Most "best YouTube channels" lists are written from memory. We had something different to work with: the lessons that real learners took inside LearnPath (learnwithpath.com) between March and early October 2026. Every lesson in a path is a YouTube video that our system searched for, read the transcript of, and judged against one specific skill at one specific level.
That gave us a library of 2,151 lessons from 1,289 channels, used in 1,505 learning paths by 1,340 learners, plus the search logs from 54,168 YouTube results our system looked at to pick them. This post is what that data says about learning from YouTube, with the caveats spelled out at the end. It follows our June look at what people actually learn from YouTube, which covered 453 paths; this one has more than three times the data and adds the search logs and how far learners watched.
Key findings
Short version, before the detail:
- Fewer than 1 in 5 search results made the shortlist. Of 54,168 results checked, 10,093 (19%) were good enough to be considered as a lesson.
- The lessons that teach are short. Median length is 13 minutes, and 71% of lessons are under 20 minutes.
- Learners finish short lessons almost twice as often. About 24% of started lessons under 20 minutes were watched to the end, against 13% for 30 to 60 minute lessons.
- Half the lessons are more than three years old. For stable basics like Python and SQL it is 62%. For AI tools it is 15%.
- Small channels carry the library. 80% of channels contributed a single lesson. The 20 most used channels supply only 18% of lessons.
- Learners passed about 7 in 10 first quizzes on the first try, whatever the lesson length.
The rest of this post goes through each one.
How we got the data
LearnPath builds a learning path from YouTube for whatever someone wants to learn. For each step, it runs a few YouTube searches, reads the transcripts of the candidates, and asks an AI reviewer a narrow question: does this video teach this skill, at this level, as a lesson rather than a trailer, a vlog or an ad? The shortlisted videos are ranked and one becomes the lesson. After the lesson, the learner takes a short quiz written from the video's own transcript.
So every number below comes from one of three places:
- Search logs. How many results each path build looked at, how many were rejected, and how many reached the shortlist. These counters exist for paths built since July 2026 (650 paths).
- The lesson library. Every video that ended up as a lesson in a path that was not deleted: its length, publish date and channel.
- Learner activity. How far learners watched (475 learners started at least one lesson) and whether they passed the quiz on the first try (433 first attempts).
This is not a random sample of YouTube. It is a sample of what a strict filter keeps, which is exactly why it is useful if you are trying to learn something from YouTube yourself.
Finding 1: fewer than 1 in 5 search results made the cut
The median path build looked at about 100 YouTube results. Across all 650 builds with counters, the totals were:
| Stage | Videos | Share of results |
|---|---|---|
| Search results checked | 54,168 | 100% |
| Rejected by the filters | 27,644 | 51% |
| Shortlisted as possible lessons | 10,093 | 19% |
The remaining results were neither rejected nor shortlisted: once a build had enough strong candidates, it stopped looking at the rest. So 19% is not "the share of YouTube that is good", it is the share that survived a strict, skill-specific check while the build still needed candidates.
The filters check simple things. Is the video about this skill or a neighbouring one? Is it pitched at the right level? Is it a lesson, or a channel trailer, a course advert or a reaction video? Does it teach the skill, or only talk about it?
What this means for you: the first page of YouTube results is a starting point, not a recommendation. Before you commit 20 minutes, skim the first two minutes and ask whether the video is teaching you to do one specific thing. If it is still introducing itself at minute two, move on.
Finding 2: the lessons that teach are short
Here is how the 2,151 lessons split by length:
| Length | Share of lessons |
|---|---|
| Under 10 minutes | 34% |
| 10 to 20 minutes | 37% |
| 20 to 30 minutes | 14% |
| 30 to 60 minutes | 10% |
| Over 60 minutes | 5% |
The median lesson is 13 minutes long. Long full courses, the multi-hour videos that dominate many recommendation lists, rarely won a step: only 5% of lessons run past an hour. That is not because they are bad. It is because a single step needs one skill taught well, and a long course spreads that skill across an hour of other material.
The channel data shows the same pattern. freeCodeCamp's videos in our library average over three hours, and they were used for 48 lessons across 16 topics. Bro Code's average 12 minutes, and they were used 296 times. Both channels are excellent. Only one of them fits a 15 minute study session.
What this means for you: if you have a long course bookmarked, use its chapter markers and treat each chapter as its own lesson. Our post on how to learn anything from YouTube walks through doing this.
Finding 3: learners finish short lessons almost twice as often
For lessons that a learner actually started, we compared how much of the video they watched:
| Length | Lessons started | Average share watched | Watched to the end |
|---|---|---|---|
| Under 10 minutes | 211 | 50% | 24% |
| 10 to 20 minutes | 236 | 43% | 24% |
| 20 to 30 minutes | 153 | 38% | 19% |
| 30 to 60 minutes | 60 | 39% | 13% |
Two things stand out. First, completion drops as lessons get longer, from about a quarter of lessons under 20 minutes to about an eighth past half an hour. Second, even short lessons are often left before the end. Some of that is learners skipping parts they already know, and some is ordinary drop-off. Either way, a 12 minute video has a much better chance of being finished than a 45 minute one.
Interestingly, quiz results barely changed with length. First-try pass rates were 68% to 75% in every length band. Learners who got to the quiz learned about as well from a short lesson as from a long one. The short lesson simply got more of them there.
What this means for you: plan study sessions around lessons you will finish. Two 12 minute lessons beat one 45 minute lesson you abandon halfway.
Finding 4: half the lessons are more than three years old
This one surprised us. Across the library:
- 50% of lessons were published more than three years ago.
- 27% were published more than five years ago.
The age depends heavily on the topic:
| Topic area | Lessons | Older than 3 years | Median length |
|---|---|---|---|
| Stable basics (Python, SQL, Excel, networking, math) | 413 | 62% | 13.5 min |
| AI tools (AI, LLMs, agents, prompting) | 251 | 15% | 16.9 min |
A loop in Python, a JOIN in SQL or a subnet mask has not changed in years, and the clearest explanation of it may well be from 2019. An AI agent framework from 2023, on the other hand, is often a different product today.
What this means for you: do not filter YouTube by "this year" for everything. For fundamentals, an older video with a clear explanation beats a newer one with a worse one. For fast-moving tools, check the date and the version number shown on screen before you start. Our lists of the best YouTube channels for Python and the best YouTube channels for SQL lean on long-running channels for exactly this reason, while the AI engineering list favours channels that update often.
Finding 5: small channels carry the library
The 2,151 lessons came from 1,289 different channels.
- 1,036 channels (80%) contributed exactly one lesson.
- The 20 most used channels account for 382 lessons, 18% of the library.
When you search for a specific skill rather than a broad topic, the best explanation often comes from a channel you have never heard of: a teacher who made one great video about one thing. Following a single big creator for a whole subject means missing most of those.
What this means for you: subscribe to a few strong channels for a subject, but search by skill when you get stuck on a specific concept. "How to use window functions in SQL" will surface better lessons than browsing a favourite channel's uploads.
Which channels show up most, by area
These are the channels whose lessons were used by the most learners in each area. They are counts from our library, not a quality score, and they reflect which topics our learners chose.
| Area | Channels used by the most learners |
|---|---|
| Python | Bro Code, Programming with Mosh, Corey Schafer, Indently, Visually Explained |
| AI and machine learning | Tech With Tim, Ryan & Matt Data Science, codebasics, IBM Technology, Dave Ebbelaar |
| Cybersecurity | Hacker Joe, NetworkChuck, PowerCert Animated Videos, TechTerms, Bogdan Stashchuk |
| System design | ByteByteGo, PowerCert Animated Videos, Ashish Pratap Singh, Gaurav Sen, ByteMonk |
| Data analysis and SQL | Maven Analytics, Alex The Analyst, Adam Finer, Bro Code, Neso Academy |
| Web development | Coding2GO, Bro Code, EdRoh, Kevin Powell, CodeLucky |
A few notes on reading this table:
- Visual explainers punch above their size. PowerCert Animated Videos shows up in both cybersecurity and system design, because short animated explanations of networking and infrastructure basics fit early steps well.
- Single-video channels appear too. Bogdan Stashchuk and Adam Finer each placed one or two videos that many learners used, which is Finding 5 in action.
- The usual names are there for a reason. ByteByteGo, NetworkChuck, Corey Schafer and Alex The Analyst lead in their areas in our data too.
For full, hand-checked rankings with a starting video for each channel, see our guides to the best YouTube channels for system design, the best YouTube channels for cybersecurity and the best YouTube channels for AI engineering.
What this means if you are learning from YouTube
Putting the five findings together, a few practical rules:
- Search by skill, not by subject. "Python list comprehensions" beats "Python course". Specific searches surface the short, focused lessons that people actually finish.
- Give a video two minutes. If it has not started teaching by then, it probably will not. Four out of five search results did not survive a careful check.
- Prefer 10 to 20 minutes. That is where most good lessons sit, and where learners finish most often.
- Judge age by topic. Old is fine for fundamentals and risky for fast-moving tools.
- Test yourself after each lesson. Learners passed about 70% of first quizzes on the first try, which means about 30% of the time, watching a lesson did not mean having learned it. A few questions right after the video show you which case you are in.
If you would rather not do the filtering yourself, that is the job LearnPath does: it builds the path from YouTube, one short lesson per step, with a quiz after each one. If you are comparing paid course platforms with learning from YouTube, our look at Coursera alternatives for self-learners covers the trade-offs.
Limits of this data
We want this to be useful to cite, so here is what it does and does not show.
- It is filtered data. Every lesson was chosen by our system. The library shows what survives a strict, skill-specific filter, not what YouTube as a whole looks like.
- The learners are our learners. Most of them are learning tech skills such as programming, data and cybersecurity. A different audience would pick different topics.
- Watch data is partial. Watching is tracked only inside our player, so a learner who opened the same video on YouTube directly is not counted. The completion numbers are best read as a comparison between lengths, not as exact rates.
- Search counters start in July 2026. The funnel in Finding 1 covers the 650 paths built after that, not the full period.
- Channel counts are popularity, not quality. A channel used by many learners was picked often for popular topics. It does not mean a smaller channel teaches worse.
Numbers were pulled on October 7, 2026. If you would like the underlying tables for your own article or research, we are happy to share them.
Frequently Asked Questions
How many YouTube search results are actually good enough to learn from?
In our data, fewer than 1 in 5. Across 650 learning paths our system looked at 54,168 search results and put 10,093 of them, about 19%, on a shortlist. The rest were off topic, at the wrong level, too thin, or a playlist trailer instead of a lesson. Picking one good lesson from a page of results takes real filtering.
How long should a YouTube lesson be?
The lessons that made it into learning paths had a median length of 13 minutes, and 71% were under 20 minutes. Learners also finished short lessons more often: about 24% of started lessons under 20 minutes were watched to the end, against 13% for lessons of 30 to 60 minutes.
Is it a problem to learn from an old YouTube video?
For stable skills, usually not. Half of all the lessons in our library are more than three years old, and for stable basics like Python, SQL, Excel and networking the share is 62%. For AI tools it drops to 15%, because the tools change every few months. Check the topic before you dismiss a video for its age.
Which YouTube channels show up most in learning paths?
Across all topics, Bro Code, Programming with Mosh, Indently, Tech With Tim and Corey Schafer appear most often. By area the leaders change: ByteByteGo and Gaurav Sen for system design, Hacker Joe and NetworkChuck for cybersecurity, Maven Analytics and Alex The Analyst for data analysis.
Do you need to watch big channels to learn well?
No. 80% of the 1,289 channels in our library contributed exactly one lesson, and the 20 most used channels account for only 18% of all lessons. For a specific concept, the best explanation is often a single video from a small channel, which is why searching by skill beats subscribing to one creator.
