In one line
Cannibalization happens when a paid ad — often brand search or retargeting, aimed at people who already know you — takes credit for conversions that would have happened organically anyway.
Why it's risky
Performance reports credit every conversion to the ad that touched it last. If some of those users would have converted organically regardless, not subtracting that share makes ad efficiency look better than it really is.
How to confirm it
Observation alone can't prove cannibalization exists. A holdout test — randomly turning the ad off for part of your audience and comparing total conversions — is the real way to confirm it.
Go deeper
Signals that suggest cannibalization, and how to confirm it, are covered in Diagnosing Cannibalization.