Digital advertising platforms have become extraordinarily good at optimisation. The problem is that they optimise for the outcome we give them, even when that outcome is a poor proxy for what the business actually needs. In an age of AI-driven media buying, measurement is no longer simply a reporting discipline. It is risk control.
One of the most dangerous moments in performance marketing is when a campaign begins to look almost too good to be true.
Cost per lead falls sharply. Conversion rates climb. Volumes accelerate. The algorithm finds more of whatever you asked it to find, and the dashboard starts producing the type of numbers that make everyone in the meeting happy.
However, sometimes the correct response to extraordinary performance is not celebration; it is suspicion.
We recently encountered precisely this situation at Offernet while analysing a pet insurance campaign. A very large proportion of the campaign’s leads began arriving through third-party app placements at a fraction of the cost we would normally expect from Facebook Ad inventory. Some placements were generating apparent conversion rates of between 15% and 25%.
On the dashboard, the online data looked exceptional, but the Offernet data analysts started connecting online data to offline results, and the offline performance metrics painted a very different picture.
When prospective customers were contacted, some told us they did not own a dog or cat, even though the submitted forms indicated otherwise. That immediately raised a more important question than cost per lead: Where is the disconnect?
Our subsequent investigation identified several very different types of app behaviour. In one category, users were promised PayPal payments or credits as rewards for interacting with advertisements and completing forms. These people were not necessarily interested in pet insurance; they were there to receive the economic motivation, the reward.
Another category included entertainment apps where users watched short, often AI generated, episodic content. Advertising interruptions formed part of the experience, with users encouraged to interact or complete actions to continue watching past each cliffhanger.
The most concerning pattern involved an app whose stated purpose was to identify and block spam calls. The irony was difficult to miss. Much of the anomalous lead activity linked to this app occurred between roughly 11pm and 6am, when users’ Android devices were likely to be inactive or charging. Our investigation found activity consistent with ads being loaded and conversion actions occurring in the background, using personal information the app already held about the user.
In other words, an app designed to protect people from unwanted calls appeared, in these instances, to contribute to generating leads that could result in exactly those unwanted calls. Beyond the immediate fraud concern, this raises a much bigger question about consent, app permissions and the unintended consequences of the mobile advertising ecosystem, one worth exploring in its own right.
We have reported the relevant evidence to Meta and are engaging with them on the matter. We are deliberately not identifying individual apps while that process is underway.
This is also not an argument that Meta Audience Network is inherently problematic. Meta describes Audience Network as a third-party mobile app inventory that includes formats such as rewarded video and interstitial advertising, and says it employs publisher reviews, fraud prevention systems and advertiser placement controls.
The lesson is much broader: No advertising ecosystem should be treated as self validating.
When the metric becomes the target
Alex Schultz captures the underlying problem particularly well in Click Here: The Art and Science of Digital Marketing and Advertising. His measurement framework begins with one clearly defined North Star metric, supported by telemetry and guardrail metrics. He also warns that metrics, once converted into targets, can be 'co-opted, gamed and hacked'.
This is essentially Goodhart’s Law applied to digital marketing. Schultz’s interpretation of Goodhart’s Law is simple: when a metric becomes the target, people and systems begin optimising towards the metric itself rather than necessarily towards the underlying objective. He provides examples ranging from meaningless activity being counted as users to apparently valuable revenue later being reversed through fraudulent chargebacks.
AI makes this problem more important, not less so.
An advertising algorithm does not inherently understand that the business wants profitable, insurable customers who genuinely requested contact. It understands the signals we provide it with.
If we tell the platform that a completed lead form represents success, then a completed lead form becomes the optimisation objective. If one group of placements can generate those form completions significantly more inexpensively than another, the system has every reason to allocate more budget towards them.
From the algorithm’s perspective, it is doing its job.
The mistake may well be ours.
Dr Augustine Fou, who has extensively researched digital advertising fraud, has argued that the economics of programmatic advertising create powerful incentives for artificial inventory and behaviour, because synthetic traffic can be produced at an enormous scale and at very low cost. He has also argued that unusually inexpensive media and apparently exceptional performance should prompt marketers to examine their underlying traffic and business outcomes rather than simply accept the platform's report.
Measurement must continue after the lead
This is why we believe one of performance marketing's most important questions consists of only four words:
What happened after that?
An advert received an impression. What happened after that?
Someone clicked. What happened after that?
They completed a form. What happened after that?
Were they contactable? Were they genuinely interested? Were they eligible? Did they receive a quote? Did they purchase? Did the sale remain valid?
At Offernet, this is the purpose behind our Touchpoint methodology. We map the customer journey beyond the advertising platform and connect upstream media activity with downstream business outcomes. Our existing measurement architecture is designed to combine digital activity with CRM and offline feedback, enabling optimisation to move beyond impressions, clicks, and submitted forms and towards commercial results.
Schultz makes essentially the same argument when discussing paid social: first decide whether the actual objective is leads, registrations, purchases or revenue, and then feed properly tagged conversion and value information back into the advertising system so that it can optimise towards something meaningful.
The speed of that feedback loop matters enormously.
If poor-quality or fraudulent leads are discovered three months later, the campaign may already have spent a substantial budget teaching the algorithm to find more of the wrong behaviour. If the anomaly is identified within hours or days, placements can be investigated, signals adjusted, and losses contained.
That is why measurement can no longer be something marketing teams build after a campaign launches.
Measurement is part of the campaign architecture.
The advertising platform has an optimisation system. Publishers have economic incentives. So do fraudsters. The advertiser, therefore, needs an independent measurement capability of its own.
Otherwise, it is entirely possible that the campaign metric improves while the business outcome deteriorates.
And that is the uncomfortable reality behind the cheapest lead in the dashboard: It may turn out to be the most expensive lead you ever buy.
