Measurement was the thread running through every significant filing of the past three days, and it ran in two directions at once. A browser began publishing how many advertisements each website shows and how much processing time those advertisements consume. A standards body moved its specifications out of static documents and into repositories that machines can read and version. A search platform documented a change to its bidding system twenty-nine days after that change had finished rolling out. Two healthcare data companies announced identity products built on hundreds of millions of patient records and published no accuracy figures for either of them. And a news agency described a paywall that decides, reader by reader, whether a subscription prompt is worth more than an advertising impression.
Set side by side, these are not five unrelated announcements. They describe an industry acquiring new instruments faster than it is agreeing on who calibrates them. In three of the five cases, the party supplying the measurement is also the party whose commercial position the measurement affects. That is not new. What is new is the speed at which the instruments are arriving, and the number of them that ship marked experimental, without thresholds, or without a methodology at all.
Chrome starts counting the advertisements on every page
On September 15, Google added four new measurements to the Chrome User Experience Report, the public dataset that already underpins Core Web Vitals. The metrics are ad count, meaning the average number of advertisements visible in the viewport; ad density, the average share of viewport area those advertisements occupy; ad weight measured as network usage in bytes; and ad weight measured as processor usage in milliseconds. Google announced them as tools for assessing ad-load experiences on the pages where advertising is bought and sold.
Luis Rijo
The mechanics matter more than the headline. Chrome does not collect new telemetry for this; it reapplies data it already gathers. The viewport is sampled once per second to derive count and density. Processor and network consumption accumulate across user sessions. Results are reported at the seventy-fifth percentile, the same convention Core Web Vitals uses, which means the figure describes a bad-but-not-worst-case visit rather than an average one. Advertising itself is identified through network-level filtering and script execution analysis, with resource requests matched against Google's ad filter list. The metrics are reachable through the CrUX API, the CrUX History API and the CrUX Vis visualisation tool.
There is one further restriction that shapes what the dataset can describe. Aggregate CrUX reporting of these metrics is limited to sites carrying an ads.txt file that names at least one authorised seller. That condition quietly defines the population: it captures sites monetised through programmatic supply chains and excludes sites that sell directly, sites behind subscriptions, and sites that carry no third-party advertising at all. A measure of ad experience across the web is, in practice, a measure of ad experience across the programmatically monetised web.
Barry Pollard, who works on web performance at Google, described the release plainly: "These metrics are a set of ad-related measurements that provide transparency into sites' ad experiences." Alex Cone, group product manager at Chrome, was more cautious about what happens next. "We want to better understand if [the metrics] are valuable for making decisions," he said, adding, "We look forward to feedback from the ecosystem." On the question of whether publishers should start changing their layouts, Cone acknowledged the sequencing problem directly: "We understand publishers may look to the buy side of the ad ecosystem to understand how it is using this data to inform ad buying decisions before taking action or optimizing."
He declined to say whether any demand-side or supply-side platform had committed to incorporating the signals. "We cannot speak for partners, but our early conversations with the ecosystem indicate an appetite for ad experience metrics," he said. Cone also confirmed that Display and Video 360 and Google Ads receive no preferential access to the data, which arrives simultaneously for independent platforms. Bill Reardon, general manager of enterprise platform at Google Ads, offered the buy-side endorsement: "Display and Video 360 supports this initiative to bring transparency to page layout quality. We look forward to exploring with the industry how these signals can improve quality and support advertiser goals."
Luis Rijo
Google's own framing was blunter. "We've all encountered overwhelming ad experiences on the web," the announcement said, adding that the company had "heard from a large cross-section of ad ecosystem stakeholders that these metrics could enhance how they value web ad opportunities." AdExchanger listed the supporters Google gathered, a group spanning The Guardian, Mediavine, Index Exchange and Adelaide, which is an unusual coalition of publisher, exchange and attention-measurement interests to align behind a browser metric.
Two features of the release deserve attention. The first is that Google has set no benchmarks and has stated it has no plans to set any. The metrics are classified as experimental and sit outside Core Web Vitals. That leaves the interpretive question open, and open questions in advertising do not stay open; they get answered by whichever buyer builds a filter first. A publisher cannot know whether an ad density of eighteen per cent is acceptable until a demand-side platform starts declining inventory above some number, at which point the number becomes the standard without anyone having proposed it.
The second is the classifier. Every figure in the dataset rests on a definition of what counts as an advertisement, and that definition is Google's filter list. It is a proprietary artefact, maintained by a company that is simultaneously the largest seller of advertising on the web, the operator of the browser doing the counting, and the owner of a demand-side platform that has publicly welcomed the signals. Nothing in the design suggests preferential treatment, and Cone's statement on equal access is explicit. The structural observation is narrower: a quality metric for publisher pages now has a denominator that publishers cannot audit.
Sitting underneath that is the awkwardness AdExchanger flagged. Chrome is offering a mechanism to improve the experience of advertising on publisher pages during a period in which generative search surfaces have reduced the traffic those pages monetise. A better-calibrated measure of ad crowding does not address a shrinking base of visits, and it arguably intensifies the pressure, since a site earning less per visitor has fewer ways to respond than by showing more advertisements.
The same week produced a quieter change to the industry's reference material. IAB Tech Lab moved its technical standards off static PDFs and onto GitHub, rewriting them as Markdown across seventy-seven public repositories. The organisation's own reasoning was that "the static nature of the PDF has become a bottleneck rather than a benefit." Mechanically, a correction to a PDF requires a full version increment and leaves no visible record of what changed between releases. The second reason is more telling: language models and autonomous agents that consult these documents during automated trading "need to confirm quickly that they are using the latest final specification," which tagged releases in a repository support and downloadable files do not.
The repositories carry OpenRTB, VAST, AdCOM and newer specifications including the Event and Conversion API finalised on May 3, 2026. The GDPR Transparency and Consent Framework repository is the most watched at 952 stars and 372 forks, with OpenRTB behind it at 519 stars and 216 forks, across an account with 443 followers. Formal versioning and working group review of pull requests continue, and final snapshots are still published for implementation. What changed is the audience the documents are written for. Between a browser API that publishes page quality as structured data and a specification library rebuilt so that agents can verify which version is authoritative, the advertising stack acquired two machine-readable reference layers in forty-eight hours.
An auction changed in August, and the documentation arrived in September
Google altered the behaviour of its target-based bidding strategies on August 17, and completed the global rollout on August 27. Advertisers learned about it on September 15, when an updated help document confirmed what had already happened. The change was first spotted by Hana Kobzová at PPC News Feed and reported by Barry Schwartz. Google's wording was procedural: "changes to target-based bid strategies rolled out globally across your campaigns that are 'Limited by budget.' This global rollout was completed on August 27, 2026."
The affected strategies are Target CPA and Target ROAS, across Search, Shopping, Performance Max and Demand Gen campaigns. The underlying mechanism is specific and consequential. Google had previously lowered bids on budget-limited campaigns to stop them overspending, which left many of those campaigns rank-limited by cost per click rather than by budget, and which had a side effect advertisers rarely complained about: constrained campaigns frequently beat their stated targets. The change removes that suppression, so performance aligns more consistently with the target an advertiser actually set.
Early analysis of what followed came from Mike Ryan at Smarter Ecommerce, who identified two effects moving in opposite directions. Budget-limited campaigns saw cost per click rise. And the inventory they had previously been bidding down on became available to everyone else. "A sizable pool of cheap clicks have been liquidated and are now available to non-limited campaigns," Ryan wrote. Kirk Williams reached a similar reading, observing that the update "improved the overall system for everyone (especially non limited-by-budget LBB competitors)."
That phrasing is worth pausing on, because improvement for the system and improvement for a given advertiser are not the same proposition here. What the change did was redistribute. Cheap impressions that had been reserved, in effect, for advertisers whose budgets could not chase them are now contested by advertisers whose budgets can. An account running at its daily cap subsidised an account running below it, and the subsidy has been withdrawn. Google's own guidance acknowledges the direction of travel without stating it: "It's recommended to review your settings and ensure they align with your business goals, especially if your 'Limited by budget' campaigns historically performed better than your targets."
The disclosure interval is the part that connects to everything else in this edition. For twenty-nine days, the mechanism governing how millions of campaigns bid had changed, and the only instrument available to detect it was performance data, read after the fact, by practitioners with enough account volume to distinguish a systemic shift from seasonality. The documentation that explained it arrived as an edit to a help centre page.
Benchmarks are the usual fallback when platform mechanics are opaque, and the week supplied a reminder of how thin that fallback can be. WordStream published Facebook advertising benchmarks on September 14 covering roughly 1,829 United States campaigns, reporting cost per lead holding at $27.39 while traffic clicks fell to $0.60. Traffic campaigns showed a click-through rate of 1.93 per cent, up 12.87 per cent year over year, against a cost per click down 14.29 per cent. Lead campaigns recorded a 2.70 per cent click-through rate, up 4.25 per cent, a cost per click of $1.80, down 6.25 per cent, and a conversion rate of 8.54 per cent, producing a cost per lead down 0.98 per cent.
The dispersion underneath those averages is the finding. Cost per lead ranged from $61.56 in dentistry to $12.30 in career services, a fivefold spread that makes the headline figure close to meaningless for any individual advertiser. The sample construction compounds the problem: 1,377 traffic campaigns spread across twenty-two industries and 452 lead campaigns across fourteen, with some lead subcategories resting on as few as four campaigns. The report also carries internal contradictions on conversion rate movement and states a lead sample window closing on June 30, 2028, a date that has not happened and is presumably a typographical error. A dataset offered as an industry reference should survive a reading of its own date fields.
Two healthcare identity products, no published accuracy
Both arrived on September 14, within hours of each other, and both describe the same ambition from different ends of the supply chain.
Chalice AI announced a partnership giving it PurpleLab's healthcare claims data, covering more than 330 million United States patient lives and upwards of three million healthcare professionals, built from longitudinal medical and pharmacy claims records spanning over fourteen billion claims a year. PurpleLab supplies the records; Chalice supplies the predictive modelling and the deployment infrastructure into connected television, social platforms and open programmatic inventory. The stated purpose is to identify audience members likely to be interested in specific drugs or therapeutic areas, uncover prescriber opportunities, predict adoption and attrition, and drive incremental growth.
Adam Heimlich, chief executive of Chalice AI, framed the pitch against the packaged-segment market: "Healthcare marketers don't need another off-the-shelf audience. They need the ability to start with the business or patient outcome they're trying to achieve and build an AI strategy specifically around that objective." Kyle Armstrong, vice president of agency and platform sales at PurpleLab, represented the data side. The announcement went out through Business Wire at 9:00 am Eastern from New York.
What the announcement does not contain is the entire technical substance of the claim. There is no launch date or availability timeline. There are no named pilot clients and no performance figures. There is no match rate between patient or provider records and the channels where the audiences are activated, which is the single number that determines whether a 330 million record base translates into addressable reach or into a fraction of it. There is no description of the privacy architecture, no statement of how claims data is handled relative to health privacy law, no disclosure of claims latency or model refresh cadence, and no clean room deployment specification.
Hours earlier, at Fierce Pharma Week 2026 in the Pennsylvania Convention Center in Philadelphia, Doceree launched its Intent Powered HCP Suite, extending what it calls Clinical Intent Signals into connected television, LinkedIn social advertising and open programmatic buying through an orchestration layer named Daily Command, available in the United States. The product ties those channels to six million verified physician identities held in an identity graph the company calls ESPYIAN, fed by 185 electronic health record and platform integrations, clinical content engagement across more than 2,000 medical publishers, and point-of-care interactions captured through a patented layer named Triggers.
The identity claim is stated in absolute terms. Doceree describes matching signals deterministically to a verified National Provider Identifier, with, in its words, no probabilistic lookalikes and no inferred audiences. Measurement includes signal-to-exposure latency reported alongside conventional reach and conversion metrics inside what the company terms a closed-loop attribution framework.
Harshit Jain, the founder and global chief executive, built the argument around decay. "The industry has accepted a strange bargain: it pays real-time prices for reach" while aiming with data "several weeks old," he said, adding that "clinical intent has a half-life" and citing therapy questions and diagnosis documentation as moments that fade within days. Kamya Elawadhi, co-founder and president, positioned the launch as closing gaps "for the channels where most of the budget sits," and said that "Care Sequence and DOOH proved that when channels take their cue from the clinic instead of the calendar, engagement and prescribing behaviour move together."
The market numbers cited are substantial enough to explain the activity. Prescription drug multiscreen television spending reached $4.95 billion between October 2025 and March 2026, a rise of 53 per cent year over year, against projected United States healthcare digital advertising spending of $20 billion for 2026. That is a category expanding fast enough to reward whoever can make its targeting defensible.
Doceree's disclosure gaps mirror Chalice's almost exactly. No clients, no pilot data, no benchmark figures. The source of the identity graph is undisclosed. The method by which LinkedIn identities are verified against National Provider Identifiers is unspecified, which is a material omission given that LinkedIn profiles are self-declared. There is no described separation between physician data flows and patient data flows. The latency figure the company proposes as a headline measurement has no published definition, and the verification methodology behind the deterministic claim is not detailed.
The juxtaposition with the week's other filing is stark. Chrome published a page-quality metric along with its sampling interval, its percentile convention, its classification method and its population restriction, and then declined to say what a good score is. Two healthcare data companies published targeting products addressing several hundred million patients and three million clinicians, asserted determinism, and published no methodology at all. Both sets of numbers will be used to price media.
Apple ships Siri AI, and Europe gets a partial version
Apple released iOS 27, iPadOS 27, macOS 27, watchOS 27, visionOS 27 and tvOS 27 on September 14, with an entirely rebuilt Siri as the headline. The geography of that rollout is the story. Siri AI shipped in English beta globally with the exception of China, and in the European Union it reached Macs and Apple Vision Pro but not iPhones, iPads or the watch. China is excluded entirely with no date offered. French, Japanese, Korean, Portuguese and Spanish are scheduled for October 2026. The assistant is unavailable to users under thirteen.
Apple attributes the European exclusion to its reading of the Digital Markets Act, arguing that a virtual assistant of this kind would require direct access to private data and control over other installed applications. The line it has drawn follows the regulation's designations rather than the hardware: macOS is not a designated core platform service, while iOS, iPadOS and Safari are. The result is a single Apple account whose assistant capability depends on which device is in hand, and an EU-based advertiser whose audience now contains systematically different assistant behaviour from the same audience elsewhere.
That is one posture toward the same regulation. Another was visible in search the same week, as Google's ecosystem carousel went live across the European Economic Area, surfacing related websites alongside weather, sports, finance and translation results on desktop and mobile, first documented publicly by Lluc B. Penycate. One company is adding surfaces in Europe to satisfy the regulation; the other is withholding a feature in Europe for the same stated reason.
Three elements of the Apple release carry direct commercial weight. Safari Intelligence now performs price-drop and restock monitoring inside the browser, which removes the need for a shopper to opt into a retailer's own alerting system. Retail alerting has been a first-party data capture mechanism as much as a service, since it requires an email address or a push permission; moving it into the browser disintermediates that exchange without replacing it.
Screen Time has been rebuilt around application categories rather than devices, letting a parent set daily minute caps for Entertainment, Games and Social Media separately instead of capping the device as a whole. Applications that declare social feeds receive a 13+ rating floor, and those declarations were made through App Store Connect in July 2026. Attention is now metered at category level against a taxonomy that publishers filled in themselves two months before the metering began. New child account setup adds an Ask to Browse flow requiring parental approval before a child reaches an unfamiliar website.
The third element concerns crawling. Applebot, which supplies the web context behind Siri AI retrieval, added 4,656 IP addresses in a single update during August 2026. Apple rewrote its crawler documentation on June 8, 2026, formalising both AI training use and real-time retrieval, and adding nosnippet controls and paywall handling. An assistant that answers from the live web needs a crawl fleet sized to match, and the address expansion is the visible proxy for that.
Apple also disclosed, without pricing, date or packaging, that expanded access to these features will carry a fee in future. And the performance claims accompanying the release are a small case study in measurement asymmetry: application launches up to 30 per cent faster, photo loading up to 70 per cent faster, AirDrop transfers up to 80 per cent faster, iPad file transfers up to five times faster, all measured by Apple on prerelease software. The device used for the application launch comparison was an iPhone 11 Pro Max, which cannot run Apple Intelligence at all.
Reuters prices each reader against the advertisement they would have seen
The clearest description of how a publisher actually values a visit came from Reuters, whose subscription business reaches its second anniversary in October. Earlier in 2026 the agency deployed a dynamic paywall, and Phil Andraos, general manager for digital, explained what it does and what it has done to advertising revenue.
The system runs machine-learning models over three inputs: the behaviour of the individual user, the performance of the specific article measured by whether it has historically converted readers into subscribers, and the estimated advertising revenue that user represents. It then decides whether to show that person a paywall. The calculation is explicitly a comparison between two revenue paths for the same visit. The gate tightens during active news cycles, when both content performance and conversion likelihood rise, and loosens on slow days when conversion rates fall. Andraos described the launch effect in four words that will be familiar to anyone who has watched a yield system switch on: "It's basically like magic. Overnight, the dollars start to flow."
Pricing sits at a flat $4 a month across 80 countries, with no introductory offers and no promotional rates, an unusually rigid structure in a category built on discounting. Paul Bascobert, the Reuters president, said in June 2025 that the company had over 100,000 subscribers; Andraos declined to update the figure, describing growth as "very news cycle dependent" and comparing its shape to "the stock market" rather than a line.
The advertising consequence is the counterintuitive part. Reuters subscribers do not get an advertisement-free experience, so conversion does not remove inventory. "Our advertising business is doing better today than it was before launching the paywall," Andraos said, arguing that engaged audiences consume more pages and generate more impressions, and that authenticated subscribers supply first-party data that improves the conversation with advertisers. His summary of the trade: "A smaller pool of highly engaged users can drive so much more ad revenue than a high base of loosely engaged readers." Similarweb estimates Reuters site traffic fell 19 per cent year over year in August 2026, which is the context that argument has to survive.
Reuters also moved to block bots by default in May 2026, while continuing to allow Google's crawlers. Bot traffic fell; monetisable traffic, by Andraos's account, did not. The evaluation criterion is revenue reciprocity, whether a given bot generates audience, advertising or subscription revenue, with those that profit from the content without returning anything facing a block. He left the position provisional: "If at some point it flips, where we feel the negatives outweigh the positives, then we'll block certain bots."
Consumer-facing subscription pricing moved in a rather less calibrated direction. Digiday's tally of every major streaming price increase since 2019 puts Netflix's standard ad-free tier at $19.99 against $10.99 in 2019, a rise of 82 per cent. Disney+ ad-free has gone from $6.99 at launch in 2019 to $18.99, up 172 per cent. Peacock Premium, which carries advertising, rose from $4.99 at its 2020 launch to $12.99 in September 2026, an increase of 160 per cent. Netflix and Disney+ have each raised standard ad-free prices five times since 2019; Apple TV+ and Peacock have done so four times. Taking the ad-free tiers of eight major streamers without bundle discounts produces a combined $151 a month.
The Peacock figure is the one that complicates the usual reading. An advertising-supported tier rising 160 per cent is not a discount holding steady while the premium tier climbs; it is the ad tier carrying its own escalation. The gap between tiers is what converts subscribers into advertising inventory, and that gap has to keep widening for the mechanism to keep working.
Which returns to the instrument problem. Reuters can price a paywall per reader because it can estimate what that reader is worth as an impression. Streaming services raise prices on a schedule and discover the ceiling by hitting it. Chrome will now report how crowded a publisher's page is without saying how crowded is too crowded. Google changed how millions of campaigns bid and documented it four weeks later. Two healthcare companies announced identity products and published no accuracy figures. The volume of measurement in the market is rising sharply. The number of parties who can check it is not.
Also noted
- September 15: Google is testing text-based sponsored links inside AI Mode responses, presented as ordinary anchor text with a Sponsored label above the generated answer on both desktop and mobile, spotted by Sachin Patel (Search Engine Roundtable).
- September 15: a settlement analysis by SuperAwesome chief executive Kate O'Loughlin notes that roughly $5.3 billion of Meta's $17-18 billion teen safety settlement is conditional on TikTok and Google adopting comparable protections, and that teen accounts now default to non-algorithmic feeds with autoplay disabled (AdExchanger).
- September 14: Amazon will require most single-ASIN bundles to be packaged by the original manufacturer or brand from January 11, 2027, with brand authorisation letters uploaded through Account Health and no pre-clearance route, leaving sellers to submit letters only after violations are recorded (PPC Land).
- September 14: IAB Ireland and RED C found that 72 per cent of Irish adults value authenticity in creator content while only 17 per cent perceive it in sponsored posts, a 55-point gap, alongside four in five wanting disclosure when AI has substantially altered content (PPC Land).
- September 15: Google acknowledged a backlog in Business Profile API access applications, with Victoria Kroll stating the company cannot provide a response timeframe and asking applicants not to submit duplicates (Search Engine Roundtable).
Discussion