I run a visa agency in Bali, and in the evenings I build scrapers on Apify. One of them reads TikTok comments: paste video links, get one row per comment with text, likes, reply count, date, language, the "liked by creator" flag and the commenter's public profile. No login, no cookies.
This post is not about how to get the comments. That part is easy — TikTok's website loads them from a plain JSON endpoint. It is about what "the first 100 comments" actually means, because I had seven runs of the same request sitting in my account and they told a story I did not expect.
The data
The Actor's example input points at one Nike video (7688359466324004110, a Caitlin Clark ad from 22 September). Between 6 and 17 November it was run seven times with the default Comments per video: 100. Every run returned exactly 100 comment rows plus one free summary row, in about 5 seconds:
Run summary: 100 row(s) delivered, 0 free; charged comment ×100 = $0.045. On 20 November I also read the same feed directly, twice, 20 seconds apart, from my own server — every page until TikTok said has_more: 0. That gave me the full list to compare the runs against.
TikTok's order is not "top comments"
The full feed today has 242 comments. Sorted by likes, the top of it looks like this:
likes comment in how many of the 7 runs 521 Nike GOAT 4 439 Is this narrated by Nick Offerman? 3 185 I love her more than anything … 0 141 Dropping the new commercial on … 0 54 CAAAAITLINNNNNN CLARK 0 44 I'm obsessed w this. 0 44 Perfection ❤️ 0 42 finally some good Nike-CC cont… 0 37 Hell yea she is. She's a gener… 0 29 I want the red Chiefs colored … 6 The third and fourth most-liked comments never appeared in any run. The two most-liked ones appeared in some runs and not others. Each run caught 57 or 58 of the 100 most-liked comments, and between 1,830 and 1,938 of the 4,338 likes on the whole thread — around 43%.
Nothing was broken. The Actor asked for "the first 100 in TikTok's order", and that is what it returned. The trap is that TikTok's order is a ranking of its own — mostly comments from the video's first days (222 of 242 were written in September), mixed with a few recent ones, and not sorted by likes or by date. The first page of my direct read went 09-23, 09-22, 09-23, 09-22, 09-23, 09-24, 09-23, 09-22, 09-22, 10-06….
If what you want is "what people say most loudly under this video", the first 100 is a sample, not the top.
The order depends on who is asking
Here is where "Nike GOAT" sat in each run:
run date position of "Nike GOAT" position of "Nick Offerman" 06 Nov 90 not in top 100 10 Nov 91 not in top 100 10 Nov 91 not in top 100 12 Nov 92 not in top 100 14 Nov not in top 100 1 17 Nov not in top 100 1 17 Nov not in top 100 1 In my direct read on 20 November, Nick Offerman was 2nd and Nike GOAT 4th. My first 100 and the Actor's last first-100 shared only 59 comments.
My own two reads, 20 seconds apart from the same address, contained exactly the same comments on every page of 50 — but not in the same order inside the page. So the page boundaries were stable for me; the ordering inside a page was not. The Actor reads from Apify's servers, and its first 100 was a different set. I cannot tell you what TikTok keys the ordering on — address, region, time, or something else — only that two readers do not see the same first 100, and the same reader does not see the same order twice.
Practical consequence: "comment #7 on this video" is not a stable fact. Do not build anything that diffs positions between runs. Diff by commentId.
275 comments, 242 rows
Every page of the feed carries a total. For this video it said 275. Pages of 50 came back with 47, 50, 46, 40, 48 and 11 comments — 242 in all — and then has_more: 0.
So even reading everything, you get 88% of the number TikTok prints. The gap is the same in both of my reads, so it is not random loss: total counts comments TikTok will not hand out publicly (hidden, filtered or under review), and it is not the reply count — the 242 comments carry 22 replies between them, and 242 + 22 is not 275 either.
One more thing about total: if you keep requesting after has_more: 0, the empty pages answer with total: 300 and total: 350 — it starts echoing the cursor. If your code reads total from the last page it fetched rather than the last page with comments, you report a number that does not exist. The Actor stops on has_more: 0, so its log line says "100 comment(s) of 275".
Likes drift within minutes
The two runs on 17 November were four minutes apart. They returned the same 100 comment ids — and 25 of them had a different like count:
change comments −1 21 −2 3 −11 1 (Nick Offerman: 438 → 427) All 25 went down, none went up. I saw the same drift across days: Nike GOAT read 520, then 510, then 520 again. My two direct reads 20 seconds apart had identical likes. I don't know whether the difference is caching on TikTok's side or unlikes being processed in batches. What I do know is that a like count is approximate to within a few percent at any moment, so ranking two comments with 17 and 18 likes is noise.
The language field is a guess
Each comment comes with a language code. In the 17 November run it said en 78, es 10, fr 4, id 2, hu 2, un 2, plus one each of ms and pt. One of the two un (undetermined) rows was this:
She is great, best person you could have to work with Nike. Great player but an ever better human. I'm 72 and can't wain to get me a pair of her signature shoes.
That's plain English. In the full feed, 43 of 242 comments were un. Use the field for a rough split, but if language matters to your analysis, run your own detection on text.
How to get the top comments
Because the feed is not sorted by likes, the only way to get the real top is to read the whole thread and sort it yourself. For this video, that means 242 comments — about 12 cents at $0.0005 each.
Cheaper: the Minimum likes filter. In the Actor's code the filter runs before the count limit, and skipped comments are not charged. So maxCommentsPerVideo: 100 with minLikes: 20 reads every page and keeps only comments with 20 or more likes. On this thread that is 33 comments covering 2,079 of the 4,338 likes — more signal than the default 100 and a third of the price.
{ "videoUrls": ["https://www.tiktok.com/@nike/video/7688359466324004110"], "maxCommentsPerVideo": 1000, "minLikes": 20, "includeReplies": false } From Python:
import requests rows = requests.post( "https://api.apify.com/v2/acts/lergassy~tiktok-comments-scraper/run-sync-get-dataset-items", params={"token": "apify_api_..."}, json={ "videoUrls": ["https://www.tiktok.com/@nike/video/7688359466324004110"], "maxCommentsPerVideo": 1000, "minLikes": 20, }, timeout=300, ).json() comments = [r for r in rows if r.get("type") == "comment"] for c in sorted(comments, key=lambda r: -r["likes"])[:10]: print(c["likes"], c["replyCount"], c["createdAt"][:10], c["authorUsername"], c["text"][:60]) Filter out type == "notice" (the free summary row) and type == "error" (a link that is not a video, or a video TikTok refused to answer for — also free, with the reason).
The Actor is TikTok Comments Scraper.
A few smaller things from the same rows
- Nobody pinned anything.
pinnedwas false on all 242 comments. On many brand videos the "pinned" comment you remember is actually the creator's own reply. - The creator liked four comments out of 242. In the 100-comment runs, exactly one of those four showed up ("Love her 🫶🏽 and her game!", 22 likes). If you are hunting for "comments the creator engaged with", you need the whole thread.
- Replies are rare here. Six of the 100 comments in the last run had any replies at all. Turning on replies would have added a handful of rows, not doubled the bill.
When this is the wrong tool
- You need a fixed, reproducible top-N. TikTok does not give one, and nothing scraping the public feed can. Read the whole thread and sort it yourself, and store the date you read it.
- Exact like counts over time. They drift by a few percent between reads minutes apart. Good for trends, not for "this comment gained 3 likes".
- The comment count in TikTok's UI.
totalincludes comments you will never receive. Report the rows you have, not the number TikTok prints. - Private or friends-only videos, and anything behind a login. The Actor reads what an anonymous visitor sees, and nothing else.
- Contacting commenters. You get public usernames and ids — nothing else, and that data is still personal under GDPR-style laws.
![UI Practice - Duolingo [1]](https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F8uawvr6a0e07ilktpkk7.png)

