How we turn YouTube data into decisions with the YT.agency Extension
Short answer: The YT.agency Extension is a private research tool, used only by yt.agency, that collects public YouTube data into structured files: 44 fields per video, 19 per comment, 14 per transcript, plus channel snapshots over time. Its value is not the data. It is that decisions which used to rest on opinion (which topic to make, which format to stop copying, which words the audience uses) can be checked against a dataset before production money is spent. We have not measured the exact hours or dollars this saves, and we show below how to measure it.
What the tool is, and is not
- Private. It is built for yt.agency's own client work. It is not sold or licensed to companies or individuals.
- Not an official YouTube product. It does not use the YouTube Data API or OAuth sign-in. It reads public pages, public captions and the same public comment feed a viewer's browser loads. Anyone running similar tooling should review YouTube's Terms of Service for their own use.
- Mostly local. The browser extension works with a small engine that runs on the operator's own machine and fetches public captions. Exports are files on that machine. (Dislike estimates are the one input that comes from an outside community dataset; see the caveat below.)
- Limited on purpose. It deliberately does not show estimated earnings or global channel rank. Earnings would be views multiplied by a revenue rate we cannot know from outside, and rank needs a proprietary database we do not have. A number we would have to invent does not go into a client report.
The data it collects: types and volume
All counts below are read from the tool's own source code as of October 2026.
| Dataset | Fields per row | What is in it |
|---|---|---|
| Video export | 44 | Title, URL, duration, views, upload date, likes, comment count, tags, thumbnail URL, channel name and subscribers, content type (long-form, Short, live, podcast), description, derived rates (engagement, like-to-view, comment-to-view, view-to-subscriber), title and description length, tag, hashtag and link counts |
| Comments export | 19 | Author, text, likes, date, reply and parent links, a heuristic category tag, comment ID, reply count, whether it is pinned, hearted or from the channel owner, sort order used, extraction time |
| Transcript export | 14 | Language, whether auto-generated, full text, timestamped text, word and segment counts, status and error for each video |
| Channel tracking | 3 exports | History (views over time), Changes (title and thumbnail changes with before and after), Growth (subscriber growth), plus "velocity movers" |
| Research | Up to 8 and 30 | Up to 8 similar channels and up to 30 similar videos per search, each with the evidence used to match it |
Derived metrics, and how they are computed
- VPH is views per hour since upload.
- Outlier multiple is a video's VPH divided by the median VPH of the same channel and the same format. Shorts are measured only against Shorts and long-form against long-form. If a format group has fewer than three videos, the tool falls back to the channel-wide median and caps the confidence at medium.
- Metric confidence is high, medium or low depending on how many underlying fields were actually captured (high needs most of them). It tells us how far to trust a row.
- Comment categories. Comments are pre-tagged into ten heuristic categories, in English and Egyptian Arabic: problem, question, complaint, praise, confusion, objection, request, testimonial, skepticism and noise. These are a first pass for a human or an analysis prompt to check, not a verdict.
How much data per run
- Comments. All comments and replies for a video are collected by paging through YouTube's comment feed until it ends, with no 1,000-comment cap. Choosing newest-first gives the most complete set; the ranked "top" feed can omit some comments.
- Batches. Comment and transcript extraction run across a filtered list of videos, several at a time, with a live count, a stop button, and per-video failure handling (a video with comments disabled is recorded and skipped, not allowed to stop the batch).
- Throttling. Requests start at 8 in parallel and back off to as few as 2 when YouTube rate-limits, then recover. A slower run is the price of not hammering the platform.
- Snapshots. Tracked channels are recorded about every 60 minutes by default (never more often than every 10). A comment in the code estimates one snapshot at roughly 18 KB for an 80-video channel. At the default interval that is 18 KB × 24 = about 432 KB per day, or about 13 MB per month for that channel. That is an estimate from a code comment, not a measured figure.
- Scale. The code has been tuned for datasets of thousands of videos (one performance fix refers to a 3,000-video build). We have not published a benchmark.
The pipeline: data first, then judgment
Our order of work is fixed: Extension, then spreadsheet, then analysis against our documented procedures, then a stored record of what we learned. We do not ask an AI model to analyze a channel it has no data for. If the dataset is missing or thin, the output says so.
Five decisions the data changes
Each row is a decision we make, the data behind it, and the cost it can avoid. The last column is a mechanism, not a measured result.
| Decision | Data used | Cost it can avoid |
|---|---|---|
| Which video ideas to make and which to drop | Video exports with VPH and outlier multiples for the client's channel and peers | Producing a video on a topic with no demonstrated demand |
| What to repeat from the channel's own history | The channel's full video export, grouped by topic and format | Repeating formats that underperformed for this channel |
| What the audience actually struggles with | Comment export, weighted by likes and replies, spam removed | Writing hooks and scripts in language the audience does not use |
| Where a competitor is weak | Competitor titles, formats, transcripts and tracked changes side by side | Entering a crowded topic with no angle of our own |
| Which questions to answer in a video | Repeated questions in comments, tagged as question or problem | Missing the question a buyer asks before buying |
Cautions we apply to every dataset
- Comments are a self-selected sample. People who comment are not all viewers. We treat patterns as signals about the vocal audience and say so.
- High views are not buyer value. A topic can pull views from people who will never buy. We check outliers against the client's buyer.
- Small samples mislead. If a dataset is too small or too repetitive, our rule is to report "data is insufficient for strong conclusions."
- Dislike figures are estimates. YouTube no longer shows public dislike counts, so the tool uses an outside community estimate and the confidence score reflects whether it was available.
What this looks like in a real niche
Our experience: Accountack serves Arab business owners and investors on US tax topics, and the speaker mixes Egyptian Arabic with English terms like IRS, W-2 and 1099. Auto-generated captions often write those English terms in Arabic letters. Before any analysis, we correct the transcript so the English terms appear as terms, because a pattern search on garbled text finds nothing. We use this process on that channel system; we do not attribute its results (2.1M+ views, 47.8K+ watch hours, 11.1K+ subscribers) to the tool alone. See the Accountack case study.
How to measure the time and money saved
We would rather give you a method than a number we cannot source:
- Time. For one research task (for example, reviewing 30 competitor videos), log the hours spent manually and the hours spent with a batch export plus analysis. Do this on at least three tasks.
- Money. Cost avoided = (cost to produce one video) × (number of ideas dropped before production because the data showed no demand). Count only ideas you can show were dropped for a data reason.
- Compare to outcomes. Track whether the videos you did make performed better than the channel's earlier baseline. One channel is not proof; treat it as one data point.
Until measured on your own work, treat any quoted saving, including ours, as unproven.
Evidence limits
- This article describes our own tooling. It is not an independent evaluation.
- Field counts, limits and formulas were read from the source code in October 2026 and may change as the tool is updated.
- We have not run a controlled comparison of manual versus tool-assisted research, so we state no time or cost savings.
- Outputs depend on what public data is available and on sample quality.
- The decisions table shows mechanisms, not guaranteed outcomes.
Sources
- YouTube Analytics API, official metric definitions (for the standard metrics the tool's derived rates build on). developers.google.com/youtube/analytics/metrics
- YouTube Terms of Service. youtube.com/t/terms
FAQ
Is the Extension available to buy or use? No. It is a private tool used only by yt.agency for client work. It is not sold or licensed to companies or individuals.
Does it use the YouTube API? No. It does not use the YouTube Data API or OAuth, and it is not an official YouTube product.
Does it replace YouTube Studio? No. Studio and the YouTube Analytics API remain the source for a channel's own performance data; the Extension supports competitor, comment and transcript research.
Can I do this without the tool? Yes, manually, but expect more hours on collection. Measure it with the method above rather than trusting a claimed saving.
Where does this fit in your process? It is the data stage of our operating system, which feeds topic, packaging and script decisions described in our approach.