A clean bill of health is not a diagnosis

The first question invites a general check. You'll be told you've picked up a virus, which is true but almost useless, because it doesn't distinguish between three conditions that look nearly identical from the outside yet call for completely different responses: who you can safely be around, what you should take, and whether this will clear up on its own or needs watching.

The second question prompts your doctor to look for something specific, using an instrument built for the purpose. A swab. A named test. It finds only what it was designed to find, which is why the question you ask determines the answer you get.

Now, treat your campaign as the patient.

One body, two verdicts

I've spent a long time now looking at digital advertising at the impression level: the individual ad event, one at a time, rather than the aggregate report. The most surprising thing I've learned isn't a number. It's this: two people can examine the same campaign and reach completely opposite conclusions.
One sees

• low fraud
  
• acceptable waste
  
• strong delivery.

The other sees

• inventory misrepresentation
  
• hidden placements
  
• attribution problems
  
• significant signal degradation.

Same campaign. Same period. Same spend. The examinations diverged because the instruments were built to detect different things.

What a single test can and cannot tell you

Most of the industry's comfort about fraud rests on IVT (invalid traffic) reporting. IVT measurement asks a precise and useful question: was the traffic behind this ad request valid, or did it show signs of being a bot, a data centre, a known traffic-based fraud scheme, or a non-human agent?

That is a good test, well designed, and I want to be careful that nothing here is read as diminishing its value.

But a specific test yields a specific result. A negative Covid swab tells you about Covid. It says nothing about your blood pressure, your liver enzymes, or the reason you've been short of breath for three weeks. Somewhere along the way, we began interpreting a low IVT percentage as a clean bill of health for the entire media buy, as though 'traffic looks valid' meant the same as 'this money bought real advertising'.

The gap between those two statements is where a great deal of value quietly disappears.

The full workup

The alternative is the examination a doctor orders when the presenting symptoms don't add up: a general check-up, together with a full blood panel. No single marker in that panel is a diagnosis. An elevated white cell count on its own means very little. Read alongside a raised inflammatory marker, a particular liver value and a patient's history, it starts to describe something.

This is the principle behind impression-level analysis. Platforms such as FouAnalytics collect up to 500 data points across all the key dimensions for a single advertising event: traffic, network, environment, publisher, device type and its efficacy, viewability and ad placement, user experience and load performance. Individually, most of those points are unremarkable. Cross-correlated, they tell a story about what actually happened.

That story is the diagnosis. Everything before it is a reading on a chart.

Where the illness originates

The mental model most of us hold is that ad fraud originates outside the advertising ecosystem: a bad actor, a bot farm, or fake traffic pushed into the pipes. An infection, in other words, contracted from the environment.

Plenty of what I see is iatrogenic – that is, harm caused by the treatment itself: the infection picked up on the ward, the reaction to a correctly prescribed drug, the complication that accompanies the procedure meant to help. Nobody set out to cause it. The machinery of care caused it anyway.

A meaningful share of what appears at the impression level works the same way. It originates within the apparatus built to deliver advertising, among parties all technically operating within it. Four examples of how that presents:

1. Inventory misrepresentation. The ad is bought as though it will appear on one property, in one environment, in front of one audience. But it is served elsewhere. The traffic can be entirely human while the placement bears no resemblance to what was purchased. Domain spoofing and app spoofing are the well-known versions; misdeclared placement type and misdeclared environment are the quieter ones.

2. Hidden and stacked placements. The impression is served, rendered and measurable. It also appears behind another ad, or outside the visible area, or in a 1x1-pixel frame, or in a place no human eye could plausibly have landed on it. The vital signs are all present. The patient never saw a thing.

3. Attribution contamination. When an impression's true placement, timing or context is misdescribed, every downstream credit assignment inherits the error. The conversion happened. The account of what caused it is wrong, and the budget moves towards whichever channel was best at claiming credit rather than the one that did the work.

4. Signal degradation. Each of the above writes a slightly misleading line into the patient's chart, which everything else in the marketing stack will read as history.
   

In all four cases, the traffic passes an IVT check, because 's this traffic valid?' was never aimed at any of them.

The better question

This is why I believe the industry has been focusing on too narrow a question. We need to shift from:

Is this traffic valid?

to:

Is this advertising event legitimate?

The second question asks whether the impression bought is the impression served; whether the described placement is the one that existed; whether a human being could plausibly have seen this ad; and whether the event record accurately reflects what happened.

Ask that, and you stop collecting a reassuring percentage and start assembling a picture of reality. There is little comfort in being the only one asking, either. Diagnostic standards work when the whole profession shares them.

Why this has become urgent

For most of the last decade, this was an efficiency argument: bad impressions waste money, so find them and stop buying them.

The argument has changed shape. Every optimisation model, every attribution framework, every AI system in the marketing stack now learns from the chart we hand it. These systems don't audit their inputs; they optimise against them. Feed a bidding algorithm a dataset in which misrepresented placements appear to perform, and it will find you more of them, faster, at greater scale, and with total confidence. A misdiagnosis, prescribed automatically, to every patient.

We have spent several years buying more intelligence to apply to our data, and comparatively little on establishing whether the data deserves it.

What I'd ask for

If you're a CMO or a media lead and want to know where you stand, here are four questions to ask your agency, your DSP (demand-side platform) partners, and your verification vendors:

1. What exactly is this report measuring, and what is it not measuring? A low IVT figure raises questions about traffic. Ask explicitly what it indicates about placement accuracy, viewability context and impression legitimacy. If the answer is 'nothing', that's fine, but now you know the shape of the gap.

2. Can we examine this at the impression level rather than in aggregate? Averages obscure cases and anomalies. Interesting findings rarely show up in a summary.

3. Can we see what caused a determination, the reasoning the data provided input to? Without the data, any determination is pure conjecture.

4. Which of these signals feed our optimisation and attribution models? A signal that is both unverified and load-bearing for automated decisions should be fixed first.
   

The biggest opportunity in marketing right now may not be finding more signals. It may be improving the quality of the signals we already have.

Which makes the first job asking better questions.