I run a visa agency in Bali. The other half of my week goes into Apify Actors, and one of them does a job that sounds trivial: give it an Instagram post link, get one row per comment — text, author handle and name, date, likes.
Comments are the field people ask me for most and the field I trust least. A profile's post list has a natural shape — here are the last 60 posts, in order, done. A comment list has no shape at all. Instagram decides what you see, re-decides it on every request, and never tells you how much it withheld.
So I left a schedule running against one post for two weeks and read the dataset instead of the docs.
The panel
One post: instagram.com/p/De_qHkUF186/, a natgeo photo published 27 October 2026 at 13:00 UTC. The data source puts its comment count at 1,088 as of today. Every read asked for 60 comments, the Actor's default.
Ten of those reads — 1 to 8 November, twice on two of the days — came back with 60 distinct comments each. 600 rows in total.
Those 600 rows held 75 different comments.
| Showed up in | Comments |
|---|---|
| all 10 reads | 44 |
| 7–9 reads | 14 |
| 2–5 reads | 14 |
| exactly 1 read | 3 |
A post with 1,088 comments has a core of 44 that always answer, a shifting fringe of about 30, and roughly a thousand comments I have never seen and have no cursor to reach. Forty pages deep is where my Actor's loop gives up; the ranked feed runs out of new material long before that.
The 75 are almost all from day one
Dates of those 75 comments:
| Written | Comments |
|---|---|
| 27 Oct (day of posting) | 41 |
| 28 Oct | 13 |
| 29 Oct | 7 |
| 30 Oct – 7 Nov | 14 |
Sixty-one of the 75 were written within three days of the post going up. The newest comment I ever saw, across every read up to 8 November, was dated 7 November — and it appeared in exactly one of the reads.
That kills the use case most people arrive with. "Watch this post and tell me when someone new comments" cannot be built on this. The feed is explicitly ranked — the source's own answer carries is_ranked: true and marks every comment is_ranked_comment: true — and ranking means recency is one weak input among many. Here is page one as it came back today, in the order given:
likes: 432, 109, 599, 401, 59, 1733, 34, 1, 561, 12, 156, 45, 199, 3530, 78 Not by likes. Not by date either: the 1-like comment in slot eight is the newest thing on the page, and the 3,530-like comment in slot fourteen is one of the oldest. There is no sort key you can reconstruct, and the API offers none — sort_options and filter_options both come back as empty arrays.
What the repeated reads are good for
One thing survived the instability, and it is more useful than it sounds. Those 44 always-present comments carry their own like counts, and over the eight days:
- 39 of 44 gained likes, 5 stayed flat, none went down
- median gain: +6.5 likes
- biggest: one comment went from 3,218 to 4,091 likes, +873
Comment-level like velocity is a measurement this tool can actually make. The set of comments is unstable; the counters on a comment you already have move smoothly and in one direction. If you are tracking which reply to a brand's post is accumulating agreement, key your storage on commentId and re-read on a schedule. If you are trying to inventory all comments, stop now.
Then page one came back as page two
On 9 November something changed upstream, and nothing changed in my code.
The comments endpoint serves 15 comments per request, so 60 rows means four pages plus one lookup to translate the post's shortcode into its numeric id — five requests to the paid source, every read, all month. Here is what came out of them:
| Reads | Rows | Distinct comments | Duplicate rows |
|---|---|---|---|
| 1–8 Nov (10 reads) | 600 | 600 | 0 |
| 9–14 Nov (8 reads) | 480 | 361 | 119 |
A quarter of the rows in the second window were the same comment delivered twice. And the duplicates were not random. In the 14 November read, I mapped every repeated row back to where it first appeared:
positions 0–14 : page one, all new positions 15+ : ...............DDDD....DD......D.....DDDD...DDD.....D...... repeat → origin: 15→0 16→7 17→2 18→1 23→4 24→5 31→10 37→11 38→8 39→9 40→3 44→12 45→13 46→14 52→6 Every origin is in 0–14. All fifteen comments from page one were served again, scrambled, inside the later pages. The cursor advanced, the ranker re-ranked, and nothing in the protocol excludes what you were already given.
It is still happening as I write this. Two raw calls today, page one then its next_page_id:
page1: 15 comments page2: 15 comments, 10 of which are already on page 1, in a different order page2 slot → page1 slot: [new, 3, 13, 0, 5, 8, new, 4, new, new, 12, new, 14, 2, 10] My Actor charges per row delivered. For six days it was charging people for 15 copies per 60-row read, and the dataset looked completely normal — 60 rows, no errors, 12 seconds, green run.
The fix, and what it costs
Remember the comment id inside the run, drop anything you have seen, and do not count a dropped row against the buyer's limit. The reads on 15 and 16 November came back 60 rows, 60 distinct comments — at seven source requests instead of five, because reaching 60 real comments now takes six pages instead of four.
That trade is the whole lesson. The duplicates were never free: I was paying for four pages and shipping three pages of content. Paying for six pages and shipping four is cheaper per useful row, and the buyer's row count finally means what it says.
While I was in there, two fields got honest treatment:
-
isAuthorReplywas empty in every comment row the Actor had ever produced — 1,323 rows across 29 runs,nullin all of them. The mapper needed the post owner's handle and nobody passed it, even though the owner is already sitting in the comments response, inside the post caption that rides along. Fixed today at no extra request, plus a newpostOwnerUsernamecolumn so the row says whose post it is without parsing the URL. -
replyCountis documented as always empty now instead of printing a number. This endpoint returns top-level comments only and carries no reply count at all — absent in all 1,323 rows. Replies to a comment are not collected. A field that is alwaysnullis annoying; a field that is always0is a lie.
One more trap worth naming: the response ships has_more_comments: false while also handing you a next_page_id that works and a has_more_headload_comments: true right next to it. Trust the cursor, not the boolean.
Calling it
curl -X POST "https://api.apify.com/v2/acts/lergassy~instagram-comments-scraper/run-sync-get-dataset-items?token=$APIFY_TOKEN" \ -H 'Content-Type: application/json' \ -d '{"postUrls":["https://www.instagram.com/p/De_qHkUF186/"],"maxItemsPerTarget":20}' A real row from today’s verification run, with the author-avatar URL dropped for width:
{ "type": "comment", "postUrl": "https://www.instagram.com/p/De_qHkUF186/", "postOwnerUsername": "natgeo", "mediaId": "4017114630619356986", "commentId": "18116666669107216", "text": "Hello my hypermobile friends", "authorUsername": "kit.n.kboodl", "authorFullName": "Kit", "authorIsVerified": false, "createdAt": "2026-10-27T15:22:29.000Z", "likeCount": 4284, "replyCount": null, "isAuthorReply": false, "scrapedAt": "2026-11-16T13:13:00.125Z" } Twenty comments took 7 seconds; 60 take 13 to 18. Post links, reel links and bare shortcodes all normalise to the same thing. A post with comments turned off, deleted or private returns one free notice row naming the reason rather than an empty dataset — filter with type != "notice". The Actor is Instagram Comments Scraper; logged out, Instagram answers a post request with a redirect to the login screen and its internal API with 401 require_login, which is why this reads through a paid source that holds the sessions instead of asking for your cookies.
Each post link is charged once as a target — the lookup costs the same whether it returns 60 comments or none — and rows are charged on top. On the Apify Free plan a run collects the first target in full and charges nothing, which is the honest way to see what 60 comments of your own post actually look like before paying for 40 of them.
When this is the wrong tool
- You need every comment. You will get the ranked head of the list. On a 1,088-comment post that was 75 comments across eight days of trying.
- You need new comments as they arrive. Day-one comments dominate the ranking forever. Eight days of reads surfaced exactly one comment written in the previous week.
- You need reply threads. Top-level comments only, no reply counts, no replies.
- You need the same rows twice. Thirty-one of the 75 comments were missing from at least one read. Re-reading is not re-fetching; store what you got.
- You want to post, like or reply. This only reads.
The habit I am taking away: count distinct keys in your own output before you trust the row count, and do it on a schedule, not once. My row count was right every single day. For six days it was describing 45 comments and calling them 60, and the only way to see that was to ask the dataset how many different comment ids were actually in it.