30% of our traffic wasn't people.

What we measured

A split test, read two ways against different denominators. Against raw recorded sessions, phantom traffic ran about 30%. A shorter forensic window on the same data came back higher. A cut against paid clicks specifically came back lower.

What the numbers showed

There is no one correct phantom rate. Every version of the number depends on which denominator you're using and how tightly you define 'engaged.' Quoting a phantom rate without saying the denominator behind it is close to meaningless, because two honest people measuring the same traffic can land on different numbers without either one being wrong.

Why it failed

Before this was filtered properly, recorded page loads on a separate page showed a hard spike at 23 seconds, 274 loads, zero scroll, zero engaged time. Read as people, that looked like a page in trouble, and real time went into chasing a performance problem that didn't exist. The page was fine. The measurement was full of ghosts: prefetchers and scanners registering as sessions without ever being a person.

What we changed

Phantoms get filtered at read time, never at collection. Absence is the tell: no page drawn, no scroll, no active time means it wasn't a person, regardless of the raw count. A session only enters any denominator once it shows a real sign of life: a scroll past a quarter, ten seconds of active time, or a click. Every report now reads through one saved, filtered view, so two people reading the same week get the same number.

How to spot it in your own account

If a page's engagement metrics look suddenly bad, check for a spike of sessions with zero scroll and zero active time before assuming the page itself has a problem. Require a real engagement signal before counting a session at all, and always state the denominator when you quote a phantom rate. Published estimates for bot traffic run as high as half of all web traffic on other networks; use that only to justify scepticism, never as your own figure.