Pas🪽

YouTube just released a 43-minute video discussing Dynamic Thumbnails, A/B testing, and YouTube’s algorithm. Here are the 7 most interesting insights so you don't need to watch it yourself: 1. Thumbnail A/B Testing YouTube doubles down on not showing CTR in A/B tests, as they don't want creators to be optimizing for the click, as it may hurt the overall watch experience of the video. 2. Shorts A/B Testing YouTube is going to allow us to pick 2 to 3 Shorts to test against each other. They will then show each of them to a different segment of the audience. You will then get retention data like watch time, but also the retention graphs for those videos. From a past interview, we already learned that you'll have up to 7 days to select which version you want. (Even though this wasn't mentioned in this interview, I wanted to add this here because it's an interesting insight that I don't think a lot of people are aware of.) 3. Dynamic Thumbnails When touching on dynamic thumbnails, Todd highlights that they are designing their system to dynamically adjust who and how many people get to see which thumbnail, which is different from the current thumbnail A/B testing (where it's pretty much an even split randomly across who sees it). He says that the way it works is that there will be some randomness introduced to make sure they learn new things, but ultimately, they want to start shifting traffic towards the direction that is working, or, in this case, the thumbnail that is working. If a particular thumbnail works well, YouTube is going to put more impressions towards that thumbnail, and by doing so, YouTube reduces the risk of messing up a thumbnail test. Over time, the test will get better at learning from failures and reducing impressions on thumbnails that fail by leaning more towards what's already working. On top of this, dynamic thumbnails will also consider audience segments and see which thumbnails work better for which segments. They will use factors such as location and whether somebody has watched the channel before. These factors are something that Todd refers to as segment features. But the conversation also kind of hinted towards the fact that it's not as rigorous as we would want it to be (for example, segment features it's based on interest). It seems to be hinting that most of the factors are going to be towards the more measurable things that YouTube already has. 4. Back Catalog Optimization In the interview, they briefly touch on back catalog optimization, but they don't go too deep into this one. We already know from a previous announcement that this has everything to do with their Ask Studio features, which allow creators to have AI involved in optimizing their back catalog for them by creating new thumbnails, etc. 5. Hyperpersonalization In the interview, they touch on another thing that they also touched on before: YouTube is actually becoming a hyper-personalized platform where different things are weighted differently for different people in the algorithm. What this means in simple terms is that, for one viewer, it may be more important how clickable a video is, while for another, it's much more important how satisfying a video is. This is something that YouTube has recently been pushing a lot more in their interviews. To translate this even further, this means that there's no single unifying metric that people should be looking at to understand what worked. Other than views ofcourse. 6. Custom Feeds Later in the video, Todd talks about custom feeds and how the algorithm will work for those custom feeds. As we currently already know a lot about how the home feed works and the other recommendations, he says that it uses similar technology to what we use for the regular home feed, but it's a lot more language-based. What this means is that when a user sets up the custom feeds, they type out or use their voice to say what they're looking for, and it can be pretty abstract. The algorithm will do an interpretation by using language modeling to understand and generalize into a vibe of what the person may actually be looking for. It's important to know here (and he mentions this as well) that this isn't just strictly keyword matching, but it is more about the interpretation of what the user is looking for. And if the user wants to see more popular stuff in their custom feeds, they can highlight this one, but it will also make it easier to match viewers with less popular stuff. 7. New Creator Discovery At some point, the conversation shifts towards how new creators can get discovered today. The first thing Todd highlights is that new creators should stop focusing on metrics, and they should focus on finding purpose and figuring out what kind of content they want to make. This hints towards the idea that metrics in Studio are not as valuable once you start out as a creator, which I do agree with. Then Todd talks about discoverability for new creators. He says that parts of the recommendation system are looking for videos that they don't know much about because maybe the channel they're uploaded to has zero subscribers. There's not really anybody who has watched the channel before. And they know that. In order to give these videos an on-ramp to success on YouTube, they need to give these videos a chance. They have something called exploration. An exploration is when viewers show up on the home feed. Every once in a while, they're going to show that person a video that YouTube themselves don't really know much about. Then, when this person likes the video, it's going to show it to more people like this person. This is how on-ramps work. YouTube is always trying to balance exploring new stuff that they don't know much about and serving stuff that has already been successful for years. Now, this breakdown was an interpretation mixed with exact things that they have referred to, but I wanted to highlight this one, as there were some interesting things in this video. I wanted to keep everybody up to date with what's going on.