I Automated YouTube Video Triage, Then Found Out I Was Optimizing the Wrong Thing

I have a tool that watches my two YouTube channels and tells me which videos are underperforming. It compares every video against the channel’s own median, sorts the losers into buckets, and for each one it writes a fresh title, description, and tags in my voice. I approve the suggestion, it pushes the update to YouTube, and 28 days later it grades whether the fix worked.

I was proud of it. It felt like the responsible, data-driven thing a serious creator does with a back catalog. Then I stopped and asked the question I should have asked before writing a single line: is triaging old videos this way even the right thing to do?

The answer turned out to be no. The diagnosis was a guess I’d dressed up as a fact, the fix was the same fix no matter what the diagnosis said, and the success metric was measuring noise. Worse than any of that, the whole activity is close to the least useful thing a small channel can spend time on.

What I built

The tool has two halves.

The first half is diagnosis. For each video it looks at click-through rate and audience retention, compares them to the channel median, and picks one of four labels. Low clicks but strong retention means the packaging is the bottleneck. High clicks but weak retention means the packaging overpromised. Low views with good retention means the video is under-exposed. High views with early drop-off means viewers are bailing in the intro.

The second half is the fix. No matter which label fired, the tool did the same thing: rewrite the title, description, and tags. Approve, push, wait, measure.

If you’ve read anything about YouTube, that quadrant logic sounds right. That was the problem: it only sounds right.

The question I skipped

Before I spent any more time validating my own reasoning, I handed the whole design to OpenAI’s Codex, running GPT-5.5 through codex exec on the command line. Different company from the one I normally build with, which was the point. I told it to be adversarial. Find what’s wrong. Don’t be agreeable.

It didn’t soften anything. Its opening line was that I was treating an old-video problem as a metadata problem, when the real decision was something else entirely: revive, redistribute, remake, or abandon. Rewriting the title was one branch of a tree I’d collapsed into a single action.

That’s a strong opinion though, and a strong opinion isn’t evidence. So I ran the whole thing down again against sources: YouTube’s own creator guidance, its help documentation, and the people who study the platform seriously. I wanted to know which of my beliefs YouTube itself confirmed, which were merely popular, and which were flat-out folklore.

Most of what I believed didn’t survive that.

The myths I had built into a tool

Editing a title or thumbnail doesn’t restart the algorithm. This is the belief that quietly justified the entire feature. Change the packaging, wake the algorithm up, get a fresh look. YouTube’s own position is that the edit itself does nothing. It doesn’t trigger new impressions and it doesn’t reset momentum. The only thing that moves is what viewers do with the packaging you changed. I was afraid an edit might reset a video’s standing, and at the same time hoping it could revive one. Neither is true.

Tags are close to worthless. My tool spent an AI pass generating tags for every video. YouTube treats tags as a very low-weight signal, useful mostly for spelling and disambiguation. I was paying for AI calls to produce something YouTube barely reads.

Deleting weak videos doesn’t clean your channel. There’s a popular idea that a back catalog of duds drags down your new work, so you should prune. YouTube judges videos individually, not by channel average. Old duds don’t suppress new uploads. And deleting a video severs your channel’s connection to everyone who ever watched it, which is why YouTube explicitly advises against it. If a video genuinely has to go, you unlist it. You almost never delete it.

The real fix for a weak intro isn’t a new title. My tool would look at a video where viewers bail in the first thirty seconds and suggest a better title. A better title can’t fix an intro. The actual fix is to trim the intro, and YouTube’s built-in trim editor does exactly that while preserving the video’s ID, its views, and its comments. That branch existed the whole time. My tool just didn’t know how to reach for it.

The best move is often not to touch the old video at all. When a video is barely being shown, the lever isn’t its metadata, it’s distribution: a Short that links to it, an end screen from a stronger recent video, a playlist that carries viewers into it. Watch time flowing from your best content is the strongest signal you can send, and none of it involves rewriting a description.

The part that stung

Here’s the finding that reframed everything. For a small channel, going through the back catalog and optimizing old videos is close to the lowest-value thing you can do.

The leverage on a small channel is in the next video: the topic you pick and the packaging you wrap it in. The advice to lovingly revive your old uploads mostly comes from tool blogs with suspiciously specific case studies and no primary source. The only old videos worth reopening are the evergreen, search-driven ones that answer a question people still type into the search bar today. Everything else is better left alone while you go make something new.

So I’d built an elaborate machine to do the thing the evidence says matters least.

The measurement was lying to me too

My tool waited 28 days, compared click-through rate before and after, and called anything above a five percent lift a win. That sounds rigorous until you look at the sample it was working from. On a channel where a video might collect a few dozen impressions in a month, one extra click clears a five percent lift by itself. I was measuring noise and writing it down as a result.

There’s a worse problem underneath that one. YouTube’s own guidance is that click-through rate can fall while a video is doing well, because a video shown to far more people converts a smaller slice of them. So the number could read as a win on a video going nowhere, and as a loss on the one video that was finally breaking out.

A before-and-after on a small channel doesn’t tell you whether your change worked. It tells you what the algorithm happened to do that month.

What good actually looks like

The rebuild isn’t four labels and one fix. It’s a router that sends each situation to the move that fits it. Packaging problem with enough traffic to judge it: generate title and thumbnail options and run YouTube’s own built-in test rather than guessing. Weak intro: suggest a trim. Under-exposed: suggest which of your stronger videos should point at it. Outdated but valuable topic: that’s not a fix, that’s your next video. Genuinely dead and browse-dependent: leave it alone.

That part is built and running. The one piece I haven’t done yet is the gate that should sit in front of all of it: is this topic still worth winning? If nobody’s searching for it, no title in the world will save it. That needs search-demand data I don’t have wired up yet, so for now it’s still me making that call by hand.

The lesson worth keeping

The tool was wrong. The process that caught it is the part worth keeping, and I’d run it again on anything I build.

Build the thing. Then, before you trust it, hand it to something adversarial and ask not “is this implemented well” but “is this even solving the right problem.” I used Codex precisely because it didn’t share the assumptions I’d been building on, and it found the load-bearing one I couldn’t see from the inside. Then I checked its claims against YouTube’s own documentation instead of taking its word for it.

Here’s the uncomfortable part. The most useful thing I did for my channels that week was admit the tool I was excited about was pointed at the wrong target. Then I rebuilt it around the router, dropped the tag generation, and made the whole thing advisory instead of letting it push updates to YouTube. Same afternoon. The rewrite was smaller than the tool it replaced.