason given.

Rohini Kasturi, the chief executive, put the thesis in one line: when infrastructure is equal, "whoever knows the account best wins," and "most vendors still sell disconnected data and call it intelligence." The problem the company describes is specific to agentic tooling rather than to dashboards. An agent drawing on isolated datasets without entity resolution produces named failure modes: one company appearing as several accounts, subsidiary spend missed or double-counted, existing customers flagged as new prospects, competitor accounts pursued as winnable, and job titles substituted for evidence of product ownership. Michael Levy of GZ Consulting said account data still lives across a dozen separate systems in most organisations, without a stated basis. Firmable has found that B2B sales teams estimate 32% of their CRM data is flawed, and an Intentsify white paper of May 2026 cited a Gartner forecast that more than 40% of agentic AI projects will be cancelled by the end of 2027.

The scale claims are company-supplied and unaudited: more than 70 million companies, more than 640 million verified installs with fifteen or more years of change history, more than 150 IT, cloud and AI spend categories, more than 200 million contacts across more than 190 countries, more than 20,000 research topics and 36,000 products tracked for intent, and technology usage by function across more than 10 million companies. The pipeline runs Collect, Connect, Validate, Confirm, Deliver, and no match rates, confidence thresholds or expert-review shares are published. A new Momentum signal tracks whether vendors, portfolios or products are gaining or losing ground, with no formula, refresh rate or weighting disclosed. Fit, Need and Intent models now update dynamically and are described as deterministic, explainable and adjustable, which is a claim about lead scoring that the screenshots only partly support: segment thresholds of 5.95, 3.67 and 2.49 produce a top band holding 10% of leads and 40% of conversions at a 23.23% conversion rate against a bottom band at 1.55%, roughly fifteen times apart, with the source of the data unstated. A second screenshot shows Workday with 17.7 thousand employees, a 2005 founding date, an estimated IT budget of $1.0bn and 109 contacts across eleven pages.

The company's own history runs through advertising rather than beside it. As HG Data it distributed technographic audiences to ad platforms through LiveRamp. It acquired TrustRadius in June 2025 and launched its first copilot in July 2025. On October 30 2025 Bombora named it one of four initial partners in curated ecosystem audiences, supplying technology adoption, spend and purchase-intent signals, segments that later reached Reddit. TrustRadius research covering 1,862 buyers found 83% shortlist three or fewer products, averaging 2.7, and a Semrush survey of 519 B2B buyers found 92% saying AI shaped their shortlist. Which is the commercial case for data enrichment priced in credits, and also the reason nobody independent has yet tested whether the records are right.

Google stops showing star ratings on healthcare pages

Google appears to have stopped displaying review rich results for healthcare web pages carrying review markup. Barry Schwartz reported the change on October 7, crediting Andrea Badder of Schema App, who first observed it and published on the subject. Schema App saw review snippet visibility drop across its healthcare customers beginning in late May, with a second drop affecting some organisations in early August, and supported the observation with a Search Console performance chart filtered by search appearance.

What makes the finding useful is the troubleshooting that failed. Schema App's customer success team tried several different ways of connecting aggregateRating, alternative property paths, and multi-typing healthcare entities with Product. None restored review snippet performance, which points at a deliberate removal on Google's side rather than at markup error on the publisher's. Schwartz noted that Google has a history of switching off specific rich results, including for particular industries, and said he does not personally see review snippets for doctors, dentists and similar queries outside local results.

The response from the affected organisations is the part with the longer tail. Badder said many healthcare customers would not change markup that is semantically accurate merely to recover a rich result, because their priority is describing physicians, locations, services, specialties and relationships correctly for both search engines and AI systems. "Rich results are still valuable when Google awards them," she wrote, which is a precise way of saying the award is Google's to withdraw. Accurately describing an organisation to machines, on Schema App's account, is now the primary job and earning any specific rich result a smaller part of it.

The precedent is well established. Google killed FAQ rich results in 2023 after four years of signals and deprecated seven structured data types to simplify results. Healthcare is also the category where YMYL standards apply most heavily, which gives a plausible reason for removing a self-reported star rating from medical results without making the removal any less expensive for the organisations that built structured data around it. No Google announcement accompanied either drop.

Who pays when an agent spends faster than the evidence arrives

The sharpest framing of the week's underlying problem came in an opinion column rather than a product release. Writing in AdExchanger on October 7, Evgeny Popov argued that an AI agent can stay inside its budget and still overspend, because evidence of performance arrives after the money is committed. He calls the interval the proof gap.

The worked example is deliberately small. An agent is handed $200,000 and spends at $1,000 an hour. The first sales readout takes 48 hours to mature. By the time it exists, $48,000 has already been committed to a change nobody has validated. Campaign budgets and pacing controls do nothing about this, because neither limits how much money can ride on an untested decision.

Popov proposes two instruments. The first is a risk budget capping what an agent may commit under an unvalidated change, with four properties: an agreed evidence standard, a time frame matched to the goal, a default to pause when evidence is late, and rules placed outside the agent's objective function so it cannot optimise around them. In his example the risk budget for testing a new mix is $12,000 rather than the full $200,000. The second is a contract guarantee assigning a defined financial loss to a funded counterparty when an agreed condition is missed. He draws the precedent from sequential liability, the early-1990s dispute over who bears the loss when money does not arrive, and from the make-good, and he distinguishes third-party-backed guarantees, which Munich Re began writing for AI vendors in 2018, from seller-funded credits of the kind TikTok offers when return on investment falls below 90% of a daily target. The distinction is whose balance sheet absorbs the miss.

His closing prediction is dated and falsifiable, which is rare in the genre: by the end of 2027, at least one media deal between independent firms will settle from escrow with a cash remedy for counting errors. Set against the rest of the week, it reads less like a forecast than a description of where the other stories are heading. Pixalate has just started publishing which properties buyers demanded refunds on. Lunio is selling education brands a per-campaign invalid traffic rate. IAB Tech Lab is signing its first Trusted Server publisher because an operating system blocked nine domains without notice. In each case the open question is the same one Popov names: when the number turns out to be wrong, whose money moves.

Also noted

  • October 7 Google is testing a Direct offer asset in Google Ads, appearing with a beta label beside Promotion, Message and Lead form, serving only on campaigns using AI Max or text customization, with an offer name, an AI-read description, incentive type and amount, optional single-use promo codes uploaded by CSV, start and end dates and an optional daily cap by claim value or count; in AI Mode it surfaces as a Claim one-time code bar under a sponsored result, spotted by Christian Cabrera. PPC News Feed
  • October 8 Frank Cartagena's four-person agency OK Future took a Goodwipes social spot from concept to execution in four days against projected production costs of $700,000, working round-the-clock shifts in which one team member threatened to quit at 2 a.m. and rejoined at 7 a.m.; the spot drew more than 750,000 Reddit views in 48 hours and almost 1.5 million across subreddits, and Cartagena said he "really wouldn't recommend that this be the norm." Digiday
  • October 6 Ferrero will run two Super Bowl LXI spots on February 14 2027, for Nutella and Kinder Bueno, doubling its Super Bowl spending inside a $50m sports marketing investment for the year, after its first Super Bowl appearance drove what it reports as a 140% year-on-year lift in volume sales and a Nutella jar filmed aboard Artemis II drove a 20% lift in Amazon unit sales. Adweek
  • October 6 Fanta reports 7 billion impressions for its Halloween work without paid media behind it, a claim that rests entirely on how impressions are counted and against which denominator. Adweek
  • October 6 Google Search Console added multi-select country filters, letting sites view several markets as a single total rather than one at a time, a small change with real consequences for anyone reporting on regional groupings. PPC Land

By the numbers

  • 9.3% Click-weighted invalid traffic on education campaigns aimed at competitor brand names, against 4.7% on generic ones. Source
  • EUR 13bn Google's stated Finnish investment across 2027 and 2028, now running ahead of two unfinished environmental assessments. Source
  • 10,000 Companies that can sit in a single Microsoft Advertising targeting or exclusion list, up from a file-import answer of no. Source
  • $22,000 Entry price for HG Insights' first Copilot tier, rare for a category that normally prints nothing. Source
  • April 2027 When App Store submissions without iPhone Duo screenshots stop being accepted. Source