What we collected from Duolingo’s Shorts tab, and how
The dataset holds 47 videos from one account: the official Duolingo YouTube channel. The window runs 2026-05-30 to 2026-08-28, the 90 days before collection. Every count was read on a single date, 2026-08-28.
The selection rule was frozen before collection, in three steps.
- Take the account: the official Duolingo YouTube channel. The channel was picked before any counting because it is a named, publicly visible brand account. One account is an editorial choice, not a sample.
- Include every video listed on that channel’s Shorts tab with an upload date inside the window. The tab lists newest first, so extraction stopped after three consecutive videos older than 2026-05-30.
- For each video, read the public counts as YouTube reported them on 2026-08-28: view count, like count, comment count, plus upload date, duration, title, and the video’s own URL.
Step 3 was done by full per-video extraction rather than by reading the counts printed on the channel listing page. Those listing-page numbers lag the per-video numbers, which means a census built from the listing alone is wrong in a direction the reader cannot see. That is the single detail most likely to be skipped by anyone repeating this, and it is the one that decides whether the figures below mean anything.
Two dates inside the dataset are worth stating plainly. The earliest upload present is 2026-06-01 and the latest is 2026-08-27. No video in the dataset carries an upload date of 2026-05-30 or 2026-05-31, so the 47 videos span 88 days rather than the full 90-day window.
Here is one video, read end to end, so the shape of a single row is concrete. The newest Short in the dataset, @KaiCenat let me teach at streamer university, was uploaded 2026-08-27, runs 10 seconds, and held 127,419 views, 4,253 likes and 545 comments as read on 2026-08-28. It was one day old. That last number is the reason the next section exists.
What the dataset supports: claims about what this account published in this window, and what the public counters said on 2026-08-28.
What it does not support: any claim about videos outside the window, videos removed before collection, other Duolingo surfaces, other brands, or the counters on any other date.
Why 13 of the 47 videos are out of the best and worst comparison
A view count is cumulative, so it only ever goes up as more playbacks happen. A count read on one date therefore measures a video’s age as well as its reach. YouTube’s own documentation confirms that the number is not fixed on publication: the platform states that its systems take time to determine which views, likes and subscriptions are legitimate, that metrics may take time to appear during the first few hours after a video is published, and that a metric count may be temporarily slowed, frozen or changed while that confirmation happens (How engagement metrics are counted, YouTube Help, retrieved 2026-08-28).
The dataset shows the effect at its own bottom end. The four lowest view counts in the full 47 belong to the four youngest videos, aged 1, 2, 3 and 4 days on the collection date. The lowest count in the whole dataset, 61,788 views, belongs to i only main multilingual brawlers (and Edgar) #brawlstars, uploaded 2026-08-26 and two days old when it was read.
Calling that video the account’s worst Short of the window would be arithmetically true and substantively wrong. It is the sentence this whole section exists to prevent.
So every best and worst comparison on this page runs on a subset: the videos at least 21 days old on 2026-08-28. That keeps 34 videos and puts 13 out. Dropping the youngest observations because they have not finished accumulating is right-censoring, which is the plain problem of measuring an unfinished count.
The 21-day figure is our choice, not a platform standard. Two things make the choice auditable rather than arbitrary.
First, the age distribution has a gap exactly where the cutoff falls. The youngest video kept is 23 days old and the oldest video excluded is 17 days old, and no video in the dataset is between 18 and 22 days old. Any cutoff from 18 through 23 days produces the same 34 videos.
Second, the named winners and losers do not move when the cutoff does. At 14 days the subset is 37 videos with a median of 834,918 views. At 21 days it is 34 videos with a median of 817,268. At 28 days it is 29 videos with a median of 789,070. The highest-view video and the lowest-view video are the same two videos at all three cutoffs.
One documented change sits inside the window and needs recording here. YouTube states that “Beginning August 24, 2026, views are counted the moment a video starts to play across all formats, including Shorts, long-form videos (VOD), and live streams” (How engagement metrics are counted, YouTube Help, retrieved 2026-08-28). Four videos in the dataset were published on or after that date, and all four are among the 13 excluded. Every one of the 47 videos accrued its final four days of views after it.
The 13 excluded videos are reported here with their counts, and no ranking claim attaches to any of them.
Counts as read on 2026-08-28. Window 2026-05-30 to 2026-08-28, n = 47, of which these 13 are excluded from every comparison below.
The spread across 34 settled Shorts
The settled subset runs from 363,062 views to 5,718,146 views. That is a ratio of 15.75 to 1 between the account’s own highest and lowest Short in this window, on videos that all had at least three weeks to accumulate.
Quartiles here use the exclusive convention, meaning each quartile is interpolated at position (n+1)p in the sorted values. That is stated so every figure in the table can be reproduced from the 47 rows published further down.
| Counter (n = 34) | Min | First quartile | Median | Third quartile | Max | Mean |
|---|---|---|---|---|---|---|
| Views | 363,062 | 612,326 | 817,268 | 1,077,264 | 5,718,146 | 1,065,427 |
| Likes | 6,455 | 9,519 | 14,655 | 20,275 | 101,891 | 20,140 |
| Comments | 367 | 931 | 1,350 | 1,825 | 9,000 | 1,798 |
Settled subset, videos at least 21 days old on 2026-08-28, counts as read on 2026-08-28.
The mean sits in that table on purpose, next to the median rather than instead of it. The subset mean of 1,065,427 views is 30% above the median of 817,268, and the gap between those two numbers is the finding.
The middle of the subset is tight. The third quartile divided by the first quartile is 1.76, so the middle half of this account’s Shorts sits inside a range of less than two to one. The ends are not tight at all. One video accounts for the difference between those two descriptions.
The median falls between two videos, habibird hamood at 799,617 views and hee hee at 834,918 views. The second of those runs 6 seconds, was 36 days old on the collection date, and holds 17,899 likes and 1,100 comments. A typical Short from this account in this window looks like that, and it is not the video anyone would show you.
Concentration is where the distribution stops being an abstraction. The 34 settled videos hold 36,224,518 views between them.
- The single highest video holds 15.8% of that total.
- The top three hold 28.7%.
- The top five hold 37.4%.
- The lower half, 17 of the 34 videos, holds 29.3%.
Half of this account’s settled output is worth roughly what its top three videos are worth. Remove the one highest video and the subset of 33 has a median of 799,617 views and a maximum of 2,860,117, which is less than half the maximum it had a moment ago.
Across the full 47 videos, including the 13 too new to compare, the counters total 41,767,047 views, 817,250 likes and 75,436 comments. The full-set quartiles for likes are 2,415 / 9,103 / 13,072 / 18,338 / 101,891 and for comments 277 / 740 / 1,100 / 1,600 / 9,000, in the order minimum, first quartile, median, third quartile, maximum.
This distribution describes this account’s Shorts in this window and nothing else. It is not a benchmark, and it says nothing about what any other account should expect from 34 videos.
The five highest-view Shorts in the window
All five come from the 34-video settled subset, so every one of them had at least three weeks to accumulate before its count was read.
| Rank | Video | Uploaded | Age on 2026-08-28 | Duration | Views | Likes | Comments |
|---|---|---|---|---|---|---|---|
| 1 | do you speak @MrBeast rizz | 2026-07-22 | 37 days | 7s | 5,718,146 | 101,891 | 8,100 |
| 2 | d’oh!lingo | 2026-08-04 | 24 days | 8s | 2,860,117 | 61,146 | 9,000 |
| 3 | all i’ve ever known is salchipapa | 2026-07-16 | 43 days | 12s | 1,810,457 | 23,487 | 1,500 |
| 4 | baba…don’t let us down | 2026-07-10 | 49 days | 12s | 1,745,091 | 30,876 | 3,800 |
| 5 | trying out a new mating call @Tinder | 2026-06-25 | 64 days | 7s | 1,416,898 | 16,486 | 1,700 |
Counts as read on 2026-08-28. Settled subset, n = 34, window 2026-05-30 to 2026-08-28.
What is publicly visible in each one, in a sentence: the top video is a spoken-language bit involving another creator’s handle; the second is a Simpsons-styled pun on the brand name; the third is a line about salchipapa; the fourth is a short piece addressed to “baba”; the fifth pairs a bird call with another brand’s handle. That is a description of what a viewer sees, not a reading of why any of it was made.
The rank order changes if you count comments instead of views. The number two video holds 9,000 comments against the number one video’s 8,100, on roughly half the views. That inversion is only visible because all three counters were read for every video rather than one headline number per video.
Nothing in this ranking establishes that any property of these five videos caused their counts. The next section tests the properties that can be observed and reports what happened.
The five lowest-view Shorts in the window
These five are the lowest view counts inside the same 34-video settled subset. The two-day-old video with 61,788 views is not among them, because it is one of the 13 excluded.
| Rank from bottom | Video | Uploaded | Age on 2026-08-28 | Duration | Views | Likes | Comments |
|---|---|---|---|---|---|---|---|
| 1 | consider this a formal petition to add this category to @LasCulturistas awards | 2026-06-04 | 85 days | 24s | 363,062 | 7,421 | 979 |
| 2 | the beautiful game? flirting in another language @Tinder | 2026-06-18 | 71 days | 29s | 422,404 | 6,455 | 367 |
| 3 | i am he who knows 10,000 languages | 2026-07-11 | 48 days | 9s | 484,235 | 7,971 | 702 |
| 4 | tell me what you got in the comments! @webtoonofficial | 2026-06-12 | 77 days | 8s | 499,545 | 12,752 | 1,800 |
| 5 | 6.10.26 | 2026-06-08 | 81 days | 10s | 515,028 | 9,199 | 722 |
Counts as read on 2026-08-28. Settled subset, n = 34, window 2026-05-30 to 2026-08-28.
The lowest-view Short in the settled subset holds 363,062 views. On this account, in this window, the floor is high. Every one of these five sits above 360,000 views and four of the five sit above 420,000.
The ranking is also about to change. The fourth video from the bottom, the one whose title asks for comments, has the highest comment rate in the entire settled subset. That figure comes two sections down.
Low relative to this account is not low in any absolute sense. No cross-account comparison exists in this dataset, and none is implied by the order of this table.
The observable properties that did not separate them
The dataset can observe two things about a video besides its counts: its duration in seconds, and the text of its title. Inside the titles, three markers are visible enough to split on. A title may tag another account with an @ handle, it may carry a hashtag, and it may credit an illustrator with a paint-palette emoji. That is the entire set of properties available. Everything else that plausibly matters, starting with what is in the frame and who is in it, is not in the dataset.
Four splits were computed on the 34 settled videos. All four report the median views on each side.
| Property | With it | Median views | Without it | Median views | Gap |
|---|---|---|---|---|---|
| Another account’s @ handle in the title | 9 videos | 789,070 | 25 videos | 834,918 | 5.5% lower |
| A hashtag in the title | 3 videos | 845,084 | 31 videos | 799,617 | 5.7% higher |
| An illustrator credit in the title | 4 videos | 838,743 | 30 videos | 817,268 | 2.6% higher |
| Duration of 10 seconds or under | 20 videos | 851,228 | 14 videos | 756,292 | 12.6% higher |
Settled subset, n = 34, counts as read on 2026-08-28.
None of the four splits separates the top of this account’s range from the bottom. The largest gap of the four is duration, at 12.6%, measured on a subset whose highest and lowest videos differ by a factor of 15.75. Across the same 34 videos, duration and views have a rank correlation of minus 0.22, computed on the collected counts with tied durations given average ranks. Duration in the settled subset runs from 5 seconds to 29 seconds with a median of 10 seconds.
The @ handle split needs one caveat, because the group is not one thing. Of the 9 settled videos whose titles carry an @ handle, 4 credit an illustrator rather than tag another account. Separating them leaves 5 videos that tag another account, with a median of 499,545 views against 834,918 for the 25 videos carrying no handle at all, and n = 5 cannot carry a finding of any kind.
The top three videos by views break the pattern between themselves. Number one carries an @ handle. Number two and number three carry none. Three of the five lowest-view videos carry an @ handle. The marker that looks like it should matter appears at both ends of the range.
Duration behaves the same way. The two longest videos in the settled subset run 29 seconds and 24 seconds, and they are also the two lowest by views, at n = 2. Two videos are two videos.
What these four splits support: on these 34 videos, none of the four observable properties separates high from low.
What they do not support: a null result on four properties is not evidence that nothing separates these videos. It is evidence that nothing the public surface exposes does. Duration here is a recorded attribute of videos that were already published, not a specification for anything.
Likes and comments rank the same videos differently
Views are not the only counter in the dataset, and the other two rank the same 34 videos in a different order. Likes per 1,000 views and comments per 1,000 views are arithmetic ratios of two collected counts, nothing more. Neither has a denominator in impressions, so neither can be compared with any published engagement-rate figure from any source.
| Rate (n = 34) | Min | First quartile | Median | Third quartile | Max |
|---|---|---|---|---|---|
| Likes per 1,000 views | 11.6 | 15.3 | 17.0 | 20.8 | 52.5 |
| Comments per 1,000 views | 0.74 | 1.29 | 1.47 | 2.02 | 3.60 |
Settled subset, arithmetic on counts as read on 2026-08-28.
Views and likes have a rank correlation of 0.89 across the settled subset. Views and comments have a rank correlation of 0.76. Both are computed on the collected counts with ties given average ranks.
The highest like rate in the window belongs to congrats it’s a streak! (now go to the app and do it for real), at 52.5 likes per 1,000 views on 967,997 views and 50,779 likes. That video ranks 11th of 34 on views and 3rd of 34 on like count. Nobody looking at the view column would pick it out.
Two more sit near the top of the like-rate order. only one week left to revive your streak for real holds 33.1 likes per 1,000 views on 1,113,281 views and 36,807 likes. omg we’re back again holds 27.7 on 613,485 views and 17,019 likes, from 26th place of 34 on views.
The lowest like rate belongs to a top-five video. trying out a new mating call @Tinder holds 11.6 likes per 1,000 views on 1,416,898 views and 16,486 likes. It ranks 5th of 34 on views and last of 34 on like rate.
Comments put a different video at the top again. tell me what you got in the comments! @webtoonofficial holds 3.60 comments per 1,000 views on 499,545 views and 1,800 comments, and it is the fourth-lowest of the 34 on views.
So the top of the view ranking and the top of the like-rate ranking are almost disjoint. The account’s highest like rate is a mid-table video by views. Its lowest like rate is a top-five video by views. Which of the two rankings a report uses changes which videos it calls successful, and the divergence is reported here without a cause attached.
What public counts cannot show
The three counters this dataset holds are the three a signed-out visitor can read. Everything a channel owner uses to judge a video sits behind YouTube Studio sign-in, which means it is absent here.
- Impressions. YouTube counts a thumbnail impression when the thumbnail is shown for more than 1 second with at least 50% of it visible, and reports impressions and click-through rate inside YouTube Studio to the channel owner (Check your YouTube thumbnail impressions and watch time, YouTube Help, retrieved 2026-08-28). No view-through rate or reach rate can be computed from public counts.
- Watch time and retention. YouTube reports watch time and average view duration in the Studio Engagement tab at video level (Get started with YouTube Analytics, YouTube Help, retrieved 2026-08-28). Public counts say nothing about whether any of these 47 videos held anyone.
- Traffic sources. YouTube reports how an audience discovered content in the Studio Reach tab, and states that traffic-source data may be limited even there (same source). Public counts cannot say where any of these views came from.
- Conversion. No public counter connects a view to an app install, a lesson, a subscription or a sale. This dataset carries no such number and neither does the public surface.
- Paid amplification. YouTube states that public views representing ad views are counted the moment a video ad starts to play across Shorts and in-stream ads (How engagement metrics are counted, YouTube Help, retrieved 2026-08-28). A promoted video and an organic one therefore look identical in this dataset. Any of the five videos in the top table could have carried paid support behind its view count, including the 5,718,146-view video at the top, and nothing in the public counts would reveal it.
- The counting definition changed inside the window. YouTube’s documented change of 2026-08-24 applies to views counted from that date onward, so all 47 videos accrued their last four days of views under it. This dataset cannot separate that effect from anything else.
- Anything internal to Duolingo. Goals, briefs, process, team and budget are not observable from public counters, and no sentence on this page infers them.
- Deleted or unlisted videos. Anything the account removed before 2026-08-28 is absent from the Shorts tab, so it is absent here, and its absence cannot be measured.
One brand does not generalize. This dataset is 47 videos from one account across 90 days, and an account with this one’s existing subscriber base generalizes least of all. Nothing here is a benchmark, a predictive model, or evidence about short-form video in general.
Two more limits belong on the record. The counts are a snapshot: every figure above was read on 2026-08-28 and has moved since, upward and unevenly. And the analysis is descriptive rather than causal. Four observable properties were tested and none of them separated the outcomes.
What we think this means
Everything above this heading is what was collected and computed. Everything below it is a reading of it, and a reader who disagrees has the same 47 rows we had.
The visible hits of an account are a poor description of its output, and this dataset puts a number on the gap. The top three of the 34 settled videos hold 28.7% of the views and the bottom 17 hold 29.3%. Anyone studying this account by looking at its most-shared video is studying about a sixth of what it did.
Copying what you can see from outside is the part that worries us most. The four properties visible on the public surface did not separate this account’s own best videos from its own worst. Someone who watches the top Short and copies the handle in the title, the hashtag, the credit line and the runtime has copied four things that did nothing to distinguish the top of this range from the bottom.
Single-metric reporting is the other thing this dataset argues against. The same 34 videos rank differently on views, on like rate and on comment rate, and a video that looks mid-table on one counter sits third on another. A report that carries one column decides in advance which videos count as wins.
The floor is the number we keep coming back to. The lowest-view Short in this window holds 363,062 views, and that is the part of a large account’s output nobody discusses. It is also why comparing your own numbers against a brand account is a trap: you are not comparing work, you are comparing subscriber bases you cannot see the size of.
What the dataset settles is the shape of one account’s 90 days. What it cannot settle is why any single video landed where it did, and on the evidence available from public counts we cannot tell. For more accounts and campaigns examined the same way, with the numbers that did not work included, see the rest of our post-by-post teardowns.
One column we collected and did not analyze
Upload date was collected for all 47 videos and used for one purpose only: computing each video’s age in days on 2026-08-28 for the censoring rule above. No analysis of publishing times, days of the week, or gaps between uploads was run, and none is reported here.
The 47 videos in full
Every figure on this page recomputes from this table. Counts as read on 2026-08-28. Window 2026-05-30 to 2026-08-28. “Settled” marks the 34 videos at least 21 days old on the collection date.
Titles are reproduced as published, with zero-width characters removed and emoji dropped for legibility. Illustrator credits appear in the original titles as a paint-palette emoji followed by a handle, rendered here as “(art: @handle)”.
Questions
How many Shorts did the channel publish in the window?
The official Duolingo YouTube channel published 47 Shorts with upload dates inside the window 2026-05-30 to 2026-08-28, all counted from the channel’s Shorts tab on 2026-08-28. The earliest upload present is 2026-06-01 and the latest is 2026-08-27.
Why are 13 videos left out of the best and worst comparison?
The 13 videos were less than 21 days old on 2026-08-28, so their view counts had not finished accumulating when they were read. The four lowest counts in the full 47 belong to the four youngest videos, at 1, 2, 3 and 4 days old. Ranking those against a 12-week-old video would measure age rather than reach.
Why 21 days and not 30?
The 21-day cutoff is our choice, and the dataset shows the choice barely matters. No video in the dataset is between 18 and 22 days old, so any cutoff from 18 through 23 days keeps the same 34 videos. The highest-view and lowest-view videos are also the same two videos at cutoffs of 14, 21 and 28 days, while the subset median moves from 834,918 to 817,268 to 789,070.
What was the lowest-view Short in the window?
The lowest-view Short in the settled subset is consider this a formal petition to add this category to @LasCulturistas awards, at 363,062 views as read on 2026-08-28. The full 47-video minimum is 61,788 views, and it belongs to a video that was two days old on the collection date, so it is not comparable with the rest.
Did shorter videos get more views?
Shorter videos in this subset had a slightly higher median, and the split does not separate the highest from the lowest. The 20 settled videos of 10 seconds or under have a median of 851,228 views against 756,292 for the 14 videos over 10 seconds, a gap of 12.6% on a subset whose extremes differ by a factor of 15.75. Duration and views have a rank correlation of minus 0.22 across the 34 videos.
Did tagging another account in the title help?
Tagging another account did not raise the median in this subset. The 9 settled videos whose titles carry an @ handle have a median of 789,070 views against 834,918 for the 25 with no handle. Four of those 9 titles credit an illustrator rather than tag another account, and separating them leaves 5 tagged videos, which is too few to support any finding.
Which Short got the most likes for its size?
congrats it’s a streak! (now go to the app and do it for real) holds the highest like rate in the settled subset, at 52.5 likes per 1,000 views on 967,997 views and 50,779 likes. It ranks 11th of 34 on views and 3rd of 34 on like count. Likes per 1,000 views is arithmetic on two public counts and not a platform engagement rate.
Can I compare these numbers with my own account?
You cannot compare them usefully. The dataset covers one account with an existing subscriber base you cannot see from outside, it has no impressions denominator, and it cannot separate paid amplification from organic views because YouTube counts ad views into the public view count. What you can reuse is the method: count every post in a window, censor the ones too new to have settled, and read the distribution rather than the average.
Will these numbers still be right when I read this?
The numbers will be lower than what you see, because view counts accumulate. Every figure on this page was read on 2026-08-28 and is labeled that way for exactly this reason. A higher count when you check is not a correction to this page, and if you recount the same 47 videos today the distribution and the concentration shares are the figures worth comparing rather than the absolute counts.
<!– BODY THAT SHIPS ENDS HERE –>
Source ledger
Every URL below was fetched in this session on 2026-08-28 with curl -L --max-time 20, a desktop browser user agent, and --write-out '%{http_code}' with the body discarded. Only 200-status URLs are cited in the body.
Primary sources cited in the body
| URL | Fetched | HTTP status | What it supports |
|---|---|---|---|
| https://www.youtube.com/@duolingo | 2026-08-28 | 200 | The account under analysis, named and linked once. |
| https://support.google.com/youtube/answer/2991785?hl=en | 2026-08-28 | 200 | Server-rendered article body read in full. Supports: views counted the moment a video starts to play from 2026-08-24 across all formats including Shorts; metrics take time to appear in the first few hours after publication; metric counts may be slowed, frozen or changed while legitimacy is confirmed; public views representing ad views are counted into the public view count across Shorts and in-stream ads. |
| https://support.google.com/youtube/answer/9314486?hl=en | 2026-08-28 | 200 | Server-rendered article body read in full. Supports: thumbnail impressions counted at over 1 second with at least 50% visible; impressions and click-through rate reported inside YouTube Studio to the channel owner; the same “views counted the moment a video starts to play” definition. |
| https://support.google.com/youtube/answer/1714323?hl=en | 2026-08-28 | 200 | Server-rendered article body read in full. Supports: watch time and average view duration in the Studio Engagement tab; audience discovery in the Studio Reach tab; YouTube’s own note that traffic-source data may be limited. |
| https://blog.recurpost.com/category/teardowns/ | 2026-08-28 | 200 | The parent category for this page. |
The 47 video URLs
All 47 https://www.youtube.com/shorts/<id> URLs in the dataset were fetched individually on 2026-08-28. All 47 returned HTTP 200. No 404, no 410, no 403, no 000. Full list in the 47-row table above; every video named anywhere on the page is linked to its own URL. Spot check without following redirects also returned 200.
Fetched, not cited
| URL | Fetched | HTTP status | Note |
|---|---|---|---|
| https://www.linkedin.com/in/dineshai/ | 2026-08-28 | 999 | LinkedIn’s block code for automated requests. The host refuses non-browser traffic, so the URL could not be resolved to 200 in this session. It appears only in the author’s Person schema sameAs, not as a body link. |
| https://support.google.com/youtube/answer/10059070?hl=en | 2026-08-28 | 200 | “Get started creating YouTube Shorts”. Reachable, deliberately not cited: it is Shorts production guidance, which this page does not carry. |
| https://blog.youtube/news-and-events/ | 2026-08-28 | 200 | Reachable, not cited. No specific dated post was resolved from it in this session, so nothing was drawn from it. |
| https://blog.recurpost.com/ | 2026-08-28 | 200 | Publication root, checked only to confirm the site resolves. Not linked in the body. |
Sources deliberately not used
- No third-party analytics site, view estimator, statistics aggregator or benchmark report was fetched or cited. Every count on the page comes from the collection described above, published in full in the 47-row table.
- No link to recurpost.com appears anywhere in the body; this page was commissioned with no product link.
