Independent benchmarks place the gap between the best- and worst-measured advertisers at a record high. For boards, that moves ad fraud out of the marketing department and into the capital-allocation conversation.
The invoice for a fraudulent advertising impression looks identical to that for a real one. Same format, same line items, same reference number, cleared by the same accounts payable process. The dashboard reporting on it glows the same shade of green. Nothing in the paper trail suggests a failure because, in the narrow terms of the contract, there was none. An impression was served. Somebody was paid.
That absence of a visible loss event is why digital advertising fraud has been filed since it ever surfaced under nuisance rather than risk. No breach notification, nothing. Marc Dhalluin, co-founder of the forensic media analytics firm TruthsetsOnline.com, puts the mechanism plainly: “Nobody in this chain has to be a criminal for the outcome to look like a crime. Every hop gets paid whether the impression was human or not.”
What has changed is not the waste. It is the distribution of it.
The spread, not the leakage
For years, the industry’s own numbers described a broadly uniform problem. The Association of National Advertisers’ widely cited 2024 programmatic transparency work found that of every $1,000 spent through programmatic channels, roughly $439 reached a consumer as a quality impression, an improvement of 7.9 percentage points on the prior year. It was a poor figure, but it was everybody’s poor figure, and a cost borne equally by all competitors is a tax, not a disadvantage.
The ANA’s Q1 2026 benchmark, built on log-level data from major advertisers and platforms, breaks that comfort. Higher-performing advertisers converted 54.0 per cent of programmatic spend into qualified impressions. Lower-performing advertisers converted 32.1 per cent. The resulting gap of 21.9 percentage points is the widest the benchmark has recorded (still by learned findings on the ‘conservative’ side, and the ANA describes the divergence as increasingly structural, driven by deliberate quality management rather than by budget size or luck.
That single number is the one worth taking to an audit committee, because it reframes the problem. Two competitors can deploy the same budget to the same audience, through the same adtech stacks, and one can get two-thirds more working media than the other; the optimised advertiser avoids the junk. The benchmark’s quality-adjusted cost measure makes the compounding visible: a headline CPM difference of $1.95 between the cohorts becomes a difference of $11.58 once waste is stripped out, $7.46 per thousand qualified impressions against $19.04. The lower cohort lost a massive 38.4 per cent of spend on media quality issues, more than double the horror 19.0 per cent lost by the higher cohort.
Applied locally, the arithmetic is unsentimental. A South African advertiser committing R10m a year to programmatic and sitting in the lower cohort loses close to R3.8m before a real consumer encounters the message.
Why the instruments failed
The gap is a measurement story before it is a crime story.
For most of the last decade, the industry’s primary quality signal has been viewability, defined by the Media Rating Council as 50 per cent of a display advertisement’s pixels on screen for one second, and two seconds for a video ad. It was designed as a floor, and it worked as one. It was never evidence that a human had the opportunity to notice anything. Eye-tracking research from Lumen Research has consistently found that around 70 per cent of technically viewable advertisements are never actually seen.
A mechanical, publicly documented metric is also a specification to build against. Ad pixels load into hidden impressions; impressions are often stacked, hidden in the background, placed off-page; bots simulate scroll behaviour, ‘touch’ screens, swipe, and clear every technical threshold with no person near the screen. Where campaign optimisation is anchored to viewability scores, the buyer has inadvertently published the instructions for defrauding them.
The second failure is quieter and more consequential. Legacy verification services typically measure a sample of impressions rather than the full campaign. When a bot strips or blocks the measurement tag, the system records nothing, and it generally reports an impression with no data as clean. No data becomes no fraud, which is a category error with a running cost attached.
None of this reflects vendor negligence. They were built to detect an earlier generation of crude, high-volume invalid traffic, and they performed against that threat. Gillian Rightford, executive director of the Association for Communication and Advertising, is right to note that a portion of invalid traffic is not malicious at all, being the residue of search crawlers and automated systems. The industry has made real progress, which the ANA data confirms. The problem is that the adversary has improved faster than the instruments.
When the fraud trains the buyer
The most expensive dynamic is not the wasted impression. It is what the wasted impression teaches.
Programmatic platforms optimise based on feedback signals from ads that were interacted with. Irrespective of whether the signal was the result of genuine human action or not. Completion and view rates, click-through rates, engagement events and landing page interactions determine where the next tranche of budget goes. Fraud operations have adapted accordingly, generating clicks behaviourally coherent enough for the algorithm to reward. Fraudulent placements produce inflated engagement; the algorithm reads that as evidence of quality, and more budget flows towards the fraud. The advertiser’s own optimisation engine becomes the distribution mechanism. The same applies to lead generation and affiliate sales, where bad actors game numbers and cost brands dearly.
The scale that this can reach was set out in September 2025 by HUMAN Security’s Satori threat intelligence team, which documented a scheme it named SlopAds: 224 mobile applications, downloaded more than 38 million times across 228 countries and territories, converting ordinary consumer devices into fraud nodes and generating a peak of 2.3 billion fraudulent bid requests a day. The commercially relevant design detail is that the applications behaved normally when downloaded organically and activated the fraud payload only for installations traced to the operators’ own campaigns. The scheme was engineered to pass inspection, not merely to perform.
The same logic exposes attribution. Loading a landing page in a hidden browser window fires the conversion pixels that trigger on page load and marks the device as exposed to the advertisement. When that real person later buys something, the fraudster claims the conversion. Return on ad spend, if built on clicks, can look excellent while delivery underneath it is substantially synthetic.
Generative AI has industrialised the supply side. Made-for-advertising exposure in the ANA benchmark rose to 1.1 per cent in Q1 2026 after holding between 0.4 and 0.6 per cent through 2025, a near doubling, with AI-generated sites built purely to harvest advertising revenue identified as an emerging subtype. Where a fake publisher site once required technical skill, thousands can now be produced in an afternoon.
The macroeconomic residue
Aggregate estimates warrant the caution owed to all industry-funded totals, and they vary widely across methodologies. An independent analysis of 105.7 billion (a mere 1 day’s worth of ad impressions bought and paid for) impressions collected through 2025 recorded a global invalid traffic rate of 20.64 per cent, implying roughly $37bn of US advertiser spend annually associated with invalid traffic. Juniper Research has projected total losses from advertising fraud to reach $172bn by 2028. The ANA identified $26.8bn of wasted programmatic spend globally in its Q2 2025 benchmark, in a quarter of measurable improvement.
The economic significance is not the headline theft. It is that a material share of the world’s marketing capital is being allocated on performance signals describing events that did not happen, in markets where that same capital funds journalism and commercial media. Claims by forensic analytics firms that clients recover between 20 and 60 per cent of programmatic spend are self-reported and not independently audited, and should be read as vendor evidence rather than a benchmark. The ANA’s own cohort data corroborates this direction.
For a chief executive, the useful question is no longer how much was stolen last year. It is a governance question with a binary answer: does the company sit in the cohort that measures its media at impression level, or the cohort that is told what happened and files the report?