YouTube Comment Questions & Content Requests

Unofficial, non-affiliated tool for finding selected public comments on YouTube videos that heuristically look like questions, content requests, or issue reports. It uses the MIT-licensed youtubei.js package (18.0.0) and does not download video or audio.

What it does

Provide up to 20 public YouTube video IDs or exact youtube.com, www.youtube.com, m.youtube.com, or youtu.be video URLs. The Actor normalizes them to canonical watch URLs, requests public top-level comments sequentially, and delivers one dataset row for each unique comment ID that matches the selected deterministic rules.

Invalid input is an explicit invalid_input outcome, not a comment-collection result. Empty, malformed, oversized, credentialed, non-HTTPS, or lookalike-host video entries end the run before any provider request, actor-start charge, or question-result charge. The Actor writes an actionable INVALID_INPUT_NOTICE record plus SUMMARY and OUTPUT to the key-value store; these notice records are not dataset rows and are never billed comment results. Storage, billing, and other unexpected runtime failures are not converted into this outcome and remain failed runs.

Input limits are shared by all users: maxCommentsPerVideo defaults to 200 and is capped at 2,000; maxResults defaults to 100 and is capped at 2,000. Newest-first is the default to surface recent how-to questions and troubleshooting; top comments remain available. Pagination has a bounded page count and a short continuation delay, so output is a sample of available top-level comment pages, not a promise to return all comments.

Detection and output

rules-v1 uses visible punctuation, English interrogative/request/issue phrases, and Korean question/request/issue patterns. It includes the signalTypes, rule IDs, and matched evidence terms that selected a comment. English/Korean labels are character-pattern heuristics; other and ambiguous text is unknown. It deliberately does not judge topicality or whether an issue is still active, so jokes and rhetorical wording can remain in results when a listed rule matches.

Every matched comment is one row even when it has several signals. Exact repeated comment IDs are de-duplicated, while different comment IDs are preserved. Each row includes a safe canonical comment URL, text, provider likeCountText and replyCountText, nullable approximateLikes and replyCount only for unambiguous English-style public counts, publishedText, observation time, and transparent priority reasons. Unparsed localized counts do not add a count-based priority reason.

Illustrative output only; this is not an observation from YouTube:

{
  "videoId": "dQw4w9WgXcQ",
  "commentId": "illustrative-comment-id",
  "text": "Could you make a tutorial about this?",
  "signalTypes": ["question", "content_request"],
  "matchedRules": ["question-mark", "english-make-video"],
  "evidenceTerms": ["?", "make a tutorial"],
  "detectedLanguage": "en",
  "replyStatus": "not_checked",
  "detectionMethod": "rules-v1"
}

The Actor does not infer unanswered status, buying intent, sentiment, author identity, profile data, or a semantic meaning beyond the listed rules. Known creator-authored comments are excluded only when the provider's author_is_channel_owner flag is explicitly true; pinned, false, and unavailable owner flags remain eligible. The provider necessarily transmits comment metadata for this decision, but the Actor does not output or store author names, IDs, or profiles. It does not follow comment links, extract contacts, or use an LLM or external model API. Email and phone-like tokens in delivered text are redacted.

Availability and compliance boundaries

Public comments can be disabled, unavailable, challenged, rate limited, or blocked by the provider. Those outcomes are reported only in SUMMARY and OUTPUT key-value records as disabled, blocked_unknown, or unknown; diagnostic records are not added to the dataset. The Actor does not log raw input URLs, credentials, or comment-linked URLs, does not use login/cookies, proof tokens, proxy rotation, anti-blocking workarounds, alternate identities, or a JavaScript interpreter.

The client injects an HTTPS-only allowlisted fetch function for exact YouTube provider hosts, disables redirects, omits credentials, and uses a timeout. It creates one fast-fail session with retrieve_player: false; it does not fetch or execute a player, download media, or make a second identity after refusal.

The implementation is based on the documented Innertube.create, getComments, Comments.contents, and Comments.getContinuation interfaces in youtubei.js. A dated 2026-09-13 smoke test verified public top-level comments and continuation for one video. This validates one observed access path at that time; it does not guarantee availability for every video or future run, and it is not customer-revenue evidence.

Pricing

Pay-per-event pricing is $0.01 for actor-start after valid input and before investigation, and $0.01 for each charged question-result. No separate platform-fee events are configured. For 100 charged result rows plus the start event, a user can set a $1.01 maximum total charge; this is an example user cap, not a platform default.

The same functionality and safety limits apply to free and paid plans. There is no free-plan-specific feature, input, or result gate. Actor.pushData(row, 'question-result') stores the row before the SDK charges its events; an error propagates and there is no storage-and-billing transaction guarantee. The Actor counts a PPE delivery only when the question-result event count increases by one. All run diagnostics remain in key-value storage.

License notices

This Actor depends on youtubei.js, distributed under the MIT License. YouTube and related marks belong to their respective owners. This Actor is unofficial and is not affiliated with, endorsed by, or sponsored by YouTube or its affiliates.