An Automated YouTube Channel

Why YouTube + AI

YouTube is the one video platform where a good piece keeps earning for years, because it is searched as well as scrolled. The cost is that a decent video takes hours. AI compresses the parts that are mechanical — research, drafting, narration, assembly, packaging — so a consistent cadence stops depending on having a free weekend.

Everything reduces to one number

Before any tactic, understand what the system is actually doing, because almost every piece of YouTube advice is downstream of it.

The platform earns from time spent watching. So it promotes videos that keep people watching and quietly stops promoting videos that do not. Every other metric — click rate, subscribers, likes — matters only as a predictor of that. This is why the two commonest strategies fail in opposite directions: a video nobody clicks never gets tested, and a video everybody clicks and abandons gets tested once and dropped.

The practical consequence is an order of priority that most creators have backwards:

1. Does anyone want this? The topic. No amount of craft rescues a video nobody was looking for.
2. Do they stay? Retention, especially the first thirty seconds.
3. Do they click? Title and thumbnail.
4. Everything else.

AI helps most with the fourth, some with the first and third, and least with the second — which is the one that decides the outcome. Knowing that stops you optimising the cheap parts and wondering why nothing moves.

Write the title and thumbnail before you make the video

This inverts how most people work and it is the highest-return change available.

If you cannot write a title and describe a thumbnail that would make you click, the video does not have a clear enough idea yet — and that is far cheaper to discover before you spend six hours producing it. Packaging first also keeps the video honest: you have committed to a promise, and the edit's job becomes delivering it rather than wandering.

Concretely: write ten titles, pick two, sketch the thumbnail concept for each, and only then outline. Generating the ten is a good use of a model. Choosing between them is not, and the criteria are in the thumbnails guide.

The first thirty seconds, structurally

The steepest part of every retention curve is at the start, and it is the only part you can reliably fix.

• Deliver on the title immediately. The viewer clicked on a promise; the first sentence should confirm they are in the right place.
• No introduction. No channel name, no "before we get started", no logo sting. These exist because television had them; on YouTube they are where people leave.
• Say what they will be able to do by the end, specifically, in one sentence.
• Show something. A finished result, a screen, the thing working. Talking heads for thirty seconds is the weakest possible opening.
• Move the subscribe ask to the middle or the end, after you have earned it.

Then watch your own retention graph on the last five videos and find the timestamp where people leave. It is almost always the same place structurally, and it is almost always something you could cut.

Topic research: two different businesses

A distinction worth making before choosing what to make, because they need different videos.

Search-driven videos answer a question people type. Demand is stable, the audience arrives with intent, and the video keeps working for years. These are tutorials, comparisons, fixes and explanations. This is where the compounding comes from.

Browse-driven videos are recommended to people who were not looking for them. Bigger ceiling, far less predictable, and dependent on the packaging doing the work of persuasion.

For anyone building with limited time, search-driven is the better foundation: the demand exists before you make the thing, and you can verify it.

• Find real queries rather than guessing at topics. What people type is rarely how you would phrase it.
• Look at what already ranks for that query — if every result is a ten-minute tutorial and you planned a two-minute overview, the format is decided for you.
• Ask a model for twenty angles on one topic, then pick using whether you can add something. See AI for SEO; the principles carry over.
• Your own comments and questions are the best source of all, because the demand is proven and specific.

Production: script to video

1. Script: AI content writing for structure and drafting — hook, body, close — with your own material in it.
2. Narration: a generated voice, or your own, which is still the better option if you can bear recording.
3. Visuals: AI video, images, stock, screen recordings, or a talking avatar.
4. Edit: assembly, b-roll, captions and music — kept well under the voice.

One rule that improves generated scripts more than any prompt tweak: cut the first paragraph and the last paragraph. Models open with context nobody needs and close with a summary of what was just said. Both are exactly where viewers leave.

Holding attention through the middle

Openings get all the attention and the middle is where most watch time is actually lost.

• Change something every few seconds — a cut, a zoom, a graphic, a change of location on screen. Static talking is where the graph slopes down.
• Signpost. "That is the first of three — the second is the one that catches people out." A named thing still to come is a reason to stay.
• Cut ruthlessly. If a section is not the reason anyone clicked, it is padding, even when it is good. Length is not a virtue; the platform rewards watch time, which is length multiplied by the share of people still there.
• Use chapters on anything long. They look like an invitation to skip and in practice they keep people who would otherwise have left entirely.
• End the moment you are done. An outro that runs after the value has been delivered is measured as people leaving.

Shorts, and how they relate to the main channel

Short vertical video sits on the same channel and behaves like a different platform, which causes a lot of confusion about why subscriber numbers move without anything else moving.

Shorts reach people fast and convert them into committed viewers slowly. Someone who watched a thirty-second clip is not the same asset as someone who watched twelve minutes, and a channel that grows on shorts frequently finds its long-form videos are shown to almost none of those subscribers — because the system has learned they watch shorts.

That is not an argument against them. It is an argument for being deliberate:

• Use shorts to demonstrate the value, not to advertise the long video. A clip that ends with "full video in the description" performs worse than one that simply delivers a complete small thing and leaves an impression of usefulness.
• Cut them from real videos rather than making them separately — the production is nearly free once the long form exists, and the material is already proven.
• Expect the subscriber-to-view relationship to break. Judge long-form on its own numbers rather than on the channel total.
• Do not let shorts set the editorial direction. What performs in a scroll feed is not what compounds in search, and chasing the first can quietly turn a useful channel into a forgettable one.

What to make first

A specific starting sequence, because "make videos consistently" is true and useless.

1. Find five questions you get asked repeatedly — by customers, colleagues, or in the comments of other people's videos in your area. Proven demand, no research needed.
2. Check each is actually searched, in the words people use rather than yours.
3. Make all five as plainly as possible. Screen recording and your voice is enough. The production can improve later; what you need now is to learn whether anyone wants these.
4. Keep the packaging consistent across the five so the comparison means something.
5. Read the retention graphs together. The pattern across five videos tells you more than any single one.
6. Then double down on whichever question had the most demand, and make three more around it. Depth in one area beats one video each on five topics — both for the algorithm and for the people deciding whether to subscribe.

Notice how little of that requires the production pipeline described above. Build the pipeline once you know what you are making every week; building it first is a comfortable way of avoiding the harder question.

Monetisation and originality

If the plan involves earning from the platform, this decides whether the plan works.

Partner programmes require content that is meaningfully your own, and explicitly exclude material that is mass-produced, repetitive or assembled from other people's work without substantial transformation. Enforcement has been tightening. A channel of templated narrated slideshows is the pattern being targeted, and the rejection tends to arrive after you have built an audience — at which point no amount of editing fixes it, because the format itself is the problem.

What tends to satisfy the requirement: your own commentary and analysis, original research, your own footage or screen recordings, a recognisable editorial point of view, and genuine variation between videos. The faceless content guide covers this in more detail — it applies identically here.

Separately, disclose synthetic content where required. Platform rules on realistic AI-generated people and voices have been expanding and are enforced independently of copyright and monetisation. Read the current policy rather than a summary.

Title, thumbnail and description

• Title: generate ten, choose one that promises something specific and deliverable.
• Thumbnail: see thumbnails with AI, and remember it should not repeat the title.
• Description: the first line or two is what appears in search and above the fold — write it for a person deciding whether to watch, not for a keyword. Timestamps and links below that.
• Clips: cut moments for short-form with repurposing, each working on its own rather than as a trailer.

Tags do very little now. The transcript matters more than the tag box, which is another reason to fix the automatic captions on anything you expect to be found by search.

A pipeline with a human in it

The stages wire together in n8n or Make: an approved idea → script → narration → assembly → title, thumbnail and description → uploaded as a private draft for review. Templates in the templates hub.

Build it so that nothing publishes without you, and be honest about which steps you are automating. Automating the mechanical work — rendering, captioning, drafting descriptions, cutting clips — is straightforwardly good. Automating the judgement, so that videos appear on your channel that you have not watched, produces exactly the output the originality policies exclude, and does so at a volume that makes the channel harder to rescue.

A reasonable division: the machine produces a complete draft; you watch it once at normal speed before approving. If you would not sit through your own video, nobody else will.

Which numbers to look at

• The retention graph, per video, and specifically where it drops. This is the only feedback that improves the next one.
• Click rate together with average view duration. Either alone misleads; the pair tells you whether the packaging matched the content.
• Impressions. Low impressions means a topic or distribution problem, not a packaging one.
• Returning viewers. The number that tells you whether you are building an audience rather than collecting views.
• Which videos bring subscribers — usually not the ones with the most views, and worth knowing before you decide what to make more of.

How a channel actually earns

Worth being concrete, because the assumed answer — advertising revenue — is usually the smallest and slowest of the options.

• Selling your own thing. A service, a product, a course. This works at audience sizes far below anything the platform will pay for, because a few hundred viewers with the exact problem you solve is a business. It is also the only route where a small, specific channel beats a large general one.
• Sponsorship. Real money once the audience is defined enough that a brand wants precisely those people. Being narrow is the asset; a sponsor pays for relevance rather than reach.
• Affiliate. Works early, in niches where people are choosing between products. Disclose it in the video and in the description — a legal requirement in most markets, not a courtesy.
• Platform advertising revenue. Requires crossing the threshold, passing the originality review, and then accumulating a great deal of watch time at a low rate. Fine as a by-product; a poor plan.

The order matters for what you make. Optimising for advertising revenue pushes towards long videos on broad topics with high view counts. Optimising for the first three pushes towards shorter videos on specific problems for people who could become customers — usually smaller numbers and considerably more money.

The comments are the research budget you already have

Most creators treat comments as a chore or a vanity metric. They are the cheapest audience research available and they compound in a way the videos themselves do not.

Three things to do with them. Answer the questions — each one is a video someone has already told you they want, phrased in their own words, which is also your title. Note the disagreements, because a point people push back on is a point worth a video of its own. And watch which parts get quoted back: the sentence people repeat in the comments is the thing your video was actually about, and it is often not the thing you thought.

A model is genuinely useful here once the volume grows — paste a few hundred comments and ask for the recurring questions and the recurring objections. That is summarising real material rather than inventing content, which is the mode these tools are reliable in.

When YouTube is the wrong channel

• You need results this quarter. The compounding is real and slow. Search visibility takes months.
• Your audience is tiny and specific. A hundred right people might be reachable by email far more cheaply.
• The content is genuinely better as text. Reference material people scan is worse as video, whatever the reach.
• You cannot sustain a cadence. One video a month, consistently, beats six in a fortnight and then silence.

Common mistakes

• Volume over value. Watch time is what is rewarded. Fast production of boring video is efficient failure.
• Packaging as an afterthought. Write the title and thumbnail first; they are the test of whether the idea is clear.
• An introduction. The commonest thirty seconds of lost retention on the platform.
• Flat narration. Direct the tone, vary the pace, and cut the breaths that go nowhere.
• Ignoring the originality and disclosure policies until the monetisation application is rejected.
• Expecting virality. This is a compounding asset or it is nothing.