How to detect ad fraud in your paid media
The signals that separate automated traffic from real buyers, and a practical order of operations for checking your own campaigns.
By T.J. Muro
Published
Updated
Short answer: You detect ad fraud by comparing what a click claims to be against what it actually did. Real buyers leave a consistent trail — a plausible device, a plausible location, time on page, scroll, a second page, sometimes a form. Automated traffic clicks and leaves, or behaves identically thousands of times in a row.
The signals that give it away
- Zero-duration sessions at scale. A click with no measurable session is the cleanest tell.
- Impossible geography. Clicks from regions you do not target, or a data-center address rather than a consumer connection.
- Repetition. Identical viewport, identical path, identical timing across hundreds of sessions.
- Click timing clusters. Traffic arriving in even intervals, or spiking at hours your market is asleep.
- Conversion-shaped noise. Form fills with disposable addresses, no engagement history, and no downstream activity.
- Placement concentration. One app, site, or placement producing a disproportionate share of clicks and none of the revenue.
Order of operations
- Segment by placement and publisher. Fraud concentrates. Sort by spend and look at the tail.
- Compare clicks to sessions. A large gap between platform-reported clicks and analytics sessions is the first hard number worth chasing.
- Check engagement depth by source. Time on page, scroll depth, and second-page rate per campaign.
- Look at hour-of-day and day-of-week curves. Human demand has a shape. Automation usually does not.
- Trace conversions downstream. A lead that never opens an email, never books, never buys is a cost, not a conversion.
- Exclude, then measure. Block the worst placements and watch whether cost per real outcome improves. If nothing changes, you removed the wrong thing.
What good looks like
After exclusions, you should see fewer clicks, similar or better qualified pipeline, and a cost per booked outcome that moves in the right direction. Volume dropping while outcomes hold is the result you want.
Where disputes fit
Networks refund invalid clicks when you present evidence in their format: the affected date range, the placements, the click identifiers, and the behavior that made them invalid. Keep the record as you go — reconstructing it after the quarter closes rarely works.