Further to the above, it sounds like a lot of the matching is...

Further to the above, it sounds like a lot of the matching is done on audio only using www.audiblemagic.com.

https://www.audiblemagic.com/wp-content/uploads/2015/04/AM_overview_datasheet_150406.pdf

In the quora link (it's quora so who knows how accurate it is) it has the following paragraph which indicates they do some sort of frame indexing too?

1. Frames from a video (most likely minus blank screens and predominately text-based frames) a sample of a section of the video is taken. The frames are being analysed when the video is being converted server side, so this isn't expensive or a separate operation. A hash is created from those sampled frames from the video. The hash would most likely resemble the one created above, a series of jumbled letters and characters that are unique to that content.