YouTube Shorts transcripts: the format most tools skip

Most transcript tools return nothing for Shorts, and they do it quietly. You pass a Shorts URL, the run completes, the output is empty, and there is no error explaining why.

The cause is usually not the URL format. It is that Shorts overwhelmingly carry auto-generated captions only, and a tool that asks for a human-written caption track and stops when there is none will report a Short as having no transcript. It has one. It was not asked for.

Shorts are ordinary videos with an awkward front door

A Short has an eleven-character video ID exactly like any other upload, and it is reachable at both youtube.com/shorts/<id> and youtube.com/watch?v=<id>. There is no separate Shorts caption system to integrate with.

What differs is discovery. Shorts live on their own channel tab rather than in the main uploads feed, so tooling built to walk a channel’s videos can miss them entirely — not because it failed, but because it never looked there. If you are collecting a channel’s Shorts, the Shorts tab is the list to walk.

Why the captions come back empty

Three reasons, and they need telling apart because only one is fixable.

Auto-generated only. Nobody uploads an SRT for a forty-second clip. If your tool requests a manually-written track in a given language and treats its absence as “no captions”, it will report almost every Short as having none. The fix is a fallback order: human-written first, then auto-generated in the language you asked for, then any track that exists.

Genuinely no captions. Some Shorts have none at all — very short clips, music-only clips, and clips where automatic speech recognition declined to produce anything. No tool can recover text that was never generated.

Blocked extraction. The request was refused rather than answered. This is an infrastructure problem and it looks identical to “no captions” unless the tool distinguishes them. If you are running from a datacenter IP, this is the likely explanation — our reliability benchmark measured 0% success unproxied against 100% from residential egress.

A tool that collapses all three into an empty cell has hidden the only information that would tell you whether to retry.

What Shorts text is good for

The interesting property is that a Short is entirely script. There is no preamble, no sponsor read, no ten-minute middle. Sixty seconds of copy, written to hold attention from the first word, published at volume by people testing what works.

That makes a channel’s Shorts corpus unusually well suited to a few things:

Hook analysis. The first sentence of a Short is the most-rewritten sentence in short-form video. Collect a few hundred first lines across the creators in a niche and the patterns are visible without any modelling.
Claim tracking. What a brand or creator is currently asserting, at the cadence they publish rather than the cadence they release long-form.
Volume. Channels publish Shorts far faster than long-form, so the same channel yields more text and more recent text through the Shorts tab.

Pair the text with the metadata that comes free — title, view count, duration, publication order — and the comparison you usually want (which hooks travelled) is available without extra collection.

Collecting a channel’s Shorts

Work from the channel, not from individual URLs. Shorts are published in volume; collecting them one at a time is the wrong unit.
Cap the count per channel. Newest-first with a ceiling is both a cost control and usually what you want — hook conventions from two years ago are not the current ones.
Set a language with a fallback, not a requirement. Asking for en and refusing anything else is the single most common reason a Shorts run returns nothing.
Leave timestamped segments off unless you need them. A Short’s plain text is a paragraph; the segment array for a channel’s worth of Shorts is a large multiple of that for no gain if you are analysing wording.
Route through residential egress. See the benchmark linked above — this decides almost everything about whether the run works at all.
Keep the failures with their reasons. A no-captions row and a blocked row mean different things, and only one of them is worth retrying.

Limitations

A Short with no caption track cannot be transcribed by any extraction tool. The text does not exist to be fetched. Generating it means running speech recognition on the audio, which is a different and much more expensive operation.

Auto-generated captions carry errors, and they cluster on exactly the words that matter in short-form: brand names, slang, product names, and anything said quickly. For hook analysis this is usually tolerable because the structure survives. For quoting someone accurately it is not — check anything you intend to publish against the video.

No punctuation or speaker labels. Automatic tracks give you a stream of words. Sentence boundaries in short-form are often inferable from the pacing, but they are inferred, not given.

Music, sound effects and on-screen text are absent. A large share of short-form meaning is visual — captions burned into the video, text overlays, what is being shown. Transcripts capture none of it, and for some niches that is most of the content.

Shorts churn. Creators delete underperformers. A corpus collected today will not fully re-collect in six months, so keep what you gather.

FAQ

Why does my transcript tool return nothing for Shorts?

Most likely it is requesting a human-written caption track and stopping when there is none. Shorts almost always have auto-generated captions only. A tool with a fallback to auto-generated tracks returns text where a stricter one returns an empty result.

Do Shorts have different URLs I need to handle?

youtube.com/shorts/<id> and youtube.com/watch?v=<id> refer to the same video, and the ID is the same eleven characters used everywhere else. Anything that parses one form can be taught the other.

Can I get every Short from a channel?

You can walk the channel’s Shorts tab, which is where they are listed rather than in the main uploads feed. Whether every Short comes back depends on what is still public — deleted and private ones are gone.

Is a Shorts transcript useful if it has no punctuation?

For counting, clustering and comparing wording, yes — those operations do not need sentence boundaries. For anything a human reads, run the text through a language model to punctuate it, and keep the raw version as the record.

How much text is a Short?

Roughly 100 to 180 words for a sixty-second clip at normal speaking pace. A channel’s last two hundred Shorts is on the order of 30,000 words, which is small enough to read in full and large enough for patterns to be real.

Hosted version

This page describes the problem. The hosted tool solves it at scale — bulk input, structured output, and no charge for videos that return nothing.

Run it on Apify