I run a visa agency in Bali. The other half of my week goes into Apify Actors, and one of them reads public Snapchat profiles: you paste usernames, profile links, Spotlight links or hashtags, and you get flat rows — subscribers, bio, website, category, Spotlight views, shares, boosts, transcripts, stories.

The field everybody buys it for is subscribers, because Snapchat's web page shows a rounded blob like "751.6k" and offers no export. The Actor returns a number instead, and because it health-checks itself with the same prefilled input every night, I am not looking at a snapshot. I have a panel: 62 successful runs between 17 October and 19 November, the same three usernames every time — mrbeast, kyliejenner, khaby00 — 179 profile rows, 116 of them carrying a subscriber number.

Here is the first thing I noticed, and it took a month of rows to see it.

116 of 116 counts end in two zeros

Every subscriber value in the panel is divisible by 100. Not one read in 33 days produced a tens or a units digit.

It is not a 1,000-rounding either: the last three digits spread evenly across all ten buckets (500 appeared 18 times, 400 seventeen, 0 thirteen). Two values from last night's run:

{ "username": "mrbeast", "subscribers": 1481900 } { "username": "kyliejenner", "subscribers": 28566900 } 

The rounding is not mine. The Actor reads the JSON Snapchat embeds in its own page and passes the field through untouched — subscriberCount: pub.subscriberCount ?? null, no parsing of the "751.6k" string anywhere. Snapchat publishes a count on a grid of 100.

The reason I am confident this is a decision about this field and not a general policy: the same embedded payload counts everything else to the unit. From the same three accounts:

field value
views on a Kylie Spotlight video 85,547
views on the next one 169,330
spotlightViewsTotal for mrbeast 6,890,314
spotlightViewsTotal for khaby00 140,944
spotlightViewsMax for khaby00 96,974

Views to the single view, subscribers to the hundred. Practical consequence: 100 subscribers is your resolution, full stop. mrbeast's count moves +200 or +300 on a normal day, so a one-day diff on a mid-size account is two or three ticks of the dial. If your dashboard says "+37 followers today", it is not reading Snapchat. Anything under 100 is invisible, and a run-to-run difference of exactly 100 could be one person or ninety-nine.

The number refreshes once a day — and two unrelated accounts tick together

25 of mrbeast's 53 consecutive read pairs returned a byte-identical count; for Kylie it was 31 of 59. The longest flat stretch covers four reads over 13.5 hours (20 Oct 20:25 → 21 Oct 09:57 UTC), all four reporting 1,471,400 and 28,718,400. Gaps between two different values run from 13.1 to 24.1 hours.

Then the part that settled it. Take the 53 runs where both accounts were read in the same run and ask, for each consecutive pair, who moved:

pairs
both counts changed 28
only mrbeast changed 0
only kyliejenner changed 0
neither changed 24

28 out of 28 changes were simultaneous. A content creator in the US and a reality-TV account do not gain subscribers on the same clock — so what I am sampling is one daily batch job on Snapchat's side, published to both profile pages at once. Each run is a fresh logged-out HTTP read with no cookies and no reuse of a previous response, so the repeats are not my cache.

If you are monitoring Snapchat, poll once a day. An hourly schedule costs twelve times as much and returns the same number eleven times out of twelve.

Counts go down, and a declining account looks exactly like a broken scraper

Over the 33 days:

first read (17 Oct) last read (19 Nov) change
mrbeast 1,470,500 1,481,900 +11,400 (+0.78%)
kyliejenner 28,794,300 28,566,900 −227,400 (−0.79%)

mrbeast: 28 moves, all 28 of them up. Kylie: 28 moves, 27 of them down, and exactly one +100.

And the decline decayed. Average change per day, by week of the panel:

week avg/day
17–24 Oct −17,799
24–31 Oct −14,016
31 Oct–7 Nov −1,719
7–14 Nov −1,812
14–19 Nov −1,224

I am not going to tell you why hers fell; I read a public page, I do not get a changelog. What the shape says is that it behaves like a correction that bled out over two weeks rather than a steady audience loss — and that is the whole point for anyone building on this. A falling subscriber count is a normal observation, not a parsing failure. If your import job has a sanity rule like "reject rows where followers dropped", you will reject the true value and keep a stale one. Mine logs the drop and moves on.

The grid makes it worse in one specific way: on an account drifting by a few hundred a day, a single pair of reads cannot distinguish a real move from the hundred-boundary. Two reads give you noise. Two weeks give you a slope.

null is not 0, and hidden does not mean small

khaby00 was in all 62 runs. 55 profile rows, 33 days, and subscribers was null every single time. Not zero, not missing — Snapchat publishes no count for him to logged-out visitors, and the Actor writes null rather than inventing a floor.

Everything else about him came back fine: 7 Spotlight videos, 140,944 total views, a best video at 96,974, engagement rate 5.94%, bio "Official Khaby Lame Snapchat … 70 MLN on Tik Tok". Two more accounts from a one-off run behaved the same way — nike (verified, website, 17 Spotlights, 306,771 views) and cnn (verified, website) — both with no subscriber count.

So the pipeline rule that costs you the most is the innocent one:

-- turns a creator with 70M TikTok followers into an empty account COALESCE(subscribers, 0) AS followers 

Sort descending on that column and the three biggest names in the sample sit at the bottom of your shortlist. Keep the null, filter on subscribers IS NOT NULL when you rank, and count how many you dropped — in my panel that is 63 of 179 rows, and it is a property of which accounts you asked for, not of how well the scrape went.

The fields that stayed and changed meaning

In November, Spotlight rows read from a profile page stopped carrying aiTitle, seoKeywords, language and the sound — present and empty, not absent; those labels now arrive only with Spotlights found through a hashtag page. Meanwhile keywords stayed filled and became a different thing. Earlier rows carried three one-word topics: ["bedtime", "relaxation", "sleep"]. The rows in this panel carry 30 and 31 long search phrases: ["DIY cupcake popper hack", "how to make sprinkle dispenser", "creative gift unboxing ideas", …]. Same key, same type, same non-empty check — and a classifier that grouped videos by keyword would have silently found zero overlap with last month's clusters one morning. If you depend on a field's content shape, assert on the shape: length, token count, vocabulary overlap with yesterday. "Not null" is not a test.

Calling it

curl -s -X POST "https://api.apify.com/v2/acts/lergassy~snapchat-profile-scraper/run-sync-get-dataset-items?token=$APIFY_TOKEN" \ -H 'Content-Type: application/json' \ -d '{"profiles":["mrbeast","kyliejenner","khaby00"], "includeSpotlights":true,"maxCommentsPerSpotlight":10}' 

Filter on type — profile, spotlight, story and comment rows share one dataset, plus free error rows ({"kind":"not-found","error":"profile not found or not public. Not charged."} — that is what the username snapchat returns) and a free closing notice row with the run's own bill on it.

When this is the wrong tool. If you need follower movement to the person, or inside a day, no public reader can give it to you — ask the creator for their own analytics. If you need a creator's full back catalogue, the logged-out page shows only the latest Spotlight videos (about 25), so you build history forward with a daily schedule, not backward. And friends-only stories and DMs are not in here at any price.

Next on my list, straight out of this panel: a subscribersHidden boolean beside the null, so nobody has to learn the COALESCE lesson the way I did.