I run a visa agency in Bali and build scrapers on Apify on the side. One of them is a YouTube Shorts transcript scraper: you give it Shorts links or a channel, it reads the caption track YouTube already publishes and returns one row per Short with the text, timestamped lines, the words in the first three seconds, and the Short's numbers.
Every evening a health check runs it with the same input: the newest 10 Shorts from @MrBeast plus one fixed Short link. Between 3 and 19 November 2026 that produced 20 runs and 220 rows covering 12 different Shorts. I never meant it as a dataset, but it answers a question people building on Shorts data keep getting wrong: which parts of a row are a fact and which parts are a snapshot?
The input
{ "videoUrls": ["https://www.youtube.com/shorts/Df5Y-2ndQyU"], "profiles": ["https://www.youtube.com/@MrBeast"], "resultsLimitPerProfile": 10, "maximumVideoDurationSeconds": 300, "includeTranscriptSegments": true } Runs took between 26 and 63 seconds for 11 Shorts, with one outlier at 189 seconds that I come back to below.
What never changed
For every one of the 12 Shorts, across every run it appeared in (4 to 20 runs each):
-
transcriptText— one distinct value per Short. Not "nearly the same"; identical strings. -
firstThreeSecondsSpeech— one distinct value per Short. -
postedAt,durationSeconds,title— one value each.
That sounds obvious until you have worked with speech recognition. On TikTok, where most videos have no captions and the audio goes through a model, I once fed the same MP4 in 27 times and got 15 different transcripts. YouTube Shorts are different because nothing is recognised on my side: the actor downloads the caption track YouTube serves, and YouTube serves the same file to everyone. All 220 rows here came back with captionsType: "auto" (YouTube's automatic captions, not the creator's own subtitles) and transcriptSource: "platform-captions".
The practical consequence: a transcript is a one-time purchase. Once you have it for a video ID, re-reading it tomorrow buys you nothing.
What moved, and how precisely
Here is kPI9Mc35mOg ("Trap Door Roulette"), posted 6 November at 17:00 UTC, as my evening runs saw it:
| Read (UTC) | Age | views | likes | comments |
|---|---|---|---|---|
| 6 Nov 20:10 | 3 h | 4,253,948 | 140,859 | 2,600 |
| 6 Nov 20:56 | 4 h | 5,140,964 | 160,730 | 2,800 |
| 7 Nov 20:12 | 27 h | 42,187,162 | 681,087 | 6,000 |
| 8 Nov 20:07 | 2 days | 55,010,685 | 856,657 | 7,600 |
| 10 Nov 20:18 | 4 days | 68,863,404 | 1,113,739 | 9,500 |
| 12 Nov 03:32 | 5.4 days | 75,857,942 | 1,231,385 | 10,000 |
| 15 Nov 20:11 | 9 days | 94,785,829 | 1,487,060 | 12,000 |
| 19 Nov 23:01 | 13 days | 112,230,670 | 1,703,938 | 13,000 |
Ten times the views between hour 4 and hour 27. Same transcript in all 13 reads.
Looking across all 12 Shorts, each counter has its own character:
views is exact but can lag. On 5 November two runs started one minute apart. 41UzbuW6wJs showed 64,812,114 views in both, while its likes went from 1,008,331 to 1,008,347. The same happened on 10 November at 20:18 and 20:19: views identical for 10 of the 11 Shorts, likes different for most of them. View counts on YouTube are refreshed in batches; likes are not. If you compute "views per minute" from two close reads, you can get zero for a Short that is clearly growing.
likes is exact and can go down. 5mU6SRS2Bxo had 469,492 likes at 22:49 on 19 November and 469,488 at 23:01. People un-like. Don't treat a drop as a scraping bug, and don't assert monotonic growth in your pipeline.
comments is rounded the way YouTube displays it. Above ten thousand it moves in steps of 1,000 (70,000 → 71,000 → 72,000 → 73,000 for LiH-P4rSkLI over 16 days); below that in steps of 100 (2,600 → 2,800). The value is a parsed "70K", not a count. If you need the exact number, this field cannot give it to you.
channelSubscribers is rounded to the million on a channel this size: 525,000,000 from 3 to 8 November, 526,000,000 from 9 November, 527,000,000 on the 19th. Fine for "how big is this channel", useless for daily growth.
Store it as two tables
Given all that, I would not keep the raw rows. Split each one into a static part, keyed by videoId and written once, and a snapshot part appended on every read:
from apify_client import ApifyClient STATIC = ("videoId", "url", "title", "author", "channelId", "postedAt", "durationSeconds", "language", "captionsType", "transcriptText", "firstThreeSecondsSpeech", "transcript") SNAPSHOT = ("views", "likes", "comments", "channelSubscribers", "scrapedAt") client = ApifyClient("<YOUR_APIFY_TOKEN>") run = client.actor("lergassy/youtube-shorts-transcript-scraper").call(run_input={ "profiles": ["https://www.youtube.com/@MrBeast"], "resultsLimitPerProfile": 10, }) videos, snapshots = {}, [] for row in client.dataset(run["defaultDatasetId"]).iterate_items(): if row.get("type") == "error": continue # error rows are free; retry them later vid = row["videoId"] videos.setdefault(vid, {k: row.get(k) for k in STATIC}) snapshots.append({"videoId": vid, **{k: row.get(k) for k in SNAPSHOT}}) Then, once a video's transcript is in your videos table, stop paying for transcripts. If all you want next week is the view curve, the official YouTube Data API returns statistics.viewCount for 50 IDs per request against your free daily quota. A transcript scraper is the wrong tool for polling numbers; it is the right tool for getting the words once.
"Newest 10" is a sliding window, not a date range
On 3 November the newest 10 Shorts on the channel were published between 31 July and 30 October: a 92-day window, because the channel posts a few Shorts a month. On 6 November "Trap Door Roulette" came out and the oldest one, from 31 July, dropped out of the list. On 13 November another new Short pushed out the one from 2 August.
Two consequences. If you want "everything since date X", set onlyPostsNewerThan instead of guessing a count; the actor stops at the first older Short and older ones are not charged. And row order is not stable: the 11 rows came back in a different order in each of the 20 runs, because they are fetched in parallel. Sort by postedAt yourself.
Two traps in the "hook" column
firstThreeSecondsSpeech holds the words whose captions start before the 3-second mark. Mostly that gives you exactly what you want:
"3 2 1 [screaming] pull!" "I just bought a $20,000 robot. Let's see how it handles" "Gross." But for CEJXqm2eiJ0 ("What's Inside My Briefcase?") it is null in every run. That is not a parsing failure: the first caption line starts at 14.56 seconds. The Short opens with 14 seconds of action and music. A null hook is a signal about the edit, so keep it as null rather than dropping the row.
The second trap is that YouTube's automatic captions use capitals for loud speech, and MrBeast is loud: "WE'RE UNBOXING THE BIGGEST TOY IN THE WORLD. WOAH, WOAH, WOAH." The actor leaves the text as YouTube sends it (it only decodes HTML entities like '). If you are counting words or comparing hooks across channels, lower-case first. Markers like [music], [screaming] and the >> speaker-change sign also stay in, which is useful for some analyses and noise for others.
The "viral" flag depends on when you ask
The actor has an optional compareWithAuthor switch that compares each Short's views with the channel's median in the same run and flags anything at 3x or more. Run it at 20:10 on 6 November and "Trap Door Roulette" sits at 4.25M against a median of 68.3M: 0.06x the channel's usual. Two days later it had 55M. Any outlier measure built on raw views is meaningless for a Short that is a few hours old. Either compare views at the same age (which needs the snapshot table above) or leave Shorts younger than about a week out of it.
When a run fails
On 12 November at 03:04, six of the 11 rows came back as error rows:
{ "type": "error", "videoId": "kPI9Mc35mOg", "errorCode": "youtube_refused", "errorMessage": "YouTube refused all 8 requests for this video from our residential addresses (... Tunnel connection failed: 402 Payment Required ...). Nothing was charged for this URL; running it again usually succeeds." } That run took 189 seconds instead of the usual 30–60, because each Short is retried up to eight times before it gives up. The cause in that message was on my side (the proxy tunnel answered 402), not YouTube. A rerun 28 minutes later returned all 11 Shorts. Error rows are not billed, and they come after the data rows, so the pattern is simple: collect the error videoIds and send them again in a later run.
When this is the wrong tool
- Shorts with no speech. Music-only Shorts usually have no caption track, so you get a free
no_captionserror row and no text. Nothing is transcribed from audio for YouTube. - Text burned into the picture. Captions are what was said, not what was written on screen.
- Exact comment counts or daily subscriber growth. The page only shows rounded values, so that is all you get.
- Tracking view counts. Get the transcript once, then poll numbers through the YouTube Data API.
For the part it is built for (the words of a Short, once, with a timestamp per line and the numbers at the moment you read it), the result was the same 220 times out of 220 in my data. Price and input options are on the actor page.