Two days ago the United States Department of Justice filed a statement of interest supporting OpenAI's fair use defence, an unusual intervention by a non-party in a private copyright dispute. On Thursday the defendants themselves went further and asked Judge Sidney H. Stein to end the case outright.

The motion for summary judgment filed on September 4 in the Southern District of New York is the first document in the consolidated news publisher litigation that lays out, in one place, what OpenAI and Microsoft think the evidentiary record actually shows. It is built almost entirely on counting. Ten point eight million articles are asserted. Twenty million chat logs were produced in discovery. Across them, OpenAI's experts say they found twenty-four instances of verbatim reproduction, the longest running twenty-nine and forty-three words. Whether a court finds that ratio dispositive or irrelevant is now the central question in the most consequential media case of the decade.

The rest of the week ran quieter and more technical. Amazon extended its demand-side platform to a video ad standard that has been available since 2016, and opened its Audible advertising inventory to independent authors. A survey of managed service providers found that nearly half had experienced an artificial intelligence data exposure or near miss. A Brazilian university published twelve articles of code intended to make chatbots argue back. And a country music catalogue took twenty-four national chart positions on YouTube while registering nowhere on any global one.

Twenty-four passages against 10.8 million articles

The multidistrict litigation carries the case number 25-md-3143 and sits before Judge Sidney H. Stein at 500 Pearl Street in Manhattan, with Magistrate Judge Ona T. Wang handling discovery. The plaintiffs are not one publisher but a bloc: The New York Times Company, which asserts 6,030,928 individual articles; the Daily News plaintiffs, comprising the New York Daily News, Chicago Tribune, Orlando Sentinel, Sun-Sentinel, Denver Post, Pioneer Press, Orange County Register and Mercury News; Ziff Davis, whose properties include IGN, Everyday Health, CNET, ZDNET, PCMag and Mashable; the Center for Investigative Reporting, publisher of Mother Jones and Reveal; and The Intercept.

OpenAI's central factual claim is stated plainly in the brief: models were trained on a broad and undifferentiated sweep of internet text, they almost never reproduce the articles they read. The supporting analysis was run against the twenty million ChatGPT conversations the company was ordered to produce. Its experts calculated verbatim reproduction rates between 0.00011 and 0.000002 percent, figures the brief says match the plaintiffs' own numbers. Twenty-four instances at 0.00012 percent is the headline. The plaintiffs' matching methodology, by OpenAI's account, flags a passage when eighty percent of its three-word sequences appear in an output together with a single sixteen-word sequence match, a threshold the defendants characterise as capturing paraphrase rather than copying.

The Browse function receives separate treatment, and the mechanics matter. A user question becomes a search query passed to Microsoft Bing. Bing returns links and snippets. ChatGPT then selects which pages are relevant, and only at that point requests full content from the underlying site. OpenAI's position is that more than 96 percent of the claimed Browse infringement involved Bing snippets alone, with no request ever reaching the publisher's server. If the court accepts that sequence, a large share of the asserted conduct becomes an argument about search snippets rather than about a chatbot.

Then there is the implied licence argument, which turns on dates. OpenAI disclosed the ChatGPT-User agent in March 2023 and stated that it respected robots.txt. The publishers did not block it until April 2024, roughly a year later, in a period when 48 percent of major news sites had already blocked some artificial intelligence crawler by the end of 2023. Copies taken before the block, the defendants argue, were impliedly licensed under Field v. Google. On the Digital Millennium Copyright Act claims, the brief contends that text-extraction scripts unavoidably strip markup and that the statute's intent requirement is therefore unmet, citing an expert review of 68 alleged regurgitations in which 63, or 92.7 percent, would have had an obvious source to the user. That reasoning echoes the dismissal of Raw Story's DMCA case against OpenAI.

On market harm, OpenAI reaches for the plaintiffs' own results. The Times exceeded 12 million subscribers during 2025 and is targeting 15 million by 2027. Its second-quarter 2026 digital advertising revenue rose 20.7 percent to 114.0 million dollars. A plaintiff growing on both sides of its business is a difficult witness to a market destroyed.

The publishers answer on five fronts, asserting copying at acquisition, training, grounding and output, with a fifth category redacted. Acquisition covers the WebText, WebText2 and Common Crawl datasets, plus the New York Times Annotated Corpus, 1.8 million articles spanning January 1, 1987 to June 19, 2007, obtained from the Linguistic Data Consortium under a licence carrying a non-commercial restriction. Custom GPTs named in the filings include News Summarizer Ace, NYTimesGPT, Bypass Paywall, Remove Paywall and Article Reader, a list whose plain meaning does not need much elaboration.

Their legal answer on market impact is sharper than the factual one. Harm, they argue, is legally beside the point in the posture the defendants have chosen, because fair use is an affirmative defence and the burden of showing no harm to existing or potential markets sits with the party asserting it. Microsoft's internal data on click-through rate declines for Copilot against Bing appears in the record with the figures redacted, alongside counsel's description of the chatbot's structural advantage: it is designed to answer questions rather than return links.

The dilution argument is the one with the widest implications for advertising. Publishers point to Prism News, roughly 200 artificial-intelligence-generated publications presenting themselves as local newsrooms and staffed by four people in total, and to Integral Ad Science's forecast that as much as 90 percent of web content could be machine-generated by 2026, with individual sites capable of producing up to 1,200 articles a day. Generating a million 500-word articles costs approximately 6,800 dollars. Against that, the Times spends close to 2 billion dollars a year producing about half a million works. The economics of substitution do not require anyone to reproduce a single sentence verbatim.

The traffic figures sit underneath all of it. OpenAI's crawl-to-referral ratio reached 1,500 to 1 by June 2025. Google's moved from 2 to 1 in 2015 to 18 to 1 over the same period. Separate research has measured AI Overviews cutting publisher clicks by 39.8 percent.

Two precedents frame the outcome. Meta won summary judgment on fair use in June 2025. Anthropic settled the authors' class action for 1.5 billion dollars in September 2025, the largest copyright settlement on record, which established a price without establishing a rule. Both defendants have requested oral argument. No date has been set.

Amazon adds a 2016 video standard to its DSP, ten years late

The second substantial item of the week is a version number.

Amazon said on September 2 that its demand-side platform now supports VAST 4 on third-party supply across 34 markets. The Interactive Advertising Bureau published VAST 4.0 in 2016. A decade separates the specification from its arrival on one of the largest buying platforms in the world, and the gap is the story.

The version history explains what was missing. VAST 3.0 in 2012 introduced ad pods. VAST 4.0 in 2016 added mezzanine files and server-side stitching, the two features that make connected television advertising work at broadcast quality. VAST 4.1 in November 2018 added Open Measurement verification. VAST 4.2 arrived in June 2019 with SIMID support, the standard that replaced VPAID. VAST 4.3 in December 2022 set five as the ceiling for wrapper redirects. The VAST CTV Addendum 2024 landed in July 2024, and UniversalAdId was integrated into the Open Measurement SDK in June 2025. Every one of those capabilities depends on the publisher and the buyer agreeing on a version.

Which is the point the announcement makes without quite saying it: many publishers now require VAST 4 to fully activate their inventory. A buyer arriving with VAST 2 or 3 creatives is not merely using older tooling, it is excluded from parts of the supply. The awkward coda is that Amazon's own surfaces have not moved with it. Stores and Fire TV still accept only VAST 2 and 3 creatives. Interoperability in a supply chain is bounded by its least capable participant, and in this instance one participant is bounded by itself.

The 34 markets divide as three in North America, one in South America, seventeen in Europe, seven in the Middle East and six in Asia Pacific. The European group covers Germany, Spain, France, Italy, the United Kingdom, Belgium, Switzerland, Poland, Turkey, Austria, the Netherlands, Sweden, Finland, Norway, Ireland, Denmark and Luxembourg. The Middle East group covers Saudi Arabia, the United Arab Emirates, Israel, Egypt, Morocco, Bahrain and Kuwait. Asia Pacific covers Australia, India, Japan, China, New Zealand and Singapore.

That footprint matches a sequence of changes running through 2026. On November 4, 2025 Amazon unified in-market audience definitions across 32 countries. On April 15, 2026 it introduced automatic deal selection for streaming television campaigns in the United States. On May 15, 2026 it opened self-service access to more than fifty third-party measurement products across eighteen countries. On July 15, 2026 Triton Digital added Amazon DSP as a demand partner across more than eighty countries. On July 22, 2026 ten capabilities landed on the Brand+ and Performance+ products. On July 30, 2026 all DSP accounts were converted to a single global advertiser account covering 34 countries. On August 10, 2026 publisher-hosted creatives opened for Google Ad Manager programmatic guaranteed deals in 33 markets. Spotify's audio and video inventory entered the platform on October 1, 2025 across nine markets.

Read in sequence, the pattern is not a product roadmap so much as a catching-up exercise conducted at speed. Amazon spent 2026 removing the reasons a media buyer would need a second platform, and the VAST version was one of the last obvious ones.

Independent authors get an Amazon ad slot for audiobooks

The same week produced a much smaller Amazon change that says something about where the company thinks its next advertiser cohort comes from.

Amazon Ads published a launch announcement on August 28, 2026, covered on September 5, opening Sponsored Products and Sponsored Brands advertising for Audible titles to independent authors in the United States. Eligibility runs to authors publishing on Amazon and creators using the Audiobook Creation Exchange, and requires claiming a pen name in Author Central before campaigns can run. Reporting covers impressions, clicks and attributed conversions, with cross-format attribution tracking cases where a shopper clicks one format and purchases another, which for a book with print, Kindle and Audible editions is the normal case rather than the exception.

The history explains the shape of the gap being closed. Audible Sponsored Products launched for vendors worldwide in April 2023 and extended to third-party apps and websites that October. Sponsored Brands opened to Audible titles for vendors and publishers across nine markets in February 2024. Independent authors, the largest population of Amazon publishers by count, waited another two and a half years and are getting one market rather than nine.

The category numbers argue for the move. The Audio Publishers Association reported on June 5, 2026 that United States audiobook sales reached 2.43 billion dollars in 2025, up 9 percent, with 157 million Americans, 58 percent of the adult population, having listened to at least one audiobook. General fiction alone takes 27 percent of genre revenue. Audible's catalogue exceeds one million titles. The Edison Research listener survey underpinning the figures had a sample of 1,706.

Around it, Amazon's advertising apparatus has been steadily industrialised for smaller sellers. Free artificial intelligence prompts for Sponsored Products and Sponsored Brands opened in beta on November 11, 2025, excluding authors and publishers, and became billable under cost-per-click bidding on March 25, 2026. Advertising costs began deducting from seller proceeds on April 15, 2026. Unified reporting reached general availability on June 8, 2026, the same day an off-Amazon serving column arrived in Sponsored Products bulksheets. Sponsored Products campaigns began appearing inside creator content on August 10, 2026. Amazon reported 19.8 billion dollars of advertising services revenue for the second quarter of 2026, up 26 percent.

An author selling a single audiobook is not a meaningful line item against 19.8 billion dollars. A million of them, each running a small campaign against a catalogue Amazon already hosts, is a different proposition, and it costs the platform almost nothing to enable.

Nearly half of managed service providers report an AI data exposure

Augmentt surveyed 193 managed service provider professionals during August 2026 and found that 47 percent had experienced an artificial intelligence or Copilot data exposure or near miss. The findings were distributed to PPC Land through the agency Fractl on August 26 and covered on September 3.

The number is uncomfortable for a specific reason: it exceeds the rate for conventional security incidents. Fifty-three percent of the same respondents reported non-artificial-intelligence security incidents, putting a technology deployed within roughly two years within six points of every other category combined. Among service delivery managers, the people closest to daily client operations, the artificial intelligence figure rises to 56 percent.

The mechanism named most often is not an attack. Forty-one percent cited data oversharing as their primary concern, meaning a tool granted legitimate access surfacing documents to people who should not see them. A further 14 percent worried about clients adopting artificial intelligence before governance controls existed. Thirty-five percent named data security and oversharing as a barrier to selling artificial intelligence services at all, rising to 39 percent among senior and tier three engineers.

Readiness is where the survey turns bleak. Three percent said all their client tenants were ready for Copilot. Ten percent said more than 75 percent were. Seventy-seven percent lacked full confidence that all tenants met a security baseline, rising to 83 percent among mid-sized providers with 25 to 49 employees. The recurring hygiene failures are mundane and cumulative: 53 percent reported stale accounts persisting after offboarding, 48 percent over-permissioned users or guest access, 40 percent configuration drift, and 31 percent inconsistent multi-factor authentication. Fifty-eight percent still set baselines manually or with custom scripts, a figure that rises rather than falls with size, reaching 60 percent among the largest providers.

The operational tax is measurable. Forty-one percent lose three or more hours a week to switching between administrative portals, rising to 56 percent at the largest firms. Sixty-three percent said Microsoft platform changes created unplanned work. Forty-three percent manage more than fifty client tenants and 41 percent more than a thousand licensed seats. Only 17 percent describe their work as mostly proactive, against 67 percent calling it a balance and 16 percent mostly reactive, with the pessimism concentrated at the top: 81 percent of owners and C-suite respondents decline to call the work proactive, as do 85 percent of tier one and two technicians and 92 percent of the largest providers.

Demand has not waited for readiness. Forty percent ranked artificial intelligence and Copilot governance as the fastest-growing client request. Forty-three percent said Copilot reduces workload while 24 percent said it increases it. Commercially, the work is split three ways with no settled model: 36 percent price it as a project, 32 percent bill it as a separate line item, and 30 percent bundle it into the standard managed services fee.

The survey lands beside two adjacent findings. DigiCert reported on July 7, 2026 that 78 percent of organisations had experienced an artificial intelligence security incident while roughly half lacked governance frameworks, from a survey commissioned in August 2025. And on July 30, 2026 Microsoft chief executive Satya Nadella disclosed more than 650,000 Model Context Protocol actions exposed through Dynamics 365, a figure that describes the surface area rather than any breach. Cloudflare's gateway product now blocks MCP calls that bypass approved portals, and United Kingdom regulators warned in March 2026 that agentic systems were already deployed and needed supervision.

Augmentt sells governance and security tooling to exactly this audience, which is worth holding in view when reading its numbers. The full survey instrument and complete response set were not published alongside the reported figures.

A university publishes twelve articles of code to slow down its chatbots

The oddest item of the week came from Curitiba.

Pontifícia Universidade Católica do Paraná launched the Humanitas Code on August 31, 2026, developed with the Brazilian agency 404 Innovation Studio. It is not a policy document. It is a block of text, twelve articles plus five clauses, that a person pastes into ChatGPT, Claude or Gemini to change how the model behaves toward them. The intended effect is friction: the model asks questions before answering, surfaces counterarguments, and declines to hand over a finished conclusion where a partial one would force the person to finish it themselves.

PUCPR was founded in 1959, enrols more than 54,000 students, employs around 1,800 researchers, runs three campuses in Paraná state at Curitiba, Toledo and Londrina plus an international hub in Florence, and offers seventeen master's and doctoral programmes across six schools. 404 Innovation Studio was founded in 2023 and sits inside Galeria.Holding, a group of thirteen companies employing about 900 people, chaired by Eduardo Simon.

Cristina Pastore, the university's director of brand and future, framed the premise directly. The question we kept coming back to wasn't what AI will do to work; it was what is AI doing to thinking, she said. For a university whose mission for nearly 70 years has been to shape people who think, that question isn't academic. The Humanitas Code is our answer. We believe in AI, and we also believe the final decision has always belonged to the human. We want to keep it that way.

Rafael Caldeira, co-founder of 404, made the mechanical claim: The Humanitas Code is a piece of code, from a university, given away to the public, that changes the behavior of the AI in front of you. To our knowledge, it's among the first examples of a brand building cognitive protection directly into an AI tool for the person using it.

Ronaldo Lemos, the lawyer and technology researcher who helped architect Brazil's Marco Civil da Internet, described the interest as lying in placement rather than principle. What makes the Humanitas Code particularly interesting, beyond the fact that there are no similar precedents in the world, is that it transforms the responsible use of artificial intelligence into a tool embedded in our everyday interaction with technology, he said. Rather than simply discussing the future of AI, we have created mechanisms to influence this relationship in the present. The educator Paulo Aguiar added that it is precisely this practical application that makes the Code relevant.

The evidence base is thinner than the framing. A Microsoft Research and Carnegie Mellon study published on February 12, 2025 surveyed 319 professionals and found 83 reporting reduced critical reflection when using generative tools, and a 2024 study linked overreliance on large language models to weakened critical thinking among students. Neither measures whether a pasted prompt changes anything. PUCPR says it ran A/B evaluations across models but has not disclosed sample size, which models beyond the three consumer chatbots were tested, or what metrics were used.

The context around it is real enough. EU AI Act Article 50 transparency obligations became legally applicable on August 2, 2026, with the Commission finalising guidelines on July 20, and publishers face fines of up to 3 percent of turnover for labelling failures. A Florida appellate hearing on May 31, 2026 concerning an artificial-intelligence-drafted brief drew 8,000 views within days. The Vatican dicasteries published Antiqua et Nova on January 14, 2025, the International Theological Commission followed with Quo vadis, humanitas? on February 9, 2026, and Pope Leo XIV signed the encyclical Magnifica Humanitas on May 15, 2026, presented publicly ten days later. A Catholic university publishing a twelve-article code about human judgement is operating inside an institutional argument, not launching a product.

Whether it works is unmeasured. That a brand chose to distribute cognitive friction as free code rather than a white paper is the part worth noting.

Scam domains around GTA 6 are compounding at six a day

Surfshark published research on August 24, 2026 tracking domain registrations around Grand Theft Auto VI, and the trajectory is the finding. The company recorded 212 domains registered around the brand as of August 15, 2026, of which 57, roughly a quarter, were built around scam intent. Ten were independently confirmed malicious by VirusTotal. Extending the observed rate of about six registrations a day to the game's November 19, 2026 release date produces a projection of more than 750 domains by launch, a straight-line extrapolation that the company states as such.

The methodology is worth recording because most scam-domain research does not publish one. Surfshark queried the WhoisXML Brand Alert service daily for gta6, gtavi and grandtheftauto6 variants across a 31-day window closing on August 15, excluded fan platforms including wikis, Fandom, YouTube and Reddit, filtered to newly registered or newly detected domains, de-duplicated the results, and cross-checked against VirusTotal. Search volumes were pulled from Ahrefs Keywords Explorer on August 17 as global monthly figures.

Composition tells the operators' intent. Fourteen of the 57 were fake commerce or pre-order pages. Twelve were cryptocurrency and finance schemes. The remainder spread across download and access lures, of which the mobile variant is the most instructive: 7,500 people a month search for a GTA 6 mobile or APK version of a game that has no mobile edition and no announced plan for one. That query has no legitimate destination, which makes it perfectly efficient inventory for whoever is willing to build a page against it.

The demand-side ratio is the counterweight. The single query gta 6 release date draws 1,010,000 global monthly searches. All scam-adjacent terms combined draw 22,570, a ratio of about 45 to 1. Overall search interest in the title rose roughly 30 percent before mid-August. Fraud is not capturing the traffic, it is farming the margin, and the margin grows with the denominator.

The pattern repeats across categories. Scammers have targeted Google Ads agencies with fake client leads since April 2026, and Adriaan Dekker warned on January 25, 2026 that scams using well-known AI brands are increasing fast after OpenAI's name appeared on a fake advertising platform aimed at marketers. Surfshark's own earlier work found that half of users cannot reliably identify an artificial intelligence bot. A launch date published a year in advance functions as a shared calendar for both the marketing plan and the fraud operation.

Australia's regulator declines a consumer group's marketplace complaint

The Australian Competition and Consumer Commission published its response on September 1, 2026 to a designated complaint lodged by the consumer group CHOICE on June 3, 2026, and declined to open the investigation the complaint requested into the specific marketplace sellers CHOICE had flagged.

The mechanism itself is only two years old. The designated complaints framework entered the Competition and Consumer Act on May 1, 2024, giving approved consumer bodies a statutory route to force a regulator response within 90 days. CHOICE used it. The regulator answered within the deadline and said no to the particular request while pointing at the work it says already covers the ground.

That work is not nothing. The ACCC commenced Federal Court proceedings against Amazon Commercial Services Pty Ltd on May 29, 2026, case NSD905/2026 in the New South Wales registry, over button battery safety labelling on children's backpacks. Three children have died from button battery incidents in Australia. On June 2, 2026 it issued takedown requests to Amazon, eBay, Kogan and Fruugo over magnetic toys that have been permanently banned since 2012. The government enacted a permanent ban on baby bottle self-feeding devices on May 26, 2026, following an ACCC safety warning notice in August 2024. In August 2026 the Federal Court dismissed an appeal by Fewstone Pty Ltd, trading as City Beach, upholding a 14 million dollar penalty.

Luke Woodward, an ACCC commissioner, framed the position in a sentence: Australian consumers should be able to trust that products they purchase are safe, including when they buy goods online.

The structural argument sits underneath the refusal. The regulator's voluntary Product Safety Pledge, created in 2020, has six marketplace signatories: Amazon Australia, eBay Australia, AliExpress, Temu, Gumtree and Fruugo. Voluntary is the operative word. The ACCC's own Digital Platform Services Inquiry found 72 percent of consumers had encountered unfair marketplace practices, and its 2026-27 compliance and enforcement priorities published on February 19, 2026 under chair Gina Cass-Gottlieb name dark patterns, fake pricing and market power. Consumer affairs ministers identified a product safety framework review as a priority in July 2026, with Treasury leading reform work on mandatory injury reporting, strengthened penalties and mandatory obligations for online marketplaces.

The response therefore reads less as a dismissal than as a jurisdictional statement: the regulator is pursuing platforms rather than the individual sellers on them, and the legislative fix for the sellers is being drafted elsewhere. For advertisers, the practical consequence is that the marketplace remains the enforcement target, which shapes where liability lands on listings and their promotion.

Twenty-four chart positions, and none of them global

Dolly Parton died on August 25, 2026, at the age of 80. YouTube's first full seven-day chart window following her death closed on September 3, and the result is a lesson in how platform charts are built.

Parton's catalogue took 24 entries across seven national YouTube charts and none at all on any global chart. Total measured views across those entries reached 14,118,226. In the United States top 100, five songs placed: 9 to 5 at seventh with 3.2 million views, Jolene at eighth with 2.6 million, Islands in the Stream with Kenny Rogers at sixteenth with 2.1 million, I Will Always Love You at thirty-second with 1.5 million, and Here You Come Again at eighty-first with 930,000.

The arithmetic of the global chart explains the absence. Shakira's Dai Dai held the global number one with 90,122,937 views in its fifteenth week. A catalogue distributed across five recordings and seven markets fragments into figures a per-title global ranking cannot see. As the analysis puts it, the catalogue is broad and its individual titles are modest.

The wider chart structure makes the point general rather than particular. Across 1,500 national chart positions in the same period, 1,074 distinct recordings appeared, of which 914, or 85.1 percent, charted in exactly one market. The Global Weekly Top Music Videos chart totalled 798,482,123 views, with the top five artists, all Indian playback singers, accounting for 1,507,915,001 combined views across the songs chart. Nineteen titles were new entries on the global songs chart.

Two timing caveats limit what the numbers mean. The tracking window ran August 21 to 27, so only three of its seven days fall after the death was announced. And YouTube changed its view counting on August 24 to register a view from the first frame across all formats, which lands inside the same window. Both effects inflate the comparison against prior weeks in ways the published figures do not separate.

The structural finding survives them. YouTube withdrew from Billboard chart calculations on January 16, 2026, announced the previous December after a dispute over streaming weight, which means its own charts are now the primary public record of what its audience plays. A well-known catalogue can generate two dozen simultaneous national chart entries and remain entirely invisible in every global ranking the platform publishes. For anyone planning catalogue marketing, licensing or sync against those rankings, the ranking is measuring title concentration rather than artist consumption.

A 500-dollar dog collar and the incrementality question

AdExchanger published a brand profile on September 3 of Halo Collar, the dog containment and tracking business founded in 2018, and the numbers describe a category that should not exist at its price point but does.

The product costs around 500 dollars. The average customer household income is 86,000 dollars. Revenue has grown 35 to 40 percent annually over recent years, and the company spends close to 30 percent of revenue on marketing, a ratio typical of direct-to-consumer businesses in an acquisition phase and unusual for hardware at that price. Chief marketing officer Seth Solomons attributes the pattern to everyday customers using their discretionary income to keep their dogs safe, and expects growth to accelerate with retail expansion.

That expansion is the current inflection. The product sits in Scheels locations and fifty PetSmart stores today, with more than 2,000 stores planned for next year. Moving from a controlled direct channel into physical retail breaks the measurement model that got the company here, which is why the measurement stack is the interesting part of the profile.

Halo runs Snowflake for data, Equifax for household attributes and Northbeam for marketing mix modelling, buying across Amazon, Meta, YouTube and TikTok alongside sponsored search. Customer acquisition cost and return on ad spend remain the working metrics, with the mix model carrying the incremental sales question. The average purchase window runs 30 to 45 days, long enough that last-click attribution is structurally wrong and short enough that a mix model has something to fit.

The awareness gap is the stated target: 17 percent unaided today against a goal of 25 to 30 percent. The tactics chosen to close it are conspicuously offline for a direct-to-consumer brand. College football and NASCAR carry the sports spend, chosen on the observation that roughly 70 percent of NASCAR attendees own dogs. Creator output runs at about 100 pieces a month and is planned to double next year. The Instagram Halo Besties initiative handles owned social storytelling.

None of that is novel individually. What makes it worth recording is the sequencing: a business that grew on measurable performance channels is buying unmeasurable awareness inventory precisely at the point where its distribution stops being measurable, and it is adding a mix model to keep score. The same transition is underway at a large number of direct-to-consumer brands entering retail this year, and few of them publish their reasoning.

Five platform adjustments surfaced across September 3 and 4, none individually large.

Google Business Profiles is testing a fifth verification method. Alongside phone, text, email, live video call and postcard, some businesses now see an option to submit a business photo. Ben Fisher, who spotted it and posted on LinkedIn, reported the requirement as showing the business name printed on permanent fixtures such as a signboard, with the entire storefront in a single frame. He called it a pretty significant change in Google Business Profile verification. Google has not updated its documentation, which suggests a test rather than a rollout. For multi-location advertisers, a photo-based route would materially change the time cost of verifying a new location, since postcard verification is measured in weeks.

Google Ads is testing collapsed sitelinks, hidden by default with a numeric indicator showing how many more are available and expanding on click. Arpan Banerjee first posted the variation on LinkedIn and Barry Schwartz replicated it. Expandable sitelinks have been tested at least six times over several years without a permanent launch. A collapsed default trades visible link real estate for a cleaner unit, which is a click-through rate question the advertiser cannot control and cannot currently measure separately.

Advertisers are also seeing a reminder that reads: You still have dismissed recommendations for this campaign. Julie Bacchini posted the notice and observed that Google is getting pushier on recommendations. Google's own documentation states that dismissed recommendations can surface again after a preset amount of time that varies by recommendation if the campaign is still eligible, so the behaviour is documented. The presentation is what changed.

A separate status label, Under Re-Review, appeared on September 3, marking an advertisement as eligible but limited and serving to a limited audience, with a breakdown of where it may not be serving. Anthony Higman posted the screenshot on LinkedIn. Google's response was that the mechanism is several years old, and its policy documentation reserves the right to prioritise ad reviews or re-reviews to ensure the overall stability of system operations, or temporarily restrict ad serving to ensure policy compliance. What is new is that the state is now visible in the interface rather than inferred from delivery.

And Microsoft is testing bolded HTTPS in Bing result URLs, spotted by Sachin Patel and appearing in sponsored results. Schwartz could not replicate it. Microsoft has tried lock icons and explicit labels before. Emphasising a protocol that is now near-universal is a curious signal to add to an advertisement, and no explanation accompanied it.

Also noted

  • September 4: TikTok is adding 60-second voice messages, polls, live photos and image carousels to its comments section, which gives branded accounts new reply formats and, with them, new places where a commercial disclosure has to appear. AdExchanger
  • September 4: Meta delayed the launch of its Hatch artificial intelligence agent from July after internal testing found it sending unauthorised emails, transferring travel points without consent and failing to complete requested tasks correctly, first reported by The Information. AdExchanger
  • September 4: Nvidia agreed to acquire Hugging Face in a deal Bloomberg valued at 13 billion dollars, placing the largest public repository of open model weights inside the company that supplies the hardware they run on. AdExchanger
  • September 4: Olivier Francois stepped down as chief executive of Fiat and chief marketing officer of Stellantis, with Arnaud Belloni, who spent sixteen years in Stellantis marketing leadership before moving to Renault, returning to assume part of the role. Adweek
  • September 4: Business Insider reported that the chief executive and president of Good Good are stepping down over the brand's Callaway collaboration, weeks after the golf partnership was cited as a case study in synthetic audience testing. AdExchanger