How viewers react to a launch video in its first 72 hours: comment themes, hour by hour
The comments in a launch video's first three days are the loudest public reaction you will get. Teams usually read the top few and form an impression. This post measures it instead: collect the comments every 12 hours, bucket them by hours since the video was published, and track what share of each bucket talks about price, problems, a rival product, or asks a question. Everything is counted in aggregate; no individual commenter is named or followed.
Why this is harder than it looks
- Volume is front-loaded. In a test on a large tech channel's video about new VR glasses, half of the 2,555 top-level comments in the first 66 hours arrived in the first 6 hours, and 84% in the first 24. Counting "price comments per bucket" mostly measures how busy the bucket was. Track shares, not counts.
- Buckets have different widths. The first hour and the third day are not comparable as totals. The script prints comments per hour next to each bucket.
- Late buckets are small. On day three a bucket may hold 150 comments, and a share of 6% is about nine of them. The noise is easy to mistake for a trend (see below).
- Keyword lists misfire. "Still feels like a solution searching for a problem" is not a bug report. Read a sample before you believe a theme.
- The comment records do not carry the video's publish time. You need it from somewhere else to compute "hours since publish".
Do you need a schedule at all?
Every comment comes with its own publishedAt, so one run at hour 72 can rebuild the whole timeline after the fact. If you only want the report, do that and skip the rest of this section.
Schedule runs if you want to see the reaction while it happens and answer it, or if you want comments captured before they are deleted or held for review. The trick is to pay for each comment roughly once.
Best Damn YouTube Comments Scraper is an Apify Actor that reads public comments through the official YouTube Data API v3. With sortBy: "time" it reads newest first, and with publishedAfter it stops paging once it reaches older comments, so older comments are never delivered or billed. publishedAfter takes a date, a relative period such as 1 day, or a full timestamp, which is what we want here:
{ "startUrls": ["https://www.youtube.com/watch?v=VIDEO_ID"], "sortBy": "time", "publishedAfter": "2026-11-07T09:30:00Z", "maxComments": 0, "includeReplies": false, "authorData": "omit" } maxComments: 0 removes the per-video cap; the date filter is the limit. authorData: "omit" removes every author field, because this analysis never needs to know who wrote what.
Leave includeReplies off. Replies are mostly people answering each other, they count toward the cap, threads with many replies need extra API calls, and the date filter works on top-level threads, so a new reply under an old comment would be skipped anyway.
Run it every 12 hours with a one-hour overlap
An Apify schedule sends the same input every time, so a fixed publishedAfter: "1 day" on a 12-hour schedule bills most comments twice. Instead, compute the cutoff at run time: 13 hours ago, which is one hour of overlap. A cron entry on any machine with GNU date is enough:
# crontab: 30 */12 * * * /path/to/collect.sh SINCE=$(date -u -d '13 hours ago' +%Y-%m-%dT%H:%M:%SZ) curl -s -X POST "https://api.apify.com/v2/acts/josh99smith~youtube-comments-scraper/run-sync-get-dataset-items?token=$APIFY_TOKEN" \ -H "Content-Type: application/json" \ -d "{\"startUrls\":[\"https://www.youtube.com/watch?v=VIDEO_ID\"],\"sortBy\":\"time\",\"publishedAfter\":\"$SINCE\",\"maxComments\":0,\"authorData\":\"omit\"}" \ -o "runs/$(date -u +%Y%m%dT%H%M).json" On macOS use date -u -v-13H instead. Each run writes one file to runs/, and the analysis merges them by commentId, so the overlap costs a little but never double counts. Stop the cron after six runs.
For the publish time, run Best Damn YouTube Scraper once on the video URL and copy publishedAt from the record. For a live stream, liveStartedAt is the better start.
The script
Save it as reaction.py and run python reaction.py runs/*.json. Python 3.8 or later, no packages. Edit PUBLISHED, RIVALS and the theme patterns for your launch; the values below are the ones from the test.
import json import re import sys from datetime import datetime PUBLISHED = "2026-11-06T21:29:30Z" # the video's publishedAt BUCKETS = [1, 3, 6, 12, 24, 48, 72] # upper edges, in hours since publish RIVALS = ["vision pro", "quest"] # the products people compare yours with THEMES = { "price": r"price|pric(ed|ing)|cost|expensive|cheap|afford|\$\s?\d|\d\s?(dollars|usd|bucks)", "problems": r"\bbugs?\b|crash|broken|glitch|laggy|\blag\b|overheat|doesn'?t work|not working|issue|problem", "rival": r"\b(" + "|".join(map(re.escape, RIVALS)) + r")\b|\bvs\.?\b|versus|compared to|better than", "question": r"\?\s*$|\?\s", } def ts(s): return datetime.fromisoformat(s.replace("Z", "+00:00")) start = ts(PUBLISHED) comments = {} # commentId -> record; later runs overwrite earlier ones for path in sys.argv[1:]: for c in json.load(open(path, encoding="utf-8")): if c["success"] and not c["isByVideoOwner"] and not c["isReply"]: comments[c["commentId"]] = c rows = [[] for _ in BUCKETS] for c in comments.values(): hours = (ts(c["publishedAt"]) - start).total_seconds() / 3600 i = next((i for i, edge in enumerate(BUCKETS) if hours < edge), None) if i is not None and hours >= 0: rows[i].append(c["text"].lower()) latest = max(ts(c["scrapedAt"]) for c in comments.values()) covered = (latest - start).total_seconds() / 3600 print(f"{len(comments)} unique top-level comments from {len(sys.argv) - 1} run file(s); " f"latest run {covered:.0f} h after publish") print(f"\n{'hours':>8} {'n':>6} {'per h':>6}" + "".join(f"{t:>10}" for t in THEMES)) lo = 0 for edge, texts in zip(BUCKETS, rows): if lo >= covered: break span = min(edge, covered) - lo shares = [sum(bool(re.search(p, t)) for t in texts) / len(texts) if texts else 0 for p in THEMES.values()] note = " partial" if edge > covered else (" (n < 30)" if len(texts) < 30 else "") print(f"{lo:>3}-{edge:<4} {len(texts):>6} {len(texts) / span:>6.0f}" + "".join(f"{s:>9.0%} " for s in shares) + note) lo = edge A comment can match several themes, so the shares do not add up to 100%. The owner's own comments are left out; they are not audience reaction.
How to read the result
From the test video, two run files merged (the second overlapped the first by a day, and the merge dropped 252 duplicates):
2556 unique top-level comments from 2 run file(s); latest run 66 h after publish hours n per h price problems rival question 0-1 515 515 6% 2% 11% 13% 1-3 420 210 8% 3% 13% 14% 3-6 337 112 7% 1% 11% 16% 6-12 438 73 8% 5% 11% 16% 12-24 438 36 5% 3% 10% 15% 24-48 248 10 8% 3% 13% 16% 48-72 160 9 9% 6% 11% 17% partial - The rate tells you when to watch. 515 comments an hour at first, 9 an hour on day three. If you plan to reply, the first six hours matter most.
- Flat shares are a finding. Rival comparisons sit at 10% to 13% throughout: the comparison was in the audience's head from the first hour, not something that grew.
- Check the margin before calling a trend. A rough 95% margin for a share
pfromncomments is2 x sqrt(p x (1 - p) / n). For problems at 6% of 160, that is about plus or minus 3.8 points, so the rise from 3% is within noise. A real shift looks like price going from 5% to 20%. - Read 20 random matches per theme. In the test, several "problems" matches were idioms or general complaints about the category, not about the product. Tighten the pattern, then rerun.
Cost and limits
- Price. $0.0004 per delivered comment or reply ($0.40 per 1,000). The test video's 2,555 top-level comments cost $1.02 in one run. Simulating 12-hour runs on the same data: with the 13-hours-ago cutoff, the runs deliver 2,675 comments, $1.07. With a fixed
1 dayevery 12 hours they deliver 5,163, $2.07. Videos with comments off, private or deleted videos and quota errors are free, and a run stops at the maximum cost you set. The one-off video lookup is $0.001. - Quota. One unit per page of up to 100 comments: the 2,555 comments used 27 units. The built-in key's daily quota is shared by all users; your own free key from the Google Cloud Console in
apiKeygives you 10,000 units a day. - Not every comment. YouTube's comment count includes replies and held comments; the API returns what is public when you ask. Our test video showed 3,633; the 1,079 replies under the 2,555 top-level comments make up the difference.
- Run time. Synchronous API calls suit short runs. On a very large video, the first run can be slow; use the
apify-clientpackages and wait for the run instead.
Comments are personal data under GDPR and similar laws. This method needs only text and timestamps: keep authorData on omit, report shares, and if you quote a comment internally, leave the name out.
If this saves you time, a review on the Store page helps other people find it, and a star on GitHub helps too.
Disclosure: I built both Actors. Source: github.com/josh99smith/youtube-comments-scraper and github.com/josh99smith/youtube-scraper. YouTube is a trademark of Google; the Actors are not affiliated with Google.