Thoughts on whether retention, a classic metric of the YouTube algorithm, matters or not. And a common misinterpretation of it.

YouTube definitely cares about retention. They’ve stated this publicly (attached from https://t.co/KRpTweoI0M ). What YouTube hasn’t publicly discussed, and I think is unfortunate because people just make the conclusion “retention doesn’t matter”, is that retention itself is an interdependent metric. People have to realise that retention is both a metric that is itself observed, but influences other metrics (impressions/views) that change the metric. A simple example that will often make YouTubers question retention is why they get low views with high retention. In that specific case, YouTube’s algo has somehow determined that despite the “high” retention for the small pool of users it was served to, the videos overall won’t lead to a predicted good expected watchtime per impression. Maybe the topic is too niche, or maybe the CTR is low, or maybe there’s too much other competition. I think retention as an objective is usually something that’s worth optimising for. However, interpretation is really complex, and there’s smarter brains than I that I know would have much more robust hard data on it. So, the correct conclusion isn’t “retention doesn’t matter”, exactly. I believe the actual correct conclusion is: the absolute AVP or AVD has a wide variety of factors that influence it, and because of this, the reported retention % doesn’t matter even if retention itself does. In MrBeast’s specific case, the sheer number of impressions means that he’s reaching an extremely broad audience, so you can’t compare his video - with literal millions upon millions of views - to someone who e.g. gets 80% AVD on a 10K view video. Otherwise, what even would be the point of YouTube justifying spending so much money to generate graphs for the millions of videos that creators upload to the platform? It would be a bizarre business decision too, and those small expenditures in software usage would add up to an extraordinary amount of unnecessary compute over time. Just my thoughts :) others have differing opinions on this subject. Open to criticism by anyone with data.