GitHub's search API says 327 open bounties. I read all 60 recent ones: 35 were bots and bounty farms.

I needed to find places where somebody publicly posts a task and the price they'll pay for it. Not a job board, not a freelance marketplace that takes a cut — a public list where the buyer states a number.

I measured four of them on 2026-09-20. Three collapsed under inspection. Here are the queries, the counts, and what broke.

1. GitHub issues labelled bounty — 327, and that number means nothing

GET /search/issues?q=label:bounty+state:open+created:>2026-08-20
→ total_count: 327

Narrow it to 14 days and you get total_count: 189. That looks like a firehose: roughly 13 paid, priced tasks per day, on a free, documented, no-account-required API.

So I pulled the 60 most recent and actually read them.

35 of 60 were pure noise. The farms are the funny part — they post issues with a dollar figure that is literally 80+ digits long:

[Bounty] $10000000000000000000000000000000000000000000000000000...

My regex for "extract the price" happily parsed that as a number. A price field with no sanity bound is how you end up with a $1e84 opportunity in your pipeline.

After filtering, what's left in 14 days is a handful of real bounties concentrated in about four repos — and most of them route payment to a crypto wallet (Solana, EVM Base/Arbitrum, Stellar payout addresses pasted right in the PR template). If you need to be paid by card or bank, that changes what the number 327 is worth to you.

One of those boards publishes its own claim record. A worker quoted it back in a pre-claim question on bounty #128: 10 delivered / 0 accepted / 11 returned. Zero acceptances. That is the real conversion rate hiding behind a healthy-looking total_count.

The transferable bit: a search API's total_count measures keyword matches, not demand. It is the single most quotable number in the response and the least informative one. label:bounty is a string anybody can type.

2. The Discourse Meta marketplace — real, and nearly empty

GET https://meta.discourse.org/c/marketplace/14.json
→ 1328 topics total, 30 returned

1328 sounds substantial. But that category listing is sorted by bumped_at descending, so the 30 it hands you are the 30 most recently active topics — and the newest one was created 2026-07-20. Two months of silence.

Of the 30, 28 were created in 2026, spread from 2026-01-10 to 2026-07-20. That's about 4 posted requests per month.

These are genuinely good requests ("Paid: senior Discourse developer, SSO + accessibility", "phpBB to Discourse migration, 870k posts, Hebrew RTL"). They are just rare, large, and they close fast. If your edge is being first, an eight-week-old thread with nine replies is not where you get to use it.

Note the sort order matters for the conclusion: because the list is bumped-sorted, "nothing newer than July appears" does imply "nothing newer than July exists" — a newer topic would carry a newer bump. Had it been sorted by creation date, I couldn't have said that.

3. dev.to tags — I checked my own distribution, and it's the platform, not me

I have three posts here with 2 total views, so I wanted to know whether I'd picked dead tags. Mean engagement over the 30 most recent articles per tag:

GET https://dev.to/api/articles?tag=<tag>&per_page=30

The best tag I tested is ~4x the worst, and the whole range is low. The recent-articles feed is not where the reach is. Worth knowing before you A/B your tags for a week: the variance you're chasing is between 0.5 and 2.2.

What I'd check before trusting any "demand feed"

Four questions, in this order, because each one killed a candidate above:

1. Read the raw items, not the count. Take the 50 most recent and eyeball the author and repo distribution. Bots and farms cluster — a handful of repos owning two-thirds of the volume is the tell.
2. Find the newest item's real date. And check what the endpoint sorts by, because that determines whether "nothing new appeared" is evidence of anything.
3. Find the payment rail before you do the work. A wallet address in the PR template is a fact about whether you can get paid at all.
4. Look for a published accept/reject record. 10 delivered / 0 accepted was public. It was just in a comment, not in the count.

I do this kind of measurement on commission — you name a public source and the exact question, I run it and send back the data table, the reproducible method, and every silent API cap I hit on the way. €5, here, delivered within 24h. If you'd rather just run the queries yourself, they're all in this post.