How the optimizer learns from real performance

After Autoretto publishes a video, the work isn't over. The platform doesn't just move to the next item on the schedule. It starts a different job. It watches what happens with that video on YouTube. It collects performance numbers from YouTube's analytics. Views, click-through rate, retention, and watch time. These four metrics become the raw material for the next release.

Each metric tells a different part of the story. Views show how far the video spread. Click-through rate is the percentage of people who saw the thumbnail and decided to click. Retention shows how many people stayed past the first few seconds and how the attention drops over time. Watch time is the total number of minutes viewers spent with the video. Together, they give a decent picture of what worked and what didn't.

But raw numbers are tricky to read. A video with 1,000 views looks small next to a viral hit, but it might be a massive success for a channel that usually gets 200. So Autoretto uses a z-score model. The z-score measures how far a result is from the average, in units of standard deviation. It's a way to say, 'this result is unusually good' or 'this is unusually bad' for that specific channel.

For example, say the average click-through rate for your channel is three percent. The standard deviation is half a percent. A new video gets four percent. That's two standard deviations above the mean. The z-score is two. The optimizer sees that as a strong positive signal. A z-score of zero means the result is exactly average. Negative z-scores mean below average. The model uses these numbers to guide changes.

The optimizer takes those z-scores and turns them into nudges. If a certain way of phrasing titles gives a high click-through z-score, the next brief leans toward that phrasing. If a specific style of intro makes retention fall off early, the brief asks for a different intro. The nudge might adjust the style prompt for the music, the subject for the artwork, or the angle for the title. It won't rewrite the rules you set. It's a gentle push in a likely direction, not a command.

Now here's the important part. Creator-set rules are the fence the optimizer works inside. If you say every video must include a certain phrase in the title, that phrase stays. If you set a genre, it's locked. If you ban specific themes, they're banned. The optimizer can suggest a dozen things, but it cannot touch anything you've marked as fixed. That's not a bug. It's the design. You stay in control.

The loop repeats for each release. After a video has been up long enough to gather real data, the z-scores update. Then the next brief gets written with the new knowledge. The model doesn't overreact to one result. It looks for patterns over multiple videos. Over time, the channel's work shifts toward what the audience actually rewards. And because your rules never bend, the shift stays within the identity you chose.