Six intermediaries, six partial logs, and no end-to-end record
A single programmatic impression — an ad bought by machine auction rather than by insertion order — may pass through the advertiser, its agency, a demand-side platform that buys, a supply-path optimization layer, an exchange, a supply-side platform that sells, the publisher's ad server, and finally the page. Each hop keeps its own log and takes a fee. No single party holds a record of the whole path.
That is the defining evidentiary feature of the channel and the reason discovery here is expensive. The advertiser's dashboard is a summary produced by one participant in a chain, and the participants downstream of it are not parties, are not necessarily identified, and have no obligation to volunteer anything. Where an expert witness on digital ads earns their keep here, it is usually in specifying — before a subpoena or audit demand goes out — which log at which hop would answer the question, and in what fields.
One scale point explains the mechanism rather than the size of the problem. The Association of National Advertisers reported that the average campaign in its programmatic study ran across roughly 44,000 top-level domains. A campaign spread that wide cannot be reviewed by a human, so quality control is delegated to systems whose outputs are the very thing in dispute.
Log-level data: what it is, and why the advertiser usually does not have it
Log-level data is the granular record of an impression as the technology vendor holds it. The ANA's definition — "the detailed record of everything about an impression obtained from the tech vendor" — is the one worth quoting, because it makes clear whose record it is. Fields typically include timestamp, placement, domain or app, the price paid at each stage, the identifiers of the intermediaries involved, and a verification vendor's determinations where one was measuring.
Getting it is a contractual question before a procedural one. Where the agency or trading desk holds a data-access right in its master services agreement, the fastest route is that right, exercised in writing, before positions harden. Otherwise the routes are a Rule 34 request to the agency, a third-party subpoena, or a negotiated audit — each slower than the retention windows vendors actually apply.
The ANA's study is the best-sourced public account of what is and is not available, and its provenance belongs in the same sentence as its findings. Published on 19 June 2023, the first look covered 21 member companies, $123 million of ad spend and 35.5 billion impressions collected between September 2022 and January 2023. It reported that made-for-advertising sites took 21 percent of study impressions and 15 percent of ad spend, and framed an $88 billion open-web programmatic market as carrying as much as $20 billion in recoverable waste. The ANA is the advertiser-side trade body, the participants volunteered their own log-level data, and the $20 billion figure is an extrapolation from a 21-advertiser sample rather than a measured loss. The ANA's own release says what it measured. Later releases in the same series report materially different figures, so a citation to the 2023 study is a citation to 2022–23 data and should be dated as such.
ads.txt and sellers.json establish authorization, not legitimacy
ads.txt — "Authorized Digital Sellers" — is a plain text file a publisher hosts at the root of its domain listing the ad systems authorized to sell its inventory and the publisher's seller account identifier in each. app-ads.txt is the equivalent for mobile and connected-TV apps. sellers.json is the counterpart published by the ad system, disclosing the sellers whose inventory it carries and whether each is a direct publisher or an intermediary. All three are IAB Tech Lab specifications, and the stated purpose of the first is to let content owners declare who is authorized to sell their inventory.
They are routinely overstated in exhibits. They are self-declarations that nobody audits; a fraudulent entry in a file the fraudster controls is not a barrier. They authorize a seller account rather than a specific impression — the file says this seller may sell for me, not this impression was mine. A domain's file can be changed at any moment and is not versioned, so establishing what it said on a past date requires a contemporaneous capture or a third-party archive, which is a preservation task nobody thinks to do until it is too late. And they address domain spoofing and unauthorized reselling only. They are silent on invalid traffic, on viewability, on brand safety, and on inventory quality, all of which occur on perfectly compliant inventory. The specification itself is explicit about its own scope limits, which is useful when opposing counsel treats compliance as a clean bill of health.
Viewability: state which standard applies before quoting a number
Viewability is the item an expert is most often asked to recite from memory, and it is where people who learned the field from blog posts get impeached. The MRC's Viewable Ad Impression Measurement Guidelines, version 2.0, dated 18 August 2015, define a viewable display impression as 50 percent or more of the advertisement's pixels on an in-focus browser tab in the viewable space of the page for one continuous second or more, post ad render. For large formats of 242,500 pixels or greater the threshold drops to 30 percent. Video requires 50 percent of pixels for two continuous seconds of play. Google's Active View restates the same thresholds and is the operative implementation for most buyers.
This is not the only MRC threshold. The MRC's Cross-Media Audience Measurement Standards (Phase I Video), final, September 2019, apply a threshold of 100 percent of pixels on screen for at least two continuous seconds, so that digital video is comparable with linear television, where the whole ad is on screen by definition. Both numbers are the MRC's, they answer different questions, and neither supersedes the other. A campaign can report 80 percent viewable under one and far less under the other with no change in delivery.
Four things the 2015 guidelines deliberately do not require are each a cross-examination point: that a human saw the ad, that it was legible or attended to, that the video was audible, and that every impression was measurable. On the last, unmeasurable impressions are excluded from the denominator in some reporting and included in others — the measured rate and the total rate are different numbers, often by a wide margin. Always establish which one an exhibit shows. The MRC guidelines should be read before testimony rather than recalled, and the MRC's current standards index checked for any superseding document.
Misplacement disputes, and the standard that no longer exists
The vocabulary is used loosely and should not be. Brand safety is avoiding content categories a brand does not want to sit beside. Brand suitability is the narrower, advertiser-specific version of that judgment. Ad misplacement is an ad appearing where it should not have. Invalid traffic is a different problem entirely.
Until recently there was a shared vocabulary. The Global Alliance for Responsible Media, an initiative of the World Federation of Advertisers, published a brand safety floor and suitability framework supplying the categories vendors worked from. GARM was discontinued on 9 August 2024, days after X Corp. sued the WFA and several member advertisers. That litigation settled on 29 July 2026, and in announcing the settlement the WFA committed to not reforming or restarting GARM or any similar initiative.
Stated neutrally and without any view on the merits of that dispute: there is currently no neutral, industry-agreed definition of brand-unsafe content. Verification vendors continue to apply taxonomies derived from GARM's categories, but in a misplacement matter the classification standard is itself contestable, and an expert should expect to be asked whose taxonomy was applied and on what authority.
The evidence in a misplacement matter is specific: the demand-side platform's log-level data showing which URL, which placement and which time; the verification vendor's classification of that URL as of that time; the campaign's configured blocklist, allowlist and category exclusions; and the change history showing when those settings moved. The hardest element is almost always the state of the page at the moment of the impression. Pages change, and a URL that is benign today may not have been.
Made-for-advertising inventory, where every metric can be green
Made-for-advertising sites are built to arbitrage traffic and monetize impressions rather than serve an audience: high ad density, autoplay video, aggressive refresh, paid acquisition, thin or aggregated content. The ANA characterized them as usually featuring low-quality content.
The definitional problem is the whole fight. There is no bright line between an aggressive but legitimate publisher and a made-for-advertising site, and vendors classify the same inventory differently. The ANA benchmark does not define the category itself; it adopts a vendor's classification. Anyone asserting a specific site is made-for-advertising is applying a contested standard, and should say whose.
The evidential interest is that this inventory is usually technically compliant: ads.txt in order, viewability high because the pages are built to be viewable, invalid-traffic rates within tolerance. It fails on outcomes and not on the measured metrics, which makes it the cleanest illustration of the theme running through this channel: every indicator can be green and the media can still have been worthless. That is an argument about media buying judgment and about what a competent buyer would have caught — a standard-of-care question — rather than an argument about whether a number is accurate.
What the display record cannot settle
Start with the one that surprises people. The advertiser has no independent record of an impression at all. It has a server-side record of a landing-page request, because the request hit its own infrastructure. Everything it knows about impressions comes from the platform or from a measurement vendor's tag. There is no advertiser-controlled artifact to check the platform's count against, which is why impression-side claims are structurally harder to work than click-side claims, and why the ad-fraud schemes that have actually produced convictions were publisher-side inventory and impression operations built out by the government with compulsory process rather than by an advertiser reading an export.
A demand-side log and a supply-side log for the same impression may not reconcile, and the gap is not necessarily misconduct — timing, deduplication rules and each party's own definitions all contribute. A viewability rate does not establish attention, and cannot be compared across vendors, whose methodologies and denominators differ; two vendors measuring the same inventory routinely report different rates, because the Media Rating Council's accreditation certifies adherence to process standards rather than identical outputs. Authorized-seller files do not establish that any impression was legitimate. And an industry waste estimate says nothing about a particular campaign: aggregate prevalence is not individual causation.
What the record does support, when it is obtained early and read carefully, is a defensible account of configuration and delivery — what was bought, at what price, through which intermediaries, onto which domains, against which exclusions, and what the measuring vendor determined at the time. That is a narrower claim than most reports in this channel make, and it is the one that survives.
Frequently Asked Questions
What is log-level data, and can an advertiser actually obtain it?
Log-level data is the impression-by-impression record held by a technology vendor — the ANA describes it as the detailed record of everything about an impression obtained from the tech vendor. Advertisers frequently cannot get it, which was the central complaint of the ANA's programmatic study. The realistic routes are a data-access right in the agency or trading-desk contract, exercised in writing; a Rule 34 request to the agency; or a third-party subpoena to the platform. All three are slower than vendor retention, so the request should go out before positions harden rather than after.Does an ads.txt file establish that an impression was legitimate?
No. Anads.txt file is a publisher's unaudited self-declaration of which ad systems may sell its inventory, and sellers.json is the seller's counterpart declaration. They authorize a seller account, not a specific impression, and they can be edited at any time without versioning, so establishing what a file said on a past date requires a contemporaneous capture. They were designed to address domain spoofing and unauthorized reselling. They say nothing about invalid traffic, viewability, brand safety or inventory quality, all of which occur on fully compliant inventory.Which viewability standard applies to an exhibit reporting a viewability rate?
Ask, because there is more than one and they produce different answers. The MRC's Viewable Ad Impression Measurement Guidelines v2.0 of 18 August 2015 require 50 percent of pixels in view on an in-focus tab for one continuous second, with a 30 percent threshold for formats of 242,500 pixels or more, and two continuous seconds for video. The MRC's Cross-Media Audience Measurement Standards (Phase I Video) of September 2019 require 100 percent of pixels for two continuous seconds. Both are MRC standards answering different questions. Also establish whether the figure is a measured rate or a total rate.Is there still an industry standard for what counts as brand-unsafe content?
Not a neutral one. The Global Alliance for Responsible Media, run under the World Federation of Advertisers, supplied the shared brand safety and suitability categories that verification vendors worked from. It was discontinued on 9 August 2024, and in settling litigation with X Corp. on 29 July 2026 the WFA committed to not restarting it or a similar initiative. Vendors continue to apply taxonomies derived from the old categories. In a misplacement matter that means the classification standard is itself contestable, and an expert should expect to be asked whose taxonomy was applied and on what authority.Should a DSP log and an SSP log for the same impression reconcile?
Not necessarily, and a failure to reconcile is not by itself evidence of misconduct. No single party in the programmatic chain holds an end-to-end record. Timing, deduplication rules, fee accounting and each participant's own metric definitions all produce legitimate divergence. The useful expert work is characterizing the size and direction of the gap, identifying which definitional differences account for how much of it, and stating what remains unexplained — rather than presenting the gap itself as a finding. An unexplained residual is a reason for further discovery, not a conclusion.What can an advertiser establish about impressions it never independently recorded?
Less than most exhibits assume. Unlike a click, which leaves a request in the advertiser's own web server log, an impression leaves no advertiser-controlled artifact. Everything known about impressions comes from the platform's count or a measurement vendor's tag, so there is no independent record against which to test the seller's own reporting. That asymmetry is why impression-side claims usually depend on third-party verification data obtained early, or on discovery from parties further down the supply chain, and why an expert should say plainly which of the two an opinion rests on.Published