A portion of the traffic recorded against any campaign was not generated by a person. The reporting consequences extend well past the money spent on those impressions.
Not all automated traffic is fraudulent
Search crawlers, uptime monitors, security scanners, link preview fetchers and internal testing tools all generate page loads that look like sessions.
These are legitimate and unavoidable, but they enter analytics as visits with no engagement, diluting averages across every metric computed per session.
Their effect is largest on low-traffic pages, where a handful of automated requests can visibly move an average that decisions are based on.
Deliberate invalid traffic is built to look human
Fraudulent traffic is designed to pass filters, so it produces mouse movement, plausible timing patterns and occasional clicks rather than obvious machine behavior.
It concentrates where detection is weakest and where payment is easiest to obtain, which historically means long-tail inventory and environments with limited measurement.
Detection depends on pattern recognition across large volumes rather than on inspecting any single session, which is why individual advertisers rarely spot it directly.
The real cost is model corruption
Automated bidding and platform optimization learn from the outcomes they observe, and invalid traffic that never converts teaches them the wrong lesson about a placement.
Worse, some invalid traffic does trigger conversion events, especially where the conversion is a form submission or a page view rather than a payment.
The system then bids up toward the source of those fake conversions, and the campaign steadily reallocates budget toward its least valuable inventory.
Filtering happens in several places at once
Buying platforms remove known invalid traffic before billing, analytics tools apply their own exclusion lists, and verification vendors add a further layer.
Because each filters differently and at a different moment, no two systems report the same totals, and reconciling them exactly is not possible.
The practical approach is to choose one system as the reference for decisions and use the others to detect movement rather than to establish truth.
Sudden changes matter more than absolute levels
Every account carries a baseline of automated traffic, and the number itself says little without a comparison point.
A sharp rise in sessions with no corresponding change in engagement, concentrated in a specific source or geography, is a far stronger signal than any single figure.
Monitoring the ratio between sessions and meaningful actions catches most of these events early, before an optimization system has spent weeks learning from the noise.