YouTube Comment Brand Monitor → Slack Alerts & Sentiment

Catch negative brand mentions on YouTube before they spread. Search by keyword (no URL needed) or scan comments under videos you paste. Each hit is scored with Twitter-RoBERTa plus likes / reply-thread size / thread negativity, filtered by a 7 / 30 / 90-day or custom window, then Slack-alerted only when it’s new and negative enough.

No YouTube login. Schedule it and sleep.

Why this Actor (not a keyword toy)

• RoBERTa sentiment — CardiffNLP Twitter-RoBERTa (social-trained), not a hand-maintained word list
• Composite risk score — blends text intensity + engagement + reply-thread size + reply-thread negativity
• Time windows — Last 7 / 30 / 90 days, custom YYYY-MM-DD range, or no limit (future dates rejected / clamped)
• Search pagination — keyword search keeps flipping pages until it hits your maxResults cap
• Two modes — keyword → videos, or paste URLs → comment scan + reply sampling
• Negative-first Slack — webhook fires only for new mentions past your threshold
• No YouTube login — public data via youtubei.js

Who it’s for

• Founders watching launch chatter under review / unboxing videos
• Small PR / marketing teams without Brandwatch budgets
• Automation builders piping mention events into Slack / Sheets / CRM

What you get (dataset)

Field	Meaning
source	youtube
keyword	which brand term matched
title, text, url	content + deep link
author, community	who / where
engagement	likes (comments) or views (search hits)
commentCount	reply-thread size
sentimentLabel	negative / neutral / positive
sentimentIntensity	RoBERTa P(negative) on the parent text
replyConcentration	mean negativity of sampled replies
sentimentScore	composite negative score (alerts + sort)
riskLevel	high / medium / low from the composite
isNew	first time seen for this monitor config
publishedAt, checkedAt	timing

How sentimentScore is built

0.50 × text intensity
+ 0.20 × intensity × engagementNorm   (log-scaled likes/views)
+ 0.15 × intensity × threadNorm       (log-scaled reply count)
+ 0.15 × replyConcentration           (sampled reply negativity)

First cloud run downloads the ONNX model into cache (slower once); later runs reuse it.

How to run

1. Enter brand keywords (search paginates until maxResults), or paste YouTube video URLs + keywords (comment scan).
2. Pick a Time range (default: last 30 days).
3. First run without Slack — inspect the dataset (sorted worst-first).
4. Tune minimumEngagement / negativeThreshold if noisy.
5. Add Slack webhook, then schedule with onlyNewMentions: true.

Example input

{
  "keywords": ["iPhone"],
  "youtubeVideoUrls": [],
  "lookbackPreset": "30",
  "maxResultsPerSource": 50,
  "negativeThreshold": 0.65,
  "minimumEngagement": 1,
  "onlyNewMentions": true,
  "slackWebhookUrl": "https://hooks.slack.com/services/…"
}

Support

Questions or volume pricing: email coohhh2408@gmail.com.