Impressions: Every exploit to “break the algorithm” targets this baffling metric

on the whole youtube creators have basically no idea how the youtube algorithm actually works i'm part of a wonderful slack workspace full of educational youtubers but it frequently sees comments like these the suggested videos algorithm is the least understood it started for no reason that i can see approximately 100 times our normal view count i expect it to drop off abruptly with no explanation we are clearly toying with forces we do not understand however it's generally agreed upon that you want your video to have impressions how many people have actually seen a thumbnail how many people click on it and how many people watch it come second if they don't see it they never even get the chance to watch a video impressions are an absolutely critical statistic that the public never gets to see basically every time that i log into youtube as a viewer youtube knows that i'm going to watch something the server is 100 aware that it has a captive audience in me and it's so friendly is to give me eight options on the home page and more if i scroll as to how exactly i want to be held captive by its platform every time you load up the home page and every time that you see a whole bunch of video thumbnails next to a video that's playing there is an impression recorded for each of those videos and that is the key statistic that every algorithm breaking exploit is actually trying to maximize i've always thought about these statistics in terms of the expectation retention time or the ctr watch time product i'm sort of making these names up as i go basically you take the click-through rate the fraction of people that click on a thumbnail when they see it and you multiply that by the number of minutes or seconds that someone on average spends watching that video and then bam you have the number of additional seconds that youtube expects to retain a particular viewer just by showing them a picture because statistics one would imagine that youtube wanting to retain viewers for as long as possible would only show you thumbnails of videos with a really high ctr watch time product and sometimes they feed you little tidbits like this to give you the impression that there's a method to the madness it just seems logical that that would be how they would want to run things but perplexingly that's not always the case the key is to somehow get youtube to push your content into the view of complete random strangers when a video goes big youtube shows the thumbnail to a lot of people but most of them don't click on it because they're outside of the original intended audience for that video a few months ago i published a video about water making a noise when it freezes i thought it was a fun video and i thought it had a suitably clickable thumbnail and apparently youtube did too because it got 260 000 impressions and 9 200 views in the first 17 days which is pretty much unheard of for my channel youtube liked showing this video to people but it had reached so far outside of my audience that the watch time ctr product currently sits at just four and a quarter seconds every time that youtube shows this thumbnail to someone they retain that person on the platform for an additional four and a quarter seconds it doesn't really seem like it's worth it of course what's actually probably going on here is that youtube flagged a video that my subscribers enjoyed and said you are an excellent test subject and cast a very very large net looking for prospective new subscribers the algorithm just isn't very efficient with this search so clearly there is a lot more data under the hood that creators let alone the general viewership has no access to and no idea how it's working which is why i thought it would be fun to crack out the pythoning skills and publish a video about impressions that literally in the thumbnail tells you how many people youtube is showing that thumbnail to right now and how many of those people are actually clicking on the video i think it adds some camaraderie that's been missing in terms of watching videos online with a whole bunch of random internet strangers and i'm really looking forward to your reactions to this video as long as my code keeps working because my code is really fragile i believe that the first big self-updating video was published by tom scott and in that video the title would update with the view count live and the view count was substantial but in that video tom explicitly warned against building a scraper to do your data collection the proper ways to use an api which makes all the sense in the world unfortunately impressions are such a neglected statistic that the youtube data api and at least to my ability to look for it the youtube analytics api neither of them can actually get you bulk impressions data on a video only weird subsets that i didn't care about so i did exactly what tom said you should never do and i built a scraper after all if i can open a website and press refresh all day long why can't a robot my first attempt was using a python library called selenium that is intended for browser automation you go to youtube you log in you navigate to the analytics page but then we've already encountered a problem because when you try to log in the google authenticator looks at your robot browser and says hey that looks kind of sus and then closes the door and locks it a few failed workarounds later i settled on probably the most grotesque option available login to my youtube analytics page by hand then start a script that takes a screenshot once a minute locates the refresh button clicks it waits 10 seconds locates the impression counter double clicks it to select the text presses control c then dumps the keyboard contents out into a local variable in python on the screen it just looks like the cursor is teleporting around really quickly it's exhausting to watch this kind of data collection is also unreasonably fragile if i minimize this window or even cover it up the script starts throwing errors because it can't find the refresh button anymore so i decided to give my refresh script its own little magical world i spun up a virtual machine installed python in the same browser with the same theme as on my work computer and then i moved my active folder of automation code over into the matrix where it could refresh the youtube analytics page to its heart's content all day every day you just have to breathe on this system to make it stop working but at least by putting it inside of a vm i've made it slightly more robust so hopefully it keeps running until no one's watching this video anymore which might not be very long hope you enjoyed it [Music] you

Channel: AlphaPhoenix

Views: 131,175

Keywords:

Id: iGW54Rr87Rg

Length: 7min 28sec (448 seconds)

Published: Fri Feb 19 2021