How to make a chatbot from your YouTube videos

A document arrives with its structure intact. It has headings, paragraphs and page numbers, and a machine can find the paragraph that answers a question and tell you where it was.

A video has none of that. It is an hour of speech with no headings, no paragraphs and no index, and the useful unit — the ninety seconds where you actually explain the thing — is buried somewhere in the middle with nothing marking it. Add the way people really talk on camera, with false starts, tangents and “as I said in the last video”, and the raw material is genuinely messier than a PDF.

The upside is that video libraries are usually far richer than their owners realise. A creator with two hundred uploads has answered most of their audience’s questions several times over, at length and well. None of it is findable. That gap is exactly what this is for.

Everything downstream depends on turning speech into text, and there are two ways it happens.

• Existing captions. If a video already has captions — yours or auto-generated — they can often be pulled directly, which is the fast path. Auto-generated captions are imperfect but usually good enough for retrieval, which tolerates the odd wrong word far better than a reader would.
• Speech-to-text. Where captions don’t exist, the audio is transcribed. Slower, and the quality depends on audio quality — a clean voice recording transcribes far better than a windy outdoor clip with music underneath.

What you should not do is transcribe a back catalogue by hand. People do consider it, and it is never worth the weeks. If a tool asks you to supply transcripts yourself, that’s a tool that hasn’t done the hard part.

One thing worth knowing before you plan around it: transcription gets the words, not the pictures. If the substance of a video is a diagram you draw silently on screen, the transcript won’t carry it — say what you’re drawing out loud, which is better for accessibility anyway.

A transcript alone gives you a searchable wall of text. What turns it into something your audience values is keeping the time information attached to every passage, so an answer can point at the exact moment it came from.

The difference is stark in practice. “You covered this in your video on position sizing” sends someone to a 40-minute video to hunt. “Here’s the answer, and you said it at 8:12 in this video” sends them to the ninety seconds that matter — and, incidentally, sends them to your video rather than away from it.

This is also the honest test of whether a tool is doing the job properly. If answers come back with no timestamp, the time data was thrown away at ingestion and no amount of prompting will get it back. On Ultranivo’s YouTube assistant — ours — every answer carries the video and the timestamp for this reason.

Pasting links one at a time is fine for five videos and absurd for two hundred, so whole-channel import is the feature that decides whether this is a project or an afternoon. It’s included on every Ultranivo plan.

Two practical notes. First, more is not automatically better: three years of uploads include things you’ve since changed your mind about, and an assistant answering from your 2023 position is a genuine problem. Consider starting with your evergreen material and adding the rest deliberately.

Second, plenty of your best material was never on YouTube — course modules, recorded sessions, member calls. Those can be uploaded directly as video files, and they’re often the material your audience would most like to be able to ask. If you gate that behind a subscription, the same assistant can sit behind a paywall.

• Stale opinions answered as current. The commonest complaint from creators. If you’ve publicly changed your view, remove the old video from the assistant or expect to be quoted on it.
• Poor audio. Wind, music beds, several people talking at once — transcription degrades and so do the answers. Check a couple of your roughest videos rather than your best one.
• Expecting a voice clone. An assistant over your videos answers with your substance and your vocabulary. It is not you, it does not do your delivery, and presenting it as though it were is the kind of thing an audience notices and resents.
• Nothing in writing. Video alone often misses the specifics — prices, dates, links, exact steps. Adding the PDFs and notes behind your content covers the gaps; documents and video feed the same assistant, so questions get answered from whichever source actually has it.