What is ad stacking?

How it works

One placement, several ads rendered in the same position, stacked in z-order. The user sees the top one. Every ad in the stack fires its impression tracker, because each one did technically render.

The publisher bills for every impression in the stack. The advertisers below the top pay for exposure that was physically impossible to see. A stack of ten multiplies the fraudulent revenue from a single placement by ten.

A related variant is pixel stuffing, where an ad is rendered in a frame one pixel square — technically present, effectively invisible.

Why it worked

Because the industry counted served impressions rather than viewable ones. If the measurement is 'was the creative returned and rendered', a stacked ad satisfies it completely.

That is the underlying lesson: fraud follows the definition of the metric. Any metric that can be satisfied without delivering the value it represents will eventually be satisfied that way at scale.

How it is stopped

• Viewability measurement, which requires a proportion of pixels visible for a minimum duration and cannot be satisfied by a buried ad.
• Buying on viewable impressions rather than served ones, which removes the economic incentive entirely.
• Verification vendors that detect overlapping placements and abnormal render conditions.
• Anomaly analysis — many impressions from one placement with essentially no clicks.

The AI parallel

There is no direct equivalent, because a generated answer has no z-order and no hidden layers to occupy. An answer insertion either appeared in the response or it did not.

The transferable lesson is about metric design rather than about stacking. The industry spent years paying for served impressions because that was what could be counted, and fraud grew to fit the gap between the metric and the value. Defining the AI-era unit around what actually reaches the user — rather than around what is easiest to log — is how that mistake is avoided a second time.