Thread by @Biasedobsrvr

Wanted to follow up @CharlesWeill tweet to explain a bit more about AVD (Average View Duration) multiplied by CTR (Impressions Click-Through Rate) (CTR*AVD), what it means and why it's one the most important metrics to Youtube's algorithm.

Thread 👇

2/13 To start, CTR*AVD is essentially Expected Watch Time Per Impression. Google outlines this in a 2016 paper linked at the end stating:

"Our final ranking objective...is generally a simple function of expected watch time per impression."

Ty to @dabidoYT for showing me this!

3/13 The result of AVD*CTR is usually btwn 0 and 60 second at the end of a videos life. (some variability) It'll look a bit like this
An important point is CTR and AVD tend to be inversely related. Table illustrates CTR changes as AVD increases by 1 minute, balancing AVD*CTR.

4/13 Further illustrating the above point, this is a graph of 2400 videos (gaming niche) for CTR and AVD. You can see as AVD gets longer CTR drops.
This is because Expected Watch Time balances out, allowing for a lower CTR because 1 view is more valuable (more watch time).

5/13 Now for the good part! Taking those 2400 videos, here's a graph of their views by CTR*AVD.
You can see clustering starting around 20 seconds to ~45 seconds. For the gaming videos I have, this tends to be the sweet spot for this niche.

For other niches it differs.

6/13 So what does this all mean?

1) CTR can and will usually be lower for videos with higher AVD, because you'll be getting more impressions and a view means more

2) YT seems to have a sweet spot below 60 seconds Expected Watch Time, meaning if you're above you're doing great!

7/13 3) CTR*AVD can be a very strong indicator of a videos success. A small teaser of an analysis to come. This is a graph of 50 videos' with First Hour CTR*AVD and their current views (end of life). You see a breaking point where if you pass it views seem to take off.

8/13 4) As everyone says, context and niches are important.

The purpose of this thread is to show people the relationship between CTR*AVD, help explain why sometimes CTR might be low but is actually a good thing, and hopefully offer additional context to a video's performance

9/13 Lastly the paper and some log transformations of the data illustrating the relationship better. Thanks to @CharlesWeill for the shoutout and help on this analysis!

The 2016 paper is 'Deep Neural Networks for YouTube Recommendations', can't seem to link it.

10/13 Per Charles rec, logging the views and Expected Watch Time we get something much more normal and linear.

11/13 Some descriptive stats about the data bucketing the Expected Watch Time into 10 second buckets to see frequency of where videos show up.

12/13 Lastly, everyone loves a good box and whiskers plot. Helps to see variability in CTR based on AVD a bit more - ideal would be seeing videos stat at some point in time (X days after publish).

13/13 I hope this was helpful to some in gaining additional insight into the YT algorithm and why some of the metrics behave certain ways!

There's hundreds of other variables, so this is just a small but important piece of the puzzle.

Follow for more analyses like this one!

Oh now I can link the paper! 😡

Well anyhues - here's the 2016 paper about Expected Watch Time. Section 4 goes into it more. research.google/pubs/pub45530/