MonetizationKit guesses whether a YouTube channel is in the Partner Program from what a logged-out visitor can see: the subscriber count, a Join button, a Store tab, and ads on recent long videos. Checks from the website, from a September sample and from a few scheduled rechecks go into one checks table, and from that table we publish a per-country breakdown under CC BY 4.0.
We rebuild a live table every Monday and publish a new version for each monthly snapshot. There are two so far. The September 2026 version on Zenodo still has the files we uploaded that month, and October is version 2 of YouTube Monetization Signals by Country on Kaggle. For each snapshot, the same CSV goes on GitHub, Zenodo and Kaggle.
The live table
Each channel has one row in a checks table: its latest check and the verdict. At 01:30 UTC every Monday a Vercel cron calls a Postgres function that groups those rows by the country the creator set on the channel. The core of it:
select
c.snapshot ->> 'country' as country,
count(*) as sample_size,
count(*) filter (where c.verdict = 'likely_monetized') as likely_monetized,
count(*) filter (where c.verdict = 'possibly_monetized') as possibly_monetized,
count(*) filter (where c.verdict = 'likely_not_monetized') as likely_not_monetized,
count(*) filter (where c.verdict = 'unable_to_determine') as unable_to_determine
from checks c
where c.snapshot ->> 'country' is not null
group by 1;
The result is upserted into country_stats, countries with no rows left are deleted, and a one-row meta table records how many checks came from lookups, from the September sample and from scheduled checks.
Every country goes into country_stats, however small. The rule that a country needs 30 checked channels before it's published is applied when the table is read, so changing it is a one-line code change, not a migration.
That table backs the page for the country-level monetization dataset and its CSV download. The live CSV can change on Monday. If you want readers to reproduce a figure, quote it from a snapshot and cite that version's DOI.
Cutting a snapshot
A snapshot is three files: the CSV, a Frictionless datapackage.json with column types, and a README whose changelog says what moved since the last one. All three go into the dataset's GitHub repo through a normal PR and are uploaded unchanged to Zenodo. Kaggle gets the same CSV, and the README's content goes into the dataset description there.
| GitHub | Zenodo | Kaggle | |
|---|---|---|---|
| A new month goes in as | a PR to main | "New version" on the existing record | "New Version" on the existing dataset |
| An old month keeps | its commit | its own DOI and files | its version number |
| What to cite | — | the version DOI, or the concept DOI for "latest" | — |
We upload to Zenodo by hand. The GitHub repo has no releases or tags, so merging to main doesn't mint a DOI. Our concept DOI is 10.5281/zenodo.23401598. We didn't generate a separate DOI on Kaggle.
We added the October DOI to the GitHub README and CITATION.cff after the record was live, so the archived README doesn't contain it. That's why the two READMEs have different checksums while the CSVs match.
In each new Kaggle version we replace the old CSV with one under the same name, youtube-monetization-signals-by-country.csv, so notebooks that search /kaggle/input for it still find it. Whether they still pass is another matter: the October notebook checks October's row and channel counts.
These two commands print the checksums for the September CSV on Zenodo and in the repo:
curl -s https://zenodo.org/api/records/23401599 \
| jq -r '.files[] | "\(.checksum) \(.key)"'
md5sum analysis/2026-10-membership-rule/data/youtube-monetization-signals-by-country-2026-09.csv
Both give 3ee96060e3d17f3159f37ecfdba044c6 for the September CSV. The October CSV is f1aae92cf00d201a48261cd04b4fa8b5 in the repo root and in Zenodo record 23601867.
Changing the schema without breaking September's readers
We stopped computing median_subscribers on 1 October 2026, when we changed how we handle YouTube API data under its 30-day retention limit. The live CSV dropped the column. The October snapshot keeps it, empty, so code written for September's file still finds it.
The scoring rule changed too. In September it put eight countries at the bottom of the ranking, which I wrote about last month. We re-scored every stored check for October, so read the README's changelog before comparing the verdict counts it lists for both months.
The Kaggle description also has a note for future snapshots. Neither month includes Namibia yet, but if it shows up, pandas will read its country code NA as missing by default. keep_default_na=False, na_values=[''] keeps NA and still treats empty cells as missing.
Why "unable to determine" stays
In the October snapshot, 1,051 of 4,382 channels (24.0%) couldn't be called either way. We keep them in their own column rather than dropping them or folding them into "likely not monetized". In an early September crawl of about 900 channels, we hit rate limits and couldn't classify 37.1% of the US channels. After we slowed down and rechecked the same sample, that fell to 14.3%. Folding those checks into "likely not monetized" would have counted rate-limited requests as negative results. We now track failure rates over time as well as in total.
It also changes the ranking. Unknowns run from 6.1% of Saudi Arabia's channels to 41.1% of Bangladesh's, so two countries side by side can rest on very different amounts of evidence.
The October notebook on Kaggle reranks the countries with the unknowns left out. The rank correlation with the published order is 0.79. It starts with three asserts that would catch a bad export:
snapshot = pd.read_csv(CSV, keep_default_na=False) # "NA" is Namibia
checked = snapshot["channels_checked"]
assert (snapshot[VERDICTS].sum(axis=1) == checked).all()
assert checked.min() >= 30
assert (len(snapshot), int(checked.sum())) == (51, 4382)
In the October snapshot, 2,247 of the 4,382 channels in published rows (51.3%) fell into the "likely monetized" bucket. That figure describes the channels we checked, not YouTube as a whole.
What sponsors can check
For a sponsorship, the simplest check is to ask the creator, since ads alone don't confirm Partner Program membership. The YouTube creator application form on Tally asks creators for their subscriber count, average views over their last 10 long-form videos, and whether they're in the YouTube Partner Program. It tells applicants a sponsor may also look for a Join button, a Store tab or ads on long videos.
What's still manual
Each snapshot still takes a PR, two uploads and a DOI pasted back into the repo. We haven't redrawn the September sample, so new channels come from lookups: the US row grew from 66 channels in September to 457 in October. The CSV gives the total without a breakdown by source.
We built MonetizationKit. Our browser extension uses the same scoring rules to show a verdict on the channel page.
I used AI to help draft this post.
Top comments (0)