YouTube channel research with an LLM:
what makes a developer video take off?
A real run: a big developer channel's 50 latest videos and transcripts, tagged by an LLM and compared, for $0.53, and why the obvious answer was wrong.
Creators and marketers ask the same question about every successful channel: what do its best videos have in common? With view counts, transcripts and a model to read them, the answer seems a few minutes away. We ran it on Fireship, a large developer-news channel, and got an answer that looked convincing and was mostly wrong, until we compared videos of the same age. Everything below is from that run on 26 October 2026, for $0.53.
The question
Across the channel's latest 50 videos, which topics, formats, lengths and titles go with more views?
The data we collected
| Source | Tool | What we asked for | Results | Cost |
|---|---|---|---|---|
| YouTube | YouTube channel videos | The channel's 50 latest videos: title, views, likes, comments, length, date | 50 videos | $0.063 |
| YouTube | YouTube transcripts | The transcript of each | 50 transcripts | $0.124 |
The videos were 2 to 187 days old, all between about 4½ and 11½ minutes long, with a median of 1.07 million views.
How the model was used
claude-sonnet-5.5 read each title and transcript in one call and tagged the topic (AI models and tools, AI business, security incidents, web frameworks…), the format (weekly news, explainer, comparison, opinion), and how the first sentences grab attention, with the opening line copied exactly (code checked all 50). That single call, over 50 transcripts, was most of the cost: $0.34. Code did the rest.
The answer that looked right
Comparing median views across the 44 videos at least two weeks old, the pattern looked clear:
(21 videos): 1.23 million, the highest of any topic. Web framework videos: 0.70 million.
(15 videos): 1.28 million, against 0.93 million without.
(9 videos): 1.92 million, against 1.0 million for 4 to 6 minutes.
That's a tidy brief: make AI videos, put a number in the title, run a bit longer. Before writing it, we checked one more thing.
What was really going on
The channel took off in September. Its monthly median was 0.8 to 0.9 million views from May to August, then 2.17 million for the 12 videos published in September. Across all 44 videos, newer ones had more views than older ones, not fewer: a rank correlation of −0.65 between age and views. And the "winning" groups were the recent ones: AI-model videos had a median age of 54 days against roughly 95 for most other topics.
So we compared videos of the same age. Among the 17 published 14 to 60 days before the run:
| Comparison (videos 14–60 days old) | With | Without |
|---|---|---|
| AI models and tools topic | 1.88 M (11 videos) | 2.14 M (6) |
| A number in the title | 1.64 M (8) | 2.35 M (9) |
| 6 to 10 minutes long | 2.58 M (6) | 1.77 M (11) |
| History or "how X works" explainer | 1.33 M (3) | 2.14 M (14) |
The topic and title effects vanish, and even reverse. Two patterns hold up in both the 14–60-day and the 14–90-day windows: 6 to 10 minute videos beat shorter ones, and explainers trail weekly news. Both rest on a handful of videos, so they're hypotheses to test, not rules.
The house style
Some things don't need a comparison because they are nearly constant, which is a finding in itself:
40 of the 44 titles end in "..." ("OpenAI's biggest math breakthrough is getting ugly...").
27 of the 50 videos open with breaking news ("Last week a brand new uncensored model entered the AI race."), and 13 with a joke or an absurd image.
No video was under 4 minutes; 33 of the 44 ran 4 to 6.
Lessons
Views keep accumulating, and channels grow or shrink, so the oldest and newest videos aren't comparable. Compare within an age window, or use views per day since publishing.
"Comparison videos get 2.1 million" was true of 2 videos. Print the count next to every median.
The model was reliable at reading transcripts and naming formats; the mistakes we nearly made were all in the counting.
Run it yourself
With an API key in LOVELY_API_KEY and the run_tool and ask helpers from the market research case study:
from datetime import datetime, timezone
videos = run_tool("youtube-channel", {"youtubeHandles": ["@Fireship"], "maxItems": 50}, 50)
scripts = run_tool("youtube-transcript", {"startUrls": [v["url"] for v in videos]}, 50)
text = {s["id"]: " ".join(x["text"] for x in s["transcript"]) for s in scripts}
tags = json.loads(ask("Tag each video's topic, format and opening hook. JSON array of {id, topic, format, hook}.",
[{"id": v["id"], "title": v["title"], "transcript": text.get(v["id"], "")[:4000]} for v in videos]))
now = datetime.now(timezone.utc)
for v in videos:
v["age_days"] = (now - datetime.fromisoformat(v["publishDate"])).days
# Compare medians only within an age window (say 14-60 days), and print the count beside each.Swap the handle for any channel, including yours. For transcripts without code, try the free YouTube transcript generator; for the tools behind it, see research agents.
Frequently asked questions
How much does it cost to analyse a YouTube channel with AI?+
This run cost $0.53: $0.19 for 50 videos' data and transcripts, and $0.34 for one call to a strong model reading all 50 transcripts. A cheaper model, or only the first few minutes of each transcript, would cut the model cost several times.
Can the model tell me why a video did well?+
It can describe what a video is and how it opens, consistently and at scale. Whether that caused the views is a statistics question, and the honest answer usually needs more videos and a fair comparison, as this run shows.
Can I compare several channels?+
Yes. Run the channel tool for each handle, tag them the same way, and compare within the same age window. With competitors' channels it becomes content research for your own.
Do I need the transcripts?+
For topic and format, the title often suffices. For how a video opens, its pacing or what it actually covers, you need the transcript, and at about a quarter of a cent each it's rarely the expensive part.