I would imagine it's simply a temporal comparison of pixel colors at predetermined locations, similar to how the Shazam algorithm[1] works? You'd just need to analyze enough pixels to reduce "collisions", coupled with the temporal aspect!

[1] https://www.ee.columbia.edu/~dpwe/papers/Wang03-shazam.pdf

It's not encoding based - its frame based, basically bitmap data. It has to be to work across the whole video delivery pipeline so it's fairly fuzzy but also accurate.

They need a source to compare too so when we worked on it masters were being sent from the network to the sync technology group. So they had source data for comparison on the first broadcast of a show.

Outside of latency there's no reason they can't match against broadcast content off cable. For user tracking they can just log the fingerprint data and compare it later to source data for analytics so this works fine.