Two reports published within hours of each other on July 20, 2026 arrived at the same uncomfortable conclusion from opposite ends of the advertising business. One simulated what happens when an identity graph links an ad exposure to the wrong person. The other measured how often the data used to find households with children is actually correct. Neither found a problem of volume. Both found a problem of truth.

The numbers are severe enough to change what a campaign looks like on a dashboard. In the first study, a campaign genuinely returning $1.50 for every dollar spent was measured at $0.43. In the second, a $1 million campaign aimed at households with children wasted $590,000 of that budget. Neither failure announces itself. Both produce clean, plausible reports that a media team would have no obvious reason to question.

What follows examines those two findings, the commercial moves happening alongside them, and the regulatory machinery now assembling around advertising data on both sides of the Atlantic.

The measurement error that looks like the truth

LiveRamp and the Marketing + Media Alliance published joint research on July 20, 2026, presenting the findings the same day at the MMA CMO + CEO Summit in Santa Barbara, California. The report carries the title The Missing Piece: Improving Confidence in Marketing Measurement. LiveRamp, the San Francisco data collaboration company listed on the New York Stock Exchange under the ticker RAMP, ran the analysis with MMA using synthetic data and its own technical staff. The MMA is a nonprofit alliance of chief marketing officers operating in 16 countries with more than 825 corporate members.

According to the research, poor identity precision collapsed campaign return on investment by approximately 70% in simulation. The distortion was severe enough to justify cancelling a campaign that was working.

The report divides measurement risk into two categories that behave differently and require different fixes. Missingness describes real ad exposures that were served and seen but never recorded, leaving gaps that underrepresent what happened. Identity mismatch describes exposures that were recorded but attributed to the wrong individual. In the second case the data looks complete. The errors sit inside it.

That distinction maps onto two analyses practitioners routinely conflate. Attribution asks which ads led to a conversion and estimates each touchpoint's contribution. Incrementality asks the harder causal question of how much advertising caused conversions that would not have happened otherwise. The Interactive Advertising Bureau and IAB Europe defined incrementality in commerce media guidelines released in November 2025 as the causal impact of marketing, distinguishing it explicitly from attribution and return-on-ad-spend calculations that show outcomes rather than causality.

Random loss is survivable, patterned loss is not

The missingness findings overturn an intuition. Random data loss barely matters. Removing 20% of impressions at random across all publishers preserved the baseline channel ranking entirely. A frequency-dependent scenario, where heavy users lost more impressions as cross-device tracking broke, stripped out 26.9% of impressions and still preserved the ranking.

Non-random loss is where the damage sits. When missing impressions concentrated among converters, the pattern produced by privacy controls or checkout flows that fail to fire a pixel, channel rankings reversed at an aggregate loss of roughly 1%. A volume most audits would treat as rounding error flipped the answer.

The distortion was invisible to standard diagnostics. According to the report, the model's own accuracy metric, the area under the curve, improved under outcome-correlated loss even as the underlying ranking corrupted. A team watching its dashboard would have seen the numbers get better while the answer got worse.

The simulations used real impression logs comprising 1.9 million exposures across 147,941 users and four publishers, paired with 12,956 transaction outcomes and a 1.82% empirical conversion rate. The attribution model was a logistic regression trained on per-publisher impression counts, with a baseline ranking across the four publishers of T over Y over G over M.

Why a high match rate offers no protection

The identity findings produce the headline number, and the mechanism explains why match rate is the wrong thing to optimise. In the incrementality simulations, the researchers held true campaign performance constant at a 25% lift and $1.50 return, then tested how different measurement designs read that ground truth under each failure mode. In the identity mismatch scenario they set campaign reach at 30% and identity precision at 50%, meaning half of all matched users were linked to the wrong person.

Relative lift, truly 25%, was measured at roughly 6.8%, a collapse of about 73%. Return on investment, truly $1.50, was measured at $0.43. On paper the campaign appeared to lose 57 cents on every dollar spent.

The report labels the 50% precision figure an illustrative lower bound demonstrating how the mechanism propagates, not a typical condition for campaigns running on deterministic, people-based identity graphs. The arithmetic is simple once stated. False positives, unexposed people wrongly labeled as treated, dilute the treatment group. False negatives, exposed people wrongly labeled as control, inflate the baseline. Both biases push measured results downward, and they compound.

Match rate, the share of records a system can link, has been the headline metric identity providers compete on for a decade. According to the report, a high match rate is insufficient if the matches themselves are wrong. Precision, not recall, governs measurement integrity. AdExchanger has been circling the same doubt from the practitioner side, asking whether effective attribution is now largely an illusion as match rates fall across the industry.

Not every method proved equally fragile. Randomized controlled trials using an intent-to-treat design, counting everyone assigned to a test regardless of confirmed exposure, remained directionally correct across every missingness scenario tested, holding return at the ground-truth $1.50 even in the outcome-correlated scenario that broke everything else. An RCT restricted to reached-only users drifted to $0.91 under the same conditions, because filtering to confirmed exposures inherits the identity graph's errors directly. Quasi-experimental and modeled designs fared worst: in the outcome-correlated loss scenario, a design of that class read a truly profitable $1.50 campaign as a loss of $2.65, flipping the sign entirely.

The commercial subtext deserves noting. The report's proposed remedy is data collaboration through identity solutions and clean rooms, which is precisely LiveRamp's business. A vendor-commissioned study concluding that the vendor's product category solves the identified problem warrants reading with that interest visible. The Federal Trade Commission cautioned in a November 2024 analysis that data clean rooms are not privacy-preserving by default. Publicis Groupe separately agreed to acquire LiveRamp for $2.5 billion in an all-cash deal announced in May 2026 at $38.50 per share, folding one of the industry's larger neutral identity providers into a holding company.

The 42% problem in children's media

Five days before the LiveRamp report, the Coalition for Innovative Media Measurement published a study that gives the same failure mode a price tag in a specific market. CIMM released When Metrics Fall Short: The Case for Children's Media Measurement on July 15, 2026, and simultaneously convened a Kids and Family Media Measurement Taskforce with ten founding members. The report was authored by Emily Horgan, an independent media consultant and former Disney executive, developed from interviews with more than 30 industry stakeholders alongside a CIMM executive roundtable held in April 2025. It runs to 47 pages.

The central finding concerns data advertisers already buy. Independent data science firm Truthset analyzed more than 4 billion US consumer records from the fourth quarter of 2024 and found that identifying whether a household contains children succeeds only 42% of the time. Broken down by the gender of the adult record, accuracy for identifying whether a woman has children present sits at 36.5%, and for men it falls to 28.7%. The same providers correctly classify gender itself 83.9% of the time for women and 82.2% for men.

The gap comes from how the two attributes are generated. Gender can be inferred with reasonable confidence from first names, survey responses, and consumer registrations. Parental status is rarely declared in commercial datasets, so providers fall back on proxies: toy purchases, diaper buys, magazine subscriptions. Those signals lapse as children age, they can reflect gifts for nieces, nephews, or grandchildren rather than children in the home, and a single diaper purchase logged against a male record can produce a false parent signal that never gets corrected. Truthset also found accuracy degrades by up to 50% with each additional data match as records move through onboarding, device graphs, and demand-side platforms.

The financial translation is explicit. At 41% accuracy, a $1 million campaign targeting households with children at a $10 CPM wastes $590,000, producing an on-target CPM of $24.39. Working with higher-accuracy data at an average 58% score cuts the effective CPM to $16 and reduces waste to $210,000. Running the same budget against the least accurate records, averaging 21% accuracy, pushes the effective CPM to $48 and burns $790,000. Truthset's measurement of roughly 1 billion hashed email addresses in the first quarter of 2025 put the average accuracy score for presence of children at 41%, with individual providers ranging from 33% to 67%, and no consistent advantage for larger vendors.

Underneath the accuracy problem sits a decade of redistributed attention. Drawing on Nielsen and Common Sense Media data, the report states that linear television, which accounted for roughly 80 to 90% of young children's video time in 2005, represented only around 6% by 2025. Premium streaming and YouTube together now capture 67% of children's video consumption, split 35% for streaming services and 32% for YouTube, with short-form content on platforms such as TikTok and YouTube Shorts adding a further 16%. Daily screen time for children aged 0 to 8 rose from about 1 hour 44 minutes in 2005 to 2 hours 27 minutes by 2025. Consumption expanded while standardized cross-platform visibility deteriorated.

Two environments account for large shares of children's time yet remain largely invisible to standard measurement. Gaming platforms such as Roblox, Minecraft, and Fortnite are interactive, session-based, and user-generated, with no standardized program units that panel-style logging can capture. YouTube supplies views, watch time, and retention at channel level, but demographic data only for users aged 13 and older because of COPPA restrictions, leaving its substantial under-13 audience without audience context. Nielsen data cited in the report shows children hold a 23.9% share of viewing to YouTube Main.

Panels and big data each fall short for structural reasons. In a panel of 50,000 households, perhaps 5,000 contain children under 12; divide that across age cohorts and then across platforms and sample cells shrink to single digits. Smart TV automatic content recognition and set-top box logs offer scale but rarely identify child viewers, since children under 12 generally lack phones, email addresses, or accounts in their own names. Such data may confirm a children's program played on a given television without revealing who was in the room.

The reliability question extends past children's data. A CIMM and Go Addressable study reported in late 2025 found IP-based household targeting can miss 87% of households, with Truthset analysis putting postal accuracy at 13% and email accuracy at 16% across nearly a billion IP records. Jon Watts, Managing Director of CIMM, framed the stakes by noting the children's media landscape has changed while the systems used to measure audiences have not kept pace.

Netflix removes the last barrier to its inventory

Against that backdrop of uncertain measurement, the commercial machinery kept opening up. Netflix inventory became available to every buyer using The Trade Desk's Sellers and Publishers 500+ marketplace on July 20, 2026, removing the private-deal requirement that had governed access since the streaming service entered programmatic advertising in 2024.

The mechanics are narrow and consequential. Advertisers previously had two routes: programmatic guaranteed deals with volume and price fixed in advance, or 1:1 private marketplace arrangements negotiated between buyer and seller. Both require direct engagement and, typically, scale commitments smaller advertisers cannot meet. According to The Trade Desk, Netflix's inclusion replaces that gated model with always-on marketplace deals accessible through existing controls inside Kokai, with no minimum spend and no workflow change. The activation path runs through Kokai's Forward tile, where buyers locate the Marketplace Summary, select Edit, and choose Sellers and Publishers 500+.

The inclusion gives buyers access to more than 250 million global monthly active users, a figure aligning with what Netflix presented at its 2026 advertising upfront on May 13, 2026. That number represents individual viewers within ad-supported households rather than subscriber accounts.

Timing matters here. Netflix has spent 2026 chasing roughly $3 billion in advertising revenue, nearly double the approximately $1.5 billion generated in 2025. Two days before the marketplace update, the company reported second-quarter revenue of $12.56 billion, a 13.4% year-over-year gain that fell short of its own guidance and sent shares down roughly 8%. The advertising arm held its pace. Netflix told investors on its first-quarter 2026 earnings call that programmatic buying was on course to become more than 50% of its non-live advertising business, confirming more than 4,000 active advertisers, a 70% year-over-year increase.

For The Trade Desk the addition strengthens a curated marketplace facing competitive pressure. Amazon has undercut it on fees, with ad sales fees of roughly 1% standing far below the 12 to 15% range that investment bank Needham and Co. attributed to The Trade Desk. The pattern is not unique to Netflix: Samsung opened its Smart TV home screens to programmatic buying in June 2026 via The Trade Desk and Google DV360.

What the announcement omits is as notable as what it contains. There is no pricing detail, no disclosure of how marketplace inventory will be priced relative to private-deal equivalents, and no data on impression volume flowing through the always-on channel. Nor does it specify which formats are included, or whether the live sports and pause-ad inventory flagged for this summer will be reachable by the same route.

The toggle that Google did not want to build

The same day, Google published Search Console documentation for a control letting website owners keep their pages out of AI Overviews, AI Mode, and generative AI features in Google Discover. The control appears under Settings in a section labeled Search generative AI, and is rolling out to a subset of owners.

The documentation describes three configurable states. Beyond include and exclude, a property can inherit its value from a parent property, and inheritance is the default whenever a parent exists. According to Google, the exclude setting stops a site appearing in the listed surfaces, and crawled content from an excluded site will not be eligible as an input to generate an AI response or preview.

The boundaries matter as much as the function. The control affects only whether content can appear in certain generative features and is not used as a ranking signal elsewhere in Search. It does not override participation in Merchant Center or Google Ads. Most consequentially, it does not affect AI training, which remains governed by Google-Extended, a separate robots.txt token covering whether crawled content may train future Gemini models. Blocking content from Search entirely still requires noindex. An owner wanting content withheld across Google's AI surfaces must therefore operate three distinct controls, each addressing a different stage of how content moves from a crawl to a generated answer.

Changes take effect within 1 to 2 days after the control goes live, though some content may take longer to disappear because of caching and propagation.

This was not a voluntary product decision. Google began testing the toggle on June 3, 2026, the same day the UK Competition and Markets Authority imposed its first binding conduct requirement under the Digital Markets, Competition and Consumers Act 2024. Substantive obligations of that Publisher Conduct Requirement take legal force on December 3, 2026. The domain-level limitation traces directly to the regulatory record: the CMA's final decision noted Google said page-level grounding controls would require additional engineering work, and the regulator set a nine-month implementation period placing their arrival on March 3, 2027. A senior Google executive had described a clean opt-out as a substantial engineering challenge at a London conference on February 11, 2026.

The stakes come from scale and from documented traffic erosion. AI Overviews count more than 2.5 billion monthly active users and AI Mode crossed one billion by May 2026. Ahrefs research found AI Overviews correlate with a 58% reduction in click-through rates for pages ranked first. A later randomized field experiment produced causal evidence: AI Overviews cut publisher clicks by 39.8% among 1,065 desktop Chrome users and raised zero-click searches by 34.5%.

Because every property begins in the include state, the burden falls on owners to notice the control, understand its layered scope, and act. Smaller operators without teams monitoring Google product launches are least equipped to exercise a right that requires attention to exercise. The European Commission opened a probe into Google's AI content practices in December 2025 over these dynamics.

That same tension surfaced elsewhere on July 21. AdExchanger's daily roundup reported on the open web's contraction, citing New York Times reporting that people spend on average between one and nine more minutes in AI Mode than standard search, and a Growth Memo study finding only about a quarter of AI Mode sessions result in a click through to a web link. The Wikimedia Foundation says human traffic to Wikipedia declined 8% over the past year even as total visits rose because of AI bots scraping the site. Nilay Patel, editor-in-chief of The Verge, told the Times that for publishers, Google Zero has already arrived.

Independent behavioural data published by Northeastern University supports the concentration thesis. Alphabet accounts for 35.0% of US desktop browsing time, according to a study of 4,608 weighted users drawing on 53,243,995 website visits collected between June 2024 and December 2025. Gmail alone took 16.4%, more than the bottom 300,000 websites combined. The top 10 sites captured 49.7% of all browsing time. Large language model tools now account for 2.9% of total online time against 2.5% for news sites, placing AI ahead of the entire news web as a browsing destination.

What AI actually costs, and what it returns

The measurement problem has a mirror image on the spending side, and Digiday examined it on July 21, 2026. Agencies are working out what AI actually costs them as usage moves past pilots into daily operations.

PMG rolled out a tool called Alli For You to its full company a month ago, pooling staff access to major models under a $50-a-day token cap per user. The cap followed months of alpha and beta testing begun in January, when testers ran without limits. Kaitlin McGrew, head of SEM at PMG, framed the reasoning around peak load, noting the agency wants enough tokens on hand for a Black Friday when ads, reporting, and agents all run at once, and describing the exercise as governance.

The harder question is what the spend delivers. Caroline Giegerich, vice president of AI and marketing innovation at the IAB, observed that the industry started at time saving and is now moving toward business impact. Tracking token spend is straightforward; knowing what it produced is not. Agencies have landed on three incompatible answers. Dept declines to pass token costs to clients, arguing that itemizing them makes the metric the tokens spent rather than the work produced. S4 Capital's Monks builds tokens directly into tech-and-subscription pricing. Large holding companies fold AI costs into broader commercial structures such as principal media deals.

Publicis CFO Loris Nold, answering a question on the company's earnings call about a 7% rise in other operating costs partly driven by AI, pointed to margin improvement as evidence the productivity benefits offset the spend. That improvement came to 17 basis points in the first half of the year, after more than 30 basis points of savings had already been reinvested into AI tools and staff training. The offset is spent before it reaches the bottom line.

Procurement expectations compound the difficulty. The assumption baked into most negotiations is that AI means lower headcount and lower cost, not better work worth paying more for. Joe Maglio, CEO of Cheil Agency Network, said the network is moving agencies toward output-based pricing, with all new business on that basis and half of existing clients transitioned. He is making that shift without having solved the measurement problem, because standing still is not an option in a market where AI aptitude is itself a selling point.

The revenue side of AI advertising is running behind projections too. MediaPost reported on July 21, 2026 that OpenAI will miss its five-year ad revenue projections by 90%, citing a Futurism report sourcing Emarketer analysis. The combined advertising total for OpenAI, Google, Microsoft and Amazon in the relevant period is projected to total less than $1 billion, against the $2.5 billion OpenAI alone had predicted reaching this year.

Regulators move on the advertising itself

The final thread running through the window concerns who becomes liable when advertising data goes wrong. Ofcom opened a consultation on 10 July 2026 setting out nearly 40 draft measures requiring the UK's largest social media and search platforms to police the paid advertising they carry. It marks the first time regulated services would face binding duties on fraudulent adverts under the Online Safety Act 2023.

The draft Fraudulent Advertising Codes of Practice apply to services designated Category 1 or Category 2A on Ofcom's 2026 register. They cover paid advertising only, explicitly excluding user-generated content and non-sponsored search results. Feedback closes 2 October 2026, with a final statement planned by mid-2027 at the latest. More than £40 billion a year is spent on digital advertising across the UK, according to Ofcom, and an estimated £200 million-plus is lost by victims to these scams annually.

The proposals lean on account-level intervention. Measure H5 would require providers to ban advertising account holders posting fraudulent adverts and take reasonable steps to prevent their return. Measure H1 would require an account checks and actions policy, reviewed at least every 12 months, verifying that holders work for the organisation they claim to represent. Measure H2 would require financial services verification confirming advertisers hold appropriate legal permissions, anchored to Financial Conduct Authority authorisation. Measures H3 and H4 address account takeover through security mechanisms and reporting routes.

Generative tools get direct treatment. Where a provider makes an advertisement generation tool available, measure F1 would require testing to identify whether it could be used to create fraudulent adverts. A separate transparency measure would require a publicly available ad library containing all adverts encounterable by UK users while live and for a year afterward, updating at least daily, with multi-criteria search and an exposed API. Penalties reach £18 million or 10% of global revenue, whichever is greater.

AdExchanger covered adjacent ground the following day, reporting expert commentary on an earlier Ofcom paper on fraudulent advertising and account integrity. Rob Leathern, who led ads integrity and privacy teams at Facebook and Google, argued the most important thing a code could mandate is platform provision of large, representative random samples of all active ads to trusted third parties, together with the total number of ads running. No major platform currently reports this. He also noted a structural misalignment: advertisers still pay when ads run on fraudulent publishers, refunds often arrive as ad credits, and victims of search or social fraud are not repaid by the platform.

Elsewhere in the window, the European Commission fined AliExpress 550 million euros over failures to remove unsafe products, with the platform given until October 20 to file a remedy plan. Brussels also ordered Google to share anonymised search query data with rivals and AI chatbots starting January 2027, with Android AI rivals waiting until August 2027 for full access.

The common thread

Four separate findings, produced by four organisations with no coordination between them, converge on one point. LiveRamp measured what happens when identity data is wrong and found a profitable campaign reading as a loss. CIMM measured how often one widely purchased data attribute is right and found 42%. Northeastern measured where attention actually goes and found a third of it inside one company. Digiday found agencies able to track AI spend precisely while unable to state what it returned.

Each describes a system generating confident outputs on inputs nobody has verified. The commercial responses arriving alongside them, Netflix dropping its spend floor and Google shipping a domain-level opt-out, both widen participation without addressing whether the underlying numbers describe reality. Ofcom's draft code is the first proposal in the window that would make somebody legally answerable for the difference, and its final form remains 12 months away at minimum.

The LiveRamp report closes on the point that generalises furthest. The most dangerous measurement error is not the one that looks like an error. It is the one that looks like the truth.

Also noted

  • July 20, 2026 - Google will force local inventory ads on all Shopping campaigns from August 31, removing the enable_local opt-out, with v25.1 code setting the field to false triggering an OPERATION_NOT_PERMITTED error at runtime. Read more, and see the advertiser email Search Engine Roundtable published on July 16
  • July 20, 2026 - Semrush surveyed 519 US professionals and found brand recognition sways just 7% of B2B buyers inside AI answers, with precise use-case fit mattering more; 84% use AI on deals above $1,000. Read more
  • July 20, 2026 - A survey of 45 studies found generative engine optimization rewrites confined to page bodies can cut a page's presence in AI retrieval results by 16%, while vendors continue selling citation gains the studies never confirmed per engine. Read more
  • July 20, 2026 - Amazon DSP added Triton audio supply across a marketplace spanning more than 80 countries, letting buyers layer Amazon first-party shopping signals onto Triton inventory following iHeartMedia's June DSP agreement. Read more
  • July 21, 2026 - Digiday's Ad Tech Briefing examined Apple's advertising ambitions after its effective discontinuation of SKAdNetwork, noting newly published Apple Maps advertising policies ahead of an expected launch later this summer. Read more