Three of the largest advertising businesses in the world reported quarterly results inside seventy-two hours this week, and all three grew. Meta Platforms posted $59.36 billion in second-quarter advertising revenue. Amazon posted $19.8 billion. Microsoft posted the slowest search advertising growth it has recorded in five quarters. Underneath those numbers, a quieter set of disclosures accumulated across six trade publications, and they pointed the other way.

A joint analysis from the Trustworthy Accountability Group and the Association of National Advertisers found that automated junk inventory scores better than human-made inventory on nearly every quality signal media buyers currently run. A tracking study of 100,000 shopping queries found that Google's AI Mode surfaces roughly 95% fewer product listings than standard search. IAB Australia published a decision matrix that abandons the idea of treating automated web traffic as a single category. A consumer survey commissioned by DoubleVerify found that people react less to whether artificial intelligence made an advertisement than to whether the result looks competent.

The through-line is not that the industry is shrinking. Revenue is at records. The through-line is that several of the instruments the industry uses to describe itself stopped describing the thing they were built to measure, and this week produced unusually direct evidence of the gap.

What follows covers the five stories that carried the most weight between Monday, July 27 and Friday, July 31, 2026, followed by a short list of smaller items.

Record quarters, thinner margins, and an unpriced input

The week's earnings sequence started with Meta. The company reported second-quarter advertising revenue of $59.36 billion, a 27% increase over the $46.56 billion recorded a year earlier, in a press release dated July 29, 2026 covering the three months ended June 30, 2026. Total company revenue reached $60.80 billion, up 28%. Net income fell 8% to $15.8 billion. Diluted earnings per share came in at $6.18, down 13% from $7.14.

The gap between rising revenue and falling profit resolves into two line items. Meta absorbed a $2.4 billion charge tied to legal proceedings and $1.18 billion in severance costs connected to a headcount reduction the company initiated in May 2026. Total costs and expenses rose 55% to $42.03 billion. Operating margin, which had held between 40% and 48% for eight consecutive quarters, dropped to 31%. The company raised the lower end of its full-year expense outlook to a range of $165 billion to $169 billion to absorb the legal charge, and it stated in the earnings release that youth-related trials scheduled in the United States this year may still produce a material loss.

The mechanics beneath the advertising line matter more to buyers than the profit figure. Ad impressions delivered across the Family of Apps rose 14% year over year, while average price per ad increased 12%. Those two components behaved very differently by geography. Asia-Pacific impressions grew 17% against 9% in the United States and Canada and 13%in Europe, while average price per ad climbed 20% in the United States and Canada and just 1% in Asia-Pacific. Regional advertising revenue split into $26.34 billion for the United States and Canada, $14.09 billion for Asia-Pacific, $10.85 billion for Europe, and $8.09 billion for the rest of the world. Family daily active people averaged 3.60 billion for June 2026, up 3%, a growth rate essentially flat since the fourth quarter of 2025. Average revenue per person reached $16.86.

Meta introduced a delivery-system change during the quarter that it named the Meta Generative Recommender, replacing individual scoring of every candidate advertisement with large language models reasoning about advertisement content and user preference together. Chief financial officer Susan Li disclosed component-level figures: a first generative model deployed into the advertisement retrieval system, early pilots using language models to interpret user preference that drove a 1% increase in app event conversions on Instagram, and improvements to user understanding models combined with the GEM ranking system that produced an 8.3% increase in advertisement clicks and a 15.7% uplift in conversions on Facebook. Those figures have a comparison point. The GEM model doubled its training cluster in the fourth quarter of 2025 and delivered a 3.5% click lift on Facebook at that stage.

Automation adoption widened alongside the ranking changes. The Advantage+ campaign suite, which bundles audience, placement and budget optimisation into a single managed workflow, reached an annual revenue run-rate above $75 billion, up from the $60 billion run-rate disclosed for the third quarter of 2025. Li illustrated the performance gap with an Indian apparel brand that moved from manual campaign setup across Facebook and Instagram to Advantage+ sales campaigns layered with audience, placement and budget optimisation, recording a 13% incremental lift in purchases and a 16% increase in add-to-cart conversions. More than 9 million small businesses now use at least one generative creative tool, up from 8 million in the first quarter of 2026, a figure that had itself doubled from 4 million roughly four months earlier. Adoption of the image generation tool, which now builds images from video assets, more than doubled inside the quarter alone.

Business Agent moved from a global consumer rollout to an enterprise platform during the same three months. More than 1 million businesses now use the system weekly to handle customer conversations or complete sales, a figure disclosed for the first time this quarter following the tool's global availability in June 2026. Li cited a deployment at Movida, a Brazilian car rental company operating close to 400 locations, which put an agent on WhatsApp to run the entire booking flow from vehicle selection through payment. Returning customers completed reservations in as few as three messages. Across one month, Movida reported a 44% increase in daily bookings through the channel against the same period a year earlier, with 85% of conversations resolved without human involvement.

Engagement milestones bear directly on the impression side of the equation. Instagram passed 2 billion daily active users. Threads crossed 500 million monthly active users, extending from the 400 million mark disclosed in January 2026 when Meta completed the global rollout of advertising on the platform. Quarterly Family of Apps other revenue, the line covering subscriptions and WhatsApp paid messaging, reached $1 billion for the first time and grew 73%, a non-advertising stream now registering at scale alongside the core business.

The capital picture carries its own tension. Capital expenditures including finance lease principal payments hit $31.08 billion for the quarter against $17.01 billion a year earlier. Free cash flow fell to $784 million from $8.55 billion. Full-year capital expenditure guidance narrowed to $130 billion to $145 billion from a prior $125 billion to $145 billion. Long-term debt issuance rose $24.91 billion during the period, leaving cash and marketable securities at $90.26 billionagainst long-term debt of $83.66 billion. Mark Zuckerberg disclosed a financing arrangement with BlackRock to develop a 1GW data centre in El Paso, Texas, announced July 28.

Analysts pressed Zuckerberg on an apparent contradiction: Meta is fielding offers to sell its own compute capacity at a premium while simultaneously buying capacity from other providers. His answer acknowledged the strain without resolving it, describing demand as far outstripping available compute and arguing that selling capacity for near-term profit would forfeit the compounding value of building on top of it. The comment connects straight back to the ranking systems described above, since the same infrastructure underwrites them. On 2027 capital expenditure, Li declined to give figures, characterising infrastructure planning as highly dynamic with a range of outcomes embedded even in the current year's outlook.

Headcount stood at 75,472 as of June 30, down 1% year over year, including roughly 8,000 employees affected by the May reduction who will drop off the count by the end of the third quarter. The effective tax rate for the quarter was 16%, with the remaining-quarters outlook rising to a range of 15% to 17% from a prior 13% to 16%. Third-quarter revenue guidance sits at $61 billion to $64 billion, assuming roughly a 1% currency headwind.

Amazon reported the following day. Advertising services revenue reached $19.8 billion, up 26%, carrying the trailing twelve-month total to $76 billion. Total net sales came in at $200.6 billion, up 20%, the company's first quarter above $200 billion. AWS contributed $42.2 billion, up 37%. Operating income rose 43% to $27.5 billion. Net income of $62.6 billion included $53.4 billion in non-operating pre-tax other income drawn principally from the company's investment in Anthropic, a detail that separates the headline profit figure from operating performance. Free cash flow ran to an outflow of $7.6 billion, driven by property and equipment purchases tied to artificial intelligence build-out.

Chief executive Andy Jassy named Sponsored Products as the largest offering and the primary growth driver, but the disclosures that drew attention concerned conversational placements. Shoppers who click a Sponsored Prompt convert 48% more often and spend 21% more on average than those who do not. Advertisers running the Ads Agent tool recorded 8% lower cost per impression and 6% lower cost per acquisition, and Amazon has extended the tool to 11 new countries this year. The company introduced more than 30 new advertisers to the NBA in its first year of streaming rights, and inventory across Thursday Night Football, NBA, WNBA and NASCAR sold out. AdExchanger noted the shift in tone across five years of Jassy earnings calls, observing that the largest advertising companies now tend to present themselves as though the advertising business were incidental.

Microsoft completed the trio with the week's outlier, reporting fiscal fourth-quarter results on July 29, 2026 for the three months ended June 30. Search advertising revenue excluding traffic acquisition costs grew 10%, two points below the previous quarter. Barry Schwartz set the figure against the preceding sequence of 12%, 10%, 16% and 21%, which makes this the second deceleration in three quarters and the weakest reading in the run.

PPC Land's detailed coverage of the same results put constant currency growth at 9% and traced the variance to third-party publisher and content-partner arrangements. Chief financial officer Amy Hood attributed the gain to higher revenue per search and volume across Edge and Bing, with those partnerships acting as a drag. The company's fiscal 2026 annual report attributes full-year search advertising growth of 9%, or 12% excluding traffic acquisition costs, partly to a benefit from the same partnerships across the twelve-month period, which is an unusually direct illustration of how sensitive a reported growth rate can be to the timing of individual publisher deals. Full-year search advertising revenue reached $15.176 billion, up from $13.878 billion in fiscal 2025 and $12.306 billion in fiscal 2024. Guidance for the first quarter of fiscal 2027 points to mid-single-digit growth.

The contrast with the rest of the company is stark. Total revenue reached $90.0 billion, up 18%. Net income rose 31% to $35.8 billion, helped by a $3.2 billion gain on Microsoft's investment in Anthropic and reduced by $480 million from the accounting treatment of its OpenAI stake. Azure and other cloud services grew 43% for the quarter and passed $100 billion across the full fiscal year for the first time. Capital expenditures including finance leases rose 70% to $41.0 billion, roughly two-thirds of it directed at short-lived assets such as processors and graphics units. Commercial remaining performance obligation reached $678 billion, up 84%. Search advertising, in other words, is now a smaller and considerably more variable contributor to a company whose financial narrative has become a cloud infrastructure story.

Three growth rates, three different stories: 27%, 26%, 10%. The comparison base matters. Alphabet had reported Search advertising revenue of $63.3 billion, up 17%, on July 23, 2026, with network revenue down 1% and capital expenditure guidance climbing to $205 billion.

The cost side of that build-out surfaced from an unexpected direction. Omnicom chief executive John Wren told analysts on the holding company's second-quarter call, reported by Digiday on July 29, that "the marketplace hasn't seen what the cost of this AI is." The remark came in response to a question about where artificial intelligence is generating savings and whether those savings return to client marketing budgets. Wren answered about cost rather than savings, on the record, on a call intended for investors, from the head of the largest advertising holding company in the world.

AdExchanger had described the commercial mechanism a day earlier. Its July 28 roundup reported that holding companies are settling the question through principal media: several are willing to absorb the full artificial intelligence infrastructure bill provided the client commits a predetermined volume of paid media to inventory the holding company bought in bulk and resells at a markup. The publication noted that the ethics of the arrangement remain contested. By July 30, AdExchanger reported Omnicom halfway through a sell-off of IPG agencies, eliminating specialist firms and overlapping agencies in certain countries.

So the infrastructure cost that Meta booked as $31.08 billion in quarterly capital expenditure, and that Alphabet guided to $205 billion for the year, has not yet been assigned a price anywhere downstream of the platforms. That is the unresolved commercial question the week produced, and no participant offered an answer to it.

The verification stack rates synthetic inventory above human work

The single most consequential technical finding of the week arrived on July 28, when the Trustworthy Accountability Group, the Association of National Advertisers and the technology firm Fiducia published the first statistically rigorous sizing of artificial-intelligence-generated junk inventory in programmatic advertising. PPC Land covered the analysis in detail. The size estimate is not the finding that matters.

The estimate itself is modest. AI slop accounts for between 1.3% and 2.4% of open web programmatic investment, close to the 1.1% measured for made-for-advertising inventory in the same quarter. Two methods produced the two bounds across the same underlying dataset. Domain-level classification, supplied by DeepSee, covered $33.97 million in matched open web spend and produced the upper figure of 2.4%. Rendered page-level evaluation, supplied by Mobian, covered 233 million URLs, 4.45 billion impressions and $14.84 million in measured spend, with more than 30,000 page-level evaluations and human verification of flags, producing the lower figure of 1.3% at a 99% confidence interval of 0.98% to 1.56%. The work sits inside the Q1 2026 ANA Programmatic Transparency Benchmark and was conducted and drafted by Scott Cunningham of Cunningham.tech Consulting.

The direction of the quality signals is the finding. Slop inventory recorded an invalid traffic rate of 0.05% against 0.32%for clean supply. Viewability came in at 77.2% against 74.9%. Once measurability was accounted for, slop graded as premium more than 70% of the time. Those scores translated into price: a TrueCPM of $7.08 for slop against $6.15 for clean inventory. Advertisers paid a premium for impressions rated highly by systems that were never designed to assess whether a human wrote anything.

Definitions came before measurement, and the definitional work is what keeps the finding from collapsing into a complaint about machine-written text generally. Through engagement with supply chain quality vendors, the analysis settled on a reading of AI slop as low-value, mass-produced content generated primarily by machine for monetisation, with little or no human input, originality or audience value. Vendors interviewed described the material as carrying zero originality and semantic shallowness, and as content unable to demonstrate what a human author contributed. Several categories built with machine assistance were explicitly excluded: generated data summaries such as box scores or earnings recaps, machine-supported editorial where a person adds edits and original perspective, high-quality machine-native products, transparent content aggregators, and design software. No vendor consulted defined the problem as simply being machine-generated content. The defining characteristic was quality, not provenance.

The mechanism is not obscure. Automated content sites are constructed to render quickly, serve measurable advertisement slots, and avoid the bot traffic that triggers fraud filters, because their revenue depends on impressions clearing verification. A page assembled by a language model and wrapped in a template performs well against tests built to catch fraud and hidden placements, and it performs badly against no test at all, because none of the standard signals asks about authorship.

One structural marker separates the two populations. The analysis found a 30.0% templated-site rate on slop inventory against 1.2% on clean inventory, a factor of 25. High viewability and low invalid traffic on template-driven domains function as a signature rather than an endorsement. Across 11,552 domains classified as slop, the topic mix followed a familiar content-farm pattern: parenting, travel, recipes, hairstyles, personal finance and how-to material, with fabricated viral stories and engagement bait produced across near-identical templates and clone networks.

Exposure varies sharply between advertisers drawing on the same market, ranging from 0.11% to 13.84% of spend. Roughly 1 in 27 impressions on unknown domains, or 3.7%, was classified as slop. Large established exchanges ran single-digit rates, smaller and native-format exchanges reached 4% to 8%, and known publishers showed effectively zero. Of slop inventory, 88% was also identified as made-for-advertising. The remaining 12% is the operational gap: inventory that escapes made-for-advertising suppression because it lacks the advertisement density and arbitrage traffic patterns those classifiers look for, and that prices above clean supply precisely because it passes every other check.

Social platforms are identified in the analysis as the primary and fastest-growing environment for the material. Platforms are putting policies and identification procedures in place, but one vendor estimated that 25% to 40% of social video inventory is misaligned, with slop forming a large and growing share of that. PPC Land has tracked the trajectory since 2025, having documented platform monetisation programmes driving mass-produced low-quality content across social media in June 2025 and one third of the YouTube Shorts feed described as machine-generated by December 2025.

The industry has been circling the problem for two years without a measured baseline. Integral Ad Science flagged slop sites as a threat to programmatic effectiveness in July 2025 and became the first company to receive the Alliance for Audited Media Ethical AI Certification the same month. DoubleVerify exposed AutoBait, a network of more than 200made-for-advertising domains producing clickbait at roughly $2.25 per article, in March 2026. IAS opened a beta for low-quality content avoidance in April 2026 and moved it to general availability in May 2026, reporting a 49% higher success rate on non-slop inventory, with coverage limited at that point to English-language text on the open web. The 12% of slop sitting outside made-for-advertising frameworks maps onto exactly that coverage limitation.

The wider benchmark shows movement elsewhere. TrueAdSpend, the share of total spend reaching impressions that are fraud-free, measurable, viewable and non-MFA, rebounded to 43.3% in Q1 2026 from 36.3% in Q4 2025. Non-viewable impressions dropped from 22.7% to 13.3%. Transaction costs moved the other way, rising to 25.9% from 23.7%. Total CPM declined to $4.42 from $5.55, and TrueCPM fell to $6.47 from $12.77. Divergence between advertisers widened: the higher-performing cohort recorded TrueAdSpend of 54.0% against 32.1% for the lower half, and the lower cohort paid $19.04 per quality impression against $7.46, a factor of 2.6. The sample covered 86 participating marketers, 20.9 billion impressions and $160 million in advertising spend, with CTV at 47.5% of spend and web at 42.3%.

A second data point landed the following day from a different direction. MediaPost reported findings from a DoubleVerify-commissioned study on July 29, titled Global Insights: Media Quality in the Age of AI, covering 22,000consumers across 22 markets and 2,020 marketers and advertisers across 21 markets. Globally, 63% of consumers said artificial-intelligence-powered tools such as search summaries, chatbots, creative automation and recommendations improve their online experience. In North America that figure drops to 50%, the most cautious region measured. Regional spread runs wide: 69% in Latin America, 68% in Asia-Pacific, 53% in Europe, the Middle East and Africa.

The advertising-specific numbers are where the study intersects with the TAG analysis. Low-quality or uncanny generated advertising negatively affects brand opinion for 42% of consumers, while 40% view polished, professional generated advertisements positively. Only about a third of North American consumers reported positive feeling toward a brand using generated advertisements that look competent. And 56% said they cannot consistently identify generated content at all.

Read against each other, the two studies describe the same failure at different points in the chain. Machine verification cannot detect the absence of human authorship and prices its absence upward. Human audiences cannot reliably detect generation either, but they do detect incompetence, and they penalise the brand for it rather than the publisher. Neither the buy-side tooling nor the audience is measuring what the industry assumed each was measuring.

Regulation arrived at the same problem from a third direction during the week, and on a fixed date. Article 50 of the EU AI Act binds on August 2, 2026, carrying penalties of up to 3% of global turnover for providers that fail to mark synthetic output. PPC Land reported on July 28 that Meta signed the EU code of practice on AI-generated content with five days remaining, a signature confirmed by Markus Reinisch six days after the compliance-presumption deadline and reversing the company's 2025 refusal. Google had signed two days past the same presumption deadline, under terms requiring providers to watermark free-form text above 200 tokens and to offer detection at no charge. The tooling arrived alongside the signatures: PPC Land documented the DV360 API gaining an AI attestation field ten days before the deadline, with advertiser-level defaults auto-filling Demand Gen and YouTube advertisements.

The penalty exposure is not limited to Europe. PPC Land set out the parallel New York regime, where synthetic performer advertisements carry $1,000 first-violation penalties already in force since June 9 and rising to $5,000, with exemptions still shielding parts of audio and streaming. Against that backdrop, a survey PPC Land covered on July 26 found generative tools lifting creative volume for 88% of marketers but quality for only 45%, with 67% still briefing models on demographic data the same respondents describe as no longer working. Volume up, quality flat, disclosure mandatory, detection unreliable: that is the compliance position the sector entered August in.

Two smaller items from the same week fill in the picture. PPC Land reported on July 27 that just 5.3% of 131 marketers surveyed by Lunio run a dedicated invalid traffic platform, with 53.5% reporting they know little about such tools and 75.6% reporting budget lost to bots. Half said bidding now optimises toward bots rather than buyers. And on July 28, PPC Land documented RedCircle turning classification models on programmatic creative after cutting read errors 75%, in response to sex toy creative arriving tagged as Shopping and political advertisements bypassing category blocks. The checks run in seconds, and they run on one host platform.

Product visibility collapses inside the answer

On July 29, the product-visibility tracking platform Productrise published a study that PPC Land covered the same day. Over 21 days in July 2026, the company compared standard Google search against AI Mode on more than 100,000 identical shopping searches, tracking over two million product listings side by side on the same day for each query, in the United States and United Kingdom only, the two markets where Google's Shopping Graph feeds AI Mode responses directly.

Three findings stack. Standard search returned at least one product on about 88% of shopping queries; AI Mode did so on about 23%. When products appeared on both surfaces, standard search displayed an average of 22.5 listings per results page against 4.3 in AI Mode, roughly 5.3 times fewer competitive slots. Multiplying the reduced frequency by the reduced count produces the headline: approximately 95% fewer product listings surfaced in AI Mode across the same query set.

The methodology matters because the headline number invites objections. Productrise ran clean, non-personalised sessions and used keyword queries entered directly rather than the longer conversational prompts sometimes associated with machine-first search behaviour, on the reasoning that AI Mode opens from the standard search box and most people therefore arrive with ordinary shopping keywords. A product was defined as one distinct listing: a single card inside a standard Shopping carousel, or a single item referenced inside an AI Mode answer. Averages excluded empty results. Titles were normalised across surfaces, since AI Mode displays cleaned names while standard search shows raw feed titles supplied by merchants. Two worked examples illustrate the gap. For the query "neck cream," standard search filled the page with a shoppable carousel while AI Mode returned a block of text linking to third-party reviews with no shoppable listing at all. For "1/2 inch plywood," standard search returned a carousel of ten products against four listed in AI Mode.

The secondary finding is arguably the sharper one. Only about 0.8% of products visible in standard search also appeared in AI Mode for the same query on the same day. Both surfaces draw on the same Shopping Graph, which Google has described as containing more than 60 billion listings with over 2 billion updated hourly. A single data source producing almost entirely different result sets points at selection logic rather than feed availability. Retailers ranking well in conventional Shopping carousels have no structural guarantee of appearing at all when the identical query runs through the conversational surface.

Productrise founder Hugo Huijer offered a mechanical explanation tied to the interface: a conversational surface that expects a follow-up question does not need to display twenty products at once, and click-through rates for the products that do clear the bar should run considerably higher. Callum Lockwood, SEO director at Re:Signal, argued that fewer products per response does not necessarily mean fewer opportunities across a full shopping journey. Jason Berkowitz of Break The Web raised the question of whether current product scarcity is a lasting property of the surface or a phase tied to monetisation plans, noting that Google is already testing sponsored product experiences there. Jamie D'Alessandro of Jakala framed the narrower result set as consistent with Google's existing product philosophy rather than a departure from it.

The findings land roughly fourteen months after AI Mode's United States launch in May 2025, and the surface has expanded from experiment to infrastructure in that window. AI Mode passed 1 billion monthly active users globally by May 2026, with query volumes more than doubling every quarter since launch and average query length running about three times longer than in traditional search. Commercial infrastructure tracked the user growth closely. Google introduced a shopping advertisement format built for the surface on February 11, 2026, when it counted more than 75 million daily active users. Sponsored store listings appeared inside product panels by March 2026. By late May 2026, sponsored listings had become visually difficult to separate from organic recommendations, distinguished only by a small label. Structured data followed: dozens of new Merchant Center attributes for conversational interfaces on January 11, 2026, several carrying an explicit note that they are intended primarily for conversational experiences, and conversational attributes established as a distinct data type at Google Marketing Live on May 20, 2026.

The click-through research sitting behind all of this has moved in one direction. Ahrefs found AI Overviews reducing organic clicks to top-ranking sites by 34.5% across 300,000 keywords in April 2025, then measured the reduction at 58%for top-ranking pages in February 2026. A randomised field experiment involving 1,065 desktop Chrome users, the first causal study on the question, found outbound organic clicks down 39.8% and zero-click searches up 34.5% where the feature appeared, with no measurable improvement in how users rated the experience. Google has consistently contested that body of work, arguing its features send substantial traffic and that clicks originating from generated summaries show higher engagement quality. The company had not publicly addressed the Productrise Shopping figures at publication.

Context arrived from Search Engine Roundtable the following morning. Barry Schwartz reported on July 30 that AI Overviews now appear on roughly 43% of queries by Similarweb data and about 48% by Semrush data, with informational categories running higher still. In the same recap, SER documented Google testing AI-generated descriptions on shopping and product advertisements, an extension of a test previously observed on standard sponsored results and confirmed by Google as small at the time. The generated layer is therefore moving into the paid unit itself, not only the surrounding answer.

SER also covered Google's response to complaints about a race-based AI Overview answer on July 30, in which the company attributed the output to the websites it drew from rather than to the generated summary. Schwartz flagged the timing against Google's separate pursuit of copyright protection for its search results, including AI Overview responses, an awkward pairing of positions on authorship. A local version of the same problem had appeared days earlier, when PPC Land reported that a Google AI Overview invented a July 31 closing date for Anna Mae's Bakery, which owner Amanda Herrfort learned about from a screenshot sent by a wedding customer. Google attributed that error to thin web data.

The publisher-side arithmetic ran through the same week. On July 27, PPC Land documented All About Berlin losing 75% of its traffic as zero-click searches reached 68%, with affiliate income remaining down 30% despite renegotiated commissions, and noted Time routing artificial intelligence crawlers to a stripped-down copy of its site. That sits on top of prior-week reporting that publisher advertisement supply fell 40% in the second quarter while Google Search advertising grew 17%, and that Reddit, USA Today, Politico and Reuters are weighing crawler blocks as search traffic drops 28%.

Not every measurement points the same way. PPC Land reported on July 28 that Chemist Warehouse gained 220% more AI Overview appearances in a quarter, with News Australia driving a 46% rise in retailer brand mentions across ChatGPT and Google AI Mode over five months. No framework yet links those citations to sales, which is the recurring problem: visibility inside generated answers is now measurable, and its commercial value is not.

A commercial dispute over the same territory surfaced at AdExchanger on July 29. The eyewear company Blacksheeplost its top organic position for its own brand name after a television segment drove a traffic surge, with Google displaying a lookalike domain leading to an error page in its place. The company responded with 25 LED billboard trucks outside Google's New York offices. The complaint covered both the broken link and the volume of sponsored results above it. That same roundup carried Luma data showing advertising technology M&A activity up 5% in the second quarter, with Fox acquiring Roku for $22 billion as the headline transaction, Walmart buying Vibe.co, and Viantacquiring TVision.

Crawlers get a taxonomy and a September deadline

On July 30, IAB Australia distributed a Bots and Crawler Guidance and Decision Matrix through its weekly member newsletter, having released it two days earlier at its Discovery: AI and Search Summit in Sydney. PPC Land covered the document in detail. The guidance is dated version 1.0, prepared July 22, 2026, and credited to Jonas Jaanimagi, the organisation's technology lead.

The framework's premise is that automated visitors should be sorted by what they do with content rather than treated as a single problem. It splits non-human traffic into five job categories and offers four verdicts for each: allow, allow with conditions, require licensing, or block.

Category A covers discovery and search indexers including Googlebot, bingbot and OAI-SearchBot, described in the document as the librarian that catalogues a site and sends readers back. Category B covers training crawlers including GPTBot, ClaudeBot and CCBot, framed as the student who memorises a text and never returns it. Category C covers live agents including ChatGPT-User, Claude-User and Google-Agent, fetching pages in real time because a person asked something. Category D covers operational and advertising infrastructure including AdsBot-Google and verification fetchers. Category E covers unknown or unverified traffic, including spoofed user agents and default library strings such as python-requests.

Recommendations diverge by category and by whether the site owner is a publisher or a brand. Search crawlers get an unconditional allow across both. Training crawlers get block-or-license for publishers and a selective stance for brands, letting product facts through while withholding premium intellectual property. Live agents get a conditional recommendation for publishers, restricted on advertisement-funded pages, and an allow for brands on product and knowledge pages. Operational infrastructure gets an unconditional allow once identity is verified. Unknown traffic gets challenge-or-block.

The traffic figures behind the framework are the reason for its timing. Cloudflare data cited in the guidance puts automated requests at 57.5% of web-page requests as of June 2026, the first crossover past human traffic on record. Roughly 52% of artificial intelligence crawler requests go toward model training, against about 2.6% representing real-time, human-triggered fetches. DataDome figures show agent requests climbing 45% quarter on quarter to 17.7 billion in the second quarter of 2026, and 80% to 88% of referral traffic from artificial intelligence sources now originating from ChatGPT even as ChatGPT's own crawling volume declined. Model Context Protocol traffic went from negligible to peaks near 500,000 requests a day. Mixed-purpose crawling fell from roughly 49% to 33% of requests across the first half of 2026 as operators split composite bots into narrower tokens, and Meta's indexing crawler overtook its own training crawler in June 2026, an inversion the document uses to argue that any allowlist may be stale on publication.

One technical distinction recurs as the document's central point. Control tokens such as Google-Extended and Applebot-Extended never independently visit a site; they are robots.txt directives honoured by an operator's main crawler, and their scope is training only. Blocking them does not remove content from answer features built on the underlying search index. Anthropic's three-bot separation illustrates the same principle from the other side: ClaudeBot for training, Claude-Userfor live queries, Claude-SearchBot for search quality, clarified in documentation updated around February 20, 2026.

Operators also diverge on whether robots.txt applies to live fetchers at all. Google documents that user-triggered fetchers including Google-Agent generally ignore it because a person requested the page. OpenAI states that it may not apply to ChatGPT-User for the same reason. Anthropic states that all of its bots honour it. The guidance concludes that for this category robots.txt is not a reliable control and that enforcement belongs at the network edge.

A substantial portion of the guidance addresses impostor traffic, on the basis that user-agent strings are voluntary text and any scraper can claim to be any operator's crawler. The recommended verification runs to three steps: match the string against a maintained directory, confirm the source address against published ranges or forward-confirmed reverse DNS, and confirm that behaviour fits the declared purpose across crawl rate, paths accessed and robots.txt compliance. A claimed major-crawler string arriving from a residential internet service provider or generic cloud range is treated as almost certainly false. The document also warns against reflexive blocking of unidentified traffic, since some unrecognised requests originate from a site owner's own procurement stack: content management systems, site search vendors, translation tools, tag managers or data-feed partners. Cross-referencing high-volume unknown traffic against internal procurement records can prevent a publisher severing a service it already pays for.

On the commercial stakes, the guidance cites a Gartner projection that agents could intermediate more than $15 trillionin business-to-business purchasing by 2028, alongside an Adobe figure putting conversion from machine-referred traffic around 42% better than non-machine visits. It cautions that both come from provider-specific networks and measurement windows and should be read as directional. The volume-versus-value split runs through the whole document: a person shopping for shoes might visit five sites, while an agent completing the same task might visit hundreds, and Cloudflare's own dashboard data has shown crawl-to-referral ratios ranging from 118 crawls per referral at the low end to nearly 50,000 at the high end for some operators.

A dated deadline sits inside the document. Before September 15, 2026, site owners operating behind Cloudflare or any managed edge or web application firewall are directed to review crawler defaults, because new default settings will block Training and Agent crawlers on advertisement-carrying pages for new domains and free-tier zones, with multi-purpose crawlers assessed under the most restrictive applicable rule. That date traces to Cloudflare's July 1, 2026 announcement, which also shifted the company's compensation model away from per-crawl charging toward paying publishers when content demonstrably contributes to a generated answer.

The document situates all of this against Australian copyright law, which has no broad fair use doctrine and relies on narrower fair dealing exceptions, none covering model training. Industry consultation has moved toward licensing through the Copyright and AI Reference Group. Licensing itself is framed not as a sixth crawler category but as a relationship that can apply to any of the five once a commercial agreement exists, with the IAB Tech Lab's Content Monetization Protocol, finalised April 28, 2026, serving as the communication layer.

The operational sequence the guidance sets out runs to five steps: audit at least 30 days of edge or server logs before changing anything; evaluate each crawler on traffic, content use, bandwidth cost, reputational exposure and strategic value; apply one of the four verdicts; document the decision, rationale, conditions and review date, with weekly checks for new high-volume unknown traffic and quarterly reassessment; and publish the resulting policy in three forms, a human-readable page, a machine-readable robots.txt, and a named licensing contact. The document's argument for that last step is commercial rather than technical: a published licensing contact converts a passive block into a negotiable relationship.

The gap between declared and enforced access is the problem the guidance names directly. Ahrefs analysis of server-log data from 137,000 domains found that 97% of llms.txt files received zero artificial intelligence requests in May 2026, even as adoption grew 8.8 times to nearly 39,000 sites, with SEO audit tools rather than assistants accounting for most of the requests that did occur.

Retail media meets its measurement ceiling in Sydney

The same IAB Australia summit produced a second strand of coverage, and it converged on a single constraint. PPC Land reported on July 28 that the Australian retail media market has reached roughly $2 billion while 85% of Australian sales still occur in store, with seven in ten advertisers planning to raise retail media budgets. Proof of incrementality now sets the ceiling on what networks can charge.

The trade body's own research, covered separately, quantified the maturity gap. Only 22% of retail media networks rate their own offering advanced. 56% of buyers keep trade and media budgets apart, which is a structural obstacle to measuring a channel that sits between the two. 60% of buyers are lifting spend regardless.

Metcash supplied the sharpest version of the argument. PPC Land reported the company warning that an 800-screen retail media network loses brand dollars to incrementality questions, with panellists arguing that single-metric measurement caps what networks can charge. The underlying audience case is unusual: Total Tools draws 85% of its sales from professional trades, a segment brands rarely reach through general media, and that specificity is precisely what a sales-lift-only metric fails to price.

The consumer behaviour running underneath the channel shifted in the same session. A Sydney panel covered by PPC Land found 75% of shoppers thinking about shopping weekly and described conversational interfaces displacing the search bar as the entry point, with Uber selling Bunnings lawnmowers in under 30 minutes and Criteo tying basket-size gains to feed depth.

Amazon supplied the counterweight to the measurement complaint in the form of inventory expansion, which is the sector's habitual answer to a pricing ceiling. PPC Land documented ten capabilities landing in Brand+ and Performance+ from July 22, including audio in 16 locales and free first-party signals, with single-deal streaming television buying removed from Brand+ and default-on models testing the company's claims about advertiser control. The audio side widened again when Octave supply opened to Amazon DSP buyers on July 24, covering talkSPORT, Times Radio, Virgin Radio UK and associated podcasts across 15.5 million United Kingdom listeners, the company's fourth audio inventory deal in a month. Location data followed a similar path, with Berlin-based Adsquare ID-Precise segments reaching the 3P Audience Hub in six markets from July 23, matched to Amazon identifiers and activatable off the shelf.

Supply expansion does not resolve the incrementality question, though, and one of the week's more uncomfortable data points concerned what shelf access actually costs a small supplier. PPC Land reported on July 26 that Scorch Marker netted $37,000 on $252,125 through Walmart before losing the account, with consolidator fees alone reaching $76,218.95, more than the cost of the product itself, during a period when Walmart advertising revenue grew 46%. That is the same structural argument the Sydney panellists were making about measurement, viewed from the supplier side of the ledger rather than the network side.

A campaign-level result gave the measurement debate something concrete to argue about. PPC Land documented Galderma recording a 9x conversion likelihood from a three-format Amazon test combining Prime Video, DSP and sponsored advertisements, lifting consideration 4.5% on a budget the brand described as small. Chemist Warehouse nonetheless still takes the larger share of the brand's spend, which is a fair summary of where retail media sits: multi-format tests produce strong relative numbers, and budget allocation continues to follow the retailer relationship rather than the measurement.

The corporate structure around one of the sector's largest independents changed the same week. PPC Land reported on July 29 that Criteo dropped its French domicile, with ordinary shares beginning direct Nasdaq trading as the company ended its American Depositary Share programme following a Luxembourg switch. A United States merger targeted for January 2027 still requires shareholder approval. The move follows July reporting at AdExchanger that Vista Equity Partners and Quinti Capital had submitted a bid for the company at roughly 50% above a market capitalisation of about $1.1 billion.

Two infrastructure items rounded out the retail and data picture. On July 29, PPC Land covered Nexxen cutting data onboarding time to 24 hours with a self-serve feature validating audiences against an identity graph the company says has nearly doubled its global footprint, ahead of second-quarter results. The same day brought a VAB guide, covered by PPC Land, running to 15 pages and naming UID2, OpenID and RampID as identifier options while directing buyers to audit vendor match-rate accuracy before signing. Both land in the same week as the TAG finding on inverted quality signals, and both concern the same underlying question of whether a vendor number describes what it claims to describe.

Screen inventory kept growing outside the grocery aisle as well. VIOOH added 7,680 gym and salon screens in Brazil through a deal with b.drops on July 28, with gyms supplying 6,241 of them and 159 million monthly impressions, a network the company describes as 5% of the Brazilian digital out-of-home market. It is the second network VIOOH has characterised at that share, which is its own small illustration of how loosely market-share claims travel in a channel without agreed denominators.

On the streaming side, Digiday's Future of TV Briefing reported on July 29 that Comcast-owned FreeWheel is rolling out show-level reporting for advertisers using its Buyer Cloud platform, addressing a decade of buy-side complaints about not knowing which programmes carried their advertisements. The same briefing recorded YouTube advertising revenue up 13% year over year in the second quarter and TelevisaUnivision United States advertising revenue down 29%. AdExchanger described the parallel workaround on July 30, reporting that third-party vendors and agencies have been jury-rigging show-level CTV reporting by connecting anonymised content identifiers to conversion logs, with contextual vendors such as Peer39 assigning an identifier to each programme. The same roundup noted PubMatic shelving its OpenWrap Web Prebid wrapper and recommending Playwire as an onboarding partner for roughly 250 publishers.

None of these five stories resolved during the week. Meta booked a legal charge it cannot yet size fully and guided to capital expenditure it has not priced downstream. The verification stack continues to rate synthetic inventory above human work, and the measured gap outside made-for-advertising tooling sits at 12% of a category that is itself growing. AI Mode carries 1 billion monthly users and roughly a twentieth of the product visibility of the surface it sits inside. Cloudflare's default-blocking change takes effect September 15, and Article 50 binds August 2. Retail media in Australia has a $2 billion market and no agreed incrementality standard. What the week supplied was measurement of the gaps rather than closure of them, which is a more useful thing to have than it sounds.

Also noted