Turn your YouTube watch history into a searchable knowledge base
You have already done hundreds of hours of research. It is sitting in a history file you have never opened, in a format that cannot answer a single question.
The short answer: export your history from Google Takeout, filter it down to the five to fifteen percent that is genuinely research, group those videos into topic corpora of fifteen to twenty-five, and run structured extraction over each corpus. The output is searchable notes with source links rather than a list of URLs you will never revisit. As of July 2026, the extraction tooling runs roughly $19-$199/month; the export itself is free.
The premise worth sitting with: if you watch technical content seriously, you have probably put several hundred hours into a subject over a couple of years. That is a graduate-level time investment stored in a format with no retrieval mechanism. You cannot ask it anything.
Step 1 — Export, and take only what you need
Google Takeout exports YouTube history with titles, channel names, and timestamps. Request the YouTube history subset specifically rather than a full account export — the difference is a manageable file versus several gigabytes of things you do not want.
{ "title": "Watched <video title>",
"titleUrl": "https://www.youtube.com/watch?v=...",
"subtitles": [{ "name": "<channel>", "url": "..." }],
"time": "2026-03-14T19:22:08Z" }Keep the raw file local. Everything downstream operates on a filtered subset, and there is no reason for the unfiltered version to leave your machine.
Step 2 — Filter hard, because most of it is noise
A typical history is overwhelmingly entertainment, music, and things watched for thirty seconds. Three filters in order do most of the work:
| Filter | Rule of thumb | Why it works |
|---|---|---|
| Duration | Drop anything under ~10 minutes | Research-grade content is long-form; shorts almost never are |
| Channel | Keep a short allow-list of channels you learn from | Your research sources are far fewer than they feel |
| Title keywords | Keep titles containing your topic vocabulary | Catches one-off videos from channels not on the list |
Expect to end up with a few hundred videos from tens of thousands of rows. That is not a failure of the filter; it is an accurate measure of how much of watching is actually research.
Videos appearing more than once in your history are worth flagging. You returned to them, which means they contained something you needed twice — a reliable marker of the genuinely useful subset.
Step 3 — Group into topic corpora
This is the step that determines whether you end up with a knowledge base or a pile. Processing four hundred unrelated videos as one blob produces a synthesis about nothing in particular, because cross-video comparison only means something within a topic.
Sort the filtered set into coherent groups of fifteen to twenty-five — one per subject you were actually researching. Small leftovers that fit nowhere are usually genuine one-offs; leave them out rather than contaminating a corpus to reach a round number.
This is the same corpus-selection discipline used when researching a new topic from scratch, described in extracting product ideas from YouTube automatically — the only difference is that here you already watched them.
Step 4 — Extract structure, not summaries
A knowledge base made of prose summaries is barely better than the original videos. What makes it queryable is consistent structure: the same fields extracted from every video, so you can search across them and count.
- Claims made — with the source video and timestamp
- Tools and techniques named — the vocabulary of the space
- Workflow steps — what was actually done, in order
- Open questions — what the video raised and did not resolve
- Contradictions — where this source disagrees with another
Attribution on every extracted item is non-negotiable. A knowledge base you cannot verify decays into a set of half-remembered assertions with the same confidence as the real ones — the failure mode described in the timestamps-and-citations test.
Which tool you use for the extraction matters less than whether it preserves structure and attribution; the category differences are laid out in the 2026 comparison of YouTube research tools.
Step 5 — Query it like a colleague
The point of the whole exercise is asking questions you cannot ask a bookmark list:
- Who disagreed about the right approach to this, and where?
- Which tools kept coming up that I never tried?
- What did people say about this six months ago versus now?
- Which questions did nobody in my sources ever answer?
That last one is the most valuable. Unanswered questions across a corpus you personally curated are the gaps in your own field of view — and often the gaps in the field itself, which is exactly the analysis in finding underserved niches through content gaps.
The knowledge base pays off most when a topic you researched a year ago suddenly becomes relevant again. Rebuilding that context from scratch is days of work; querying it is minutes — and you already paid the watching cost.
Step 6 — Keep it alive
A knowledge base built once and abandoned is a time capsule. The maintenance loop is small: add new videos to the relevant topic project as you watch them, and re-run synthesis periodically so the summary reflects current understanding rather than a frozen January snapshot.
Re-running also surfaces something a static archive never will — where your own sources have started contradicting their earlier positions. Volume tiers exist for exactly this shape of use: two standing projects on the entry plan, eight on the middle, twenty on the largest.
Who this is actually worth doing for
The process is an afternoon, so it is worth being honest about who gets a return on it. Four profiles where it pays clearly:
| Profile | What they get out of it |
|---|---|
| Builders researching a space | A year of watching becomes a product thesis with sources attached, instead of a vague sense that you know the area |
| Consultants and analysts | Client questions answered from your own curated corpus with citations, rather than a fresh search each time |
| Career switchers | An honest map of what you have covered and what you have never seen explained — the gaps are the study plan |
| Anyone writing publicly | Source material with attribution ready, which removes the worst part of writing: reconstructing where you learned something |
And two cases where it is not worth it: if your watching is mostly entertainment (the filter will leave you with thirty videos and no corpus), or if you are researching something so fast-moving that anything older than six months is misleading. In the second case, build the knowledge base from fresh videos only and let the history go.
The temptation is to process the whole export because storage is cheap. Resist it. A knowledge base's value comes from the ratio of signal to volume — a corpus of twenty carefully chosen videos answers better questions than four hundred indiscriminate ones, and costs less to process.
Group your filtered history into topic projects, get structured notes per video and a synthesis across each corpus, then chat with the whole knowledge base. Export everything as markdown whenever you want it. 7-day free trial.
Closing thought
Nobody sets out to build a research archive by watching YouTube. But if you have been serious about a subject for a year or two, you built one anyway — just in a format optimized for the platform's recommendations rather than your recall.
Converting it costs an afternoon of filtering and a processing run. And once a topic corpus is structured, it stops being a personal archive and becomes an input: the same material that answers what do I already know also answers what should I build, which is where the seven-day playbook picks up.
Frequently asked
How can I turn my YouTube watch history into a knowledge base?
Export your history from Google Takeout, filter it down to the small subset that is actually research, group those videos into topic corpora, and run structured extraction over each corpus so the output is searchable notes with source links rather than a list of URLs. The filtering is the important step — most watch histories are 90% entertainment. Tooling for the extraction runs roughly $19-$199/month as of July 2026.
Where do I get my watch history data?
Google Takeout exports YouTube history as JSON or HTML including titles, channels, and timestamps. Request only the YouTube history subset rather than everything; the full export is large and mostly irrelevant.
How much of a watch history is worth keeping?
Typically five to fifteen percent. Filter by duration first — anything under about ten minutes is rarely research — then by channel and title keyword. The remainder is usually a manageable few hundred videos across several topics.
Should I process the whole history at once?
No. A single undifferentiated pile produces a knowledge base with no structure. Group filtered videos into topic corpora of fifteen to twenty-five and process each separately, so cross-video synthesis has something coherent to compare.
What does a knowledge base give me that bookmarks do not?
Bookmarks are addresses; a knowledge base is content you can search and question. The difference shows up when you need something you cannot name — you remember the idea but not which video it was in, which is exactly the query a bookmark list cannot answer.
Can I keep adding to it over time?
Yes, and that is where it becomes valuable. Add new videos to the relevant topic project as you watch, and re-run synthesis periodically so the knowledge base reflects what you know now rather than what you knew in January.
Is there a privacy concern with uploading watch history?
Your history is personal data, so only send the filtered subset you actually want analyzed rather than the whole export, and drop anything unrelated to the research topic before processing. Keep the raw Takeout file local.