# How to research a topic on YouTube

_By Nakul Kelkar, September 28, 2026_

A YouTube research brief should show which videos you selected, what you read or watched, and where each observation came from. Search results are useful for choosing the videos; they are not enough to summarize an entire argument.

This workflow uses Vaaya to discover candidates and retrieve available supporting data. It ends with a source-linked brief and a list of videos that still need direct review.

## Choose the research question

Write one question, a language and a publication window. Decide whether you need tutorials, product demonstrations, interviews or audience reactions. Mixing them without labels makes comparisons difficult.

After following the [Vaaya installation guide](https://vaaya.ai/install), use this prompt:

```text
Research [question] using YouTube videos published between [dates], in [language].
Use Vaaya consult to select current routes and quote the work. Budget: [amount].
Find a small candidate set, then choose [number] relevant videos to read deeply.
Keep video URLs, channels, publication dates and metric retrieval times.
Base content claims on retrieved captions or reviewed footage, with timestamps
where available. Mark metadata-only videos. Return findings and disagreements.
```

## 1. Search and select the videos

The current [Vaaya catalog](https://vaaya.ai/api/catalog) includes `apify/youtube-videos`. It requires `searchQueries` and accepts an optional `maxItems` up to 50. Include the topic and format in your query, then inspect the returned publication dates rather than assuming the request enforced your window.

TikHub also lists `/api/v1/youtube/web/search_video` through `tikhub/fetch`, with `search_query` required. Use `vaaya/discover` to confirm the endpoint and quote before calling it. A different YouTube route may use a different input name.

Select videos because they address the question and add evidence. Keep the channel and format attached to each row. A product launch announcement and an independent demonstration may discuss the same feature from different positions.

## 2. Retrieve the content you will summarize

Check available video metadata and caption options for the selected IDs. The TikHub catalog includes `/api/v1/youtube/web/get_video_info`, requiring `video_id`, and `/api/v1/youtube/web/get_video_subtitles`, requiring `subtitle_url`.

The subtitle route needs a subtitle resource URL, not the video's watch URL. Inspect the metadata for available caption information or ask consult for a current caption route and its required inputs. If no usable captions are returned, mark the gap.

For media you have permission to process, `deepgram/transcribe` accepts a hosted audio or video `url`. A web page containing a player is not necessarily a usable media URL. Confirm the input and quoted duration before choosing transcription as a fallback.

Preserve transcript language and available timestamps. Automatic captions can mishear names, numbers and technical terms. Check the footage around any precise claim you plan to quote or use as a decision input.

## 3. Read comments as a separate source

If audience reaction is part of the question, collect comments only for the chosen videos. `apify/youtube-comments` requires `startUrls` and allows an explicit `maxItems` cap. The [social-data guide](https://vaaya.ai/docs/reference/social-data) describes TikHub discovery for more targeted reads.

Record how many comments you received and their available sort or pagination context. Comments are a selected set of reactions; they cannot establish what every viewer thought. A comment describing a bug is a report to investigate, not independent confirmation that the bug affects everyone.

Keep these observations separate from the video's claims so the brief does not attribute a viewer's statement to the creator.

## 4. Build a claim-to-source table

Use an illustrative schema like this:

| Field | What belongs in it |
| --- | --- |
| Observation | A specific answer to the research question |
| Video and channel | Original watch URL and creator |
| Evidence location | Timestamp or transcript segment when available |
| Evidence type | Reviewed footage, captions, description or comment |
| Timing | Publication date and retrieval date |
| Qualification | Disagreement, uncertainty or unread material |

Write the brief from this table. Group sources that agree, then describe where they differ. Avoid calling a feature proven because several videos repeat the same announcement without demonstrating it.

## Close the remaining gaps

A failed caption request should leave the video marked unread or metadata-only. A successful search with a small result set should remain a small sample, not a survey of all YouTube coverage.

Keep actual charges and transaction IDs from [Vaaya's responses](https://vaaya.ai/docs/reference/making-calls) in the working notes. Hand the reader the final brief plus the few timestamps or videos that still need direct inspection. That lets them review the evidence without replaying every search you made.

## Questions

**Can a video title support a detailed summary?**

No. A title and description help select videos. Claims about the video’s argument should come from viewed content or a transcript you actually retrieved.

**How do I fetch subtitles through the listed TikHub route?**

The catalog lists get_video_subtitles with subtitle_url required. Obtain a valid subtitle URL from available video metadata or a supported caption route; a YouTube watch URL is not the same input.

**What if a transcript is unavailable?**

Mark that video as metadata-only or review it directly. Do not invent quotations or infer its full argument from the title and comments.