Measurement · 30,729 tokens probed · 2026-09-03
Nearly half of the ATS board links on the public web are dead ends.
We harvested every Greenhouse, Ashby, Lever, Breezy, Recruitee and Teamtailor board link in 17 Common Crawl indexes — 30,729 distinct company tokens — then called every one of them against the vendor's own public API. 18,164 answered with at least one posting open. 12,565 did not — 40.9% of the list: 8,607 are gone outright, and 3,958 still answer with nothing open. Everyone building a job feed harvests these tokens. This is what nobody publishes about them.
What survived, by vendor
Solid is what answered with an open board. Hatched is what did not — a token that was linked somewhere on the public web, and today has no open role behind it.
-
Greenhouse 59.2% of 10,091 tokens still live
5,971 live882 answered, nothing open3,238 gone
-
Ashby 77.8% of 4,386 tokens still live
3,413 live301 answered, nothing open672 gone
-
Lever 38.8% of 4,961 tokens still live
1,925 live424 answered, nothing open2,612 gone
-
Breezy 45.5% of 4,562 tokens still live
2,077 live1,459 answered, nothing open1,026 gone
-
Recruitee 68.8% of 3,554 tokens still live
2,444 live386 answered, nothing open724 gone
-
Teamtailor 73.5% of 3,175 tokens still live
2,334 live506 answered, nothing open335 gone
| Vendor | Candidates | Live | Empty | Gone | Live rate | Postings |
|---|---|---|---|---|---|---|
| Greenhouse | 10,091 | 5,971 | 882 | 3,238 | 59.2% | 196,407 |
| Ashby | 4,386 | 3,413 | 301 | 672 | 77.8% | 59,609 |
| Lever | 4,961 | 1,925 | 424 | 2,612 | 38.8% | 69,565 |
| Breezy | 4,562 | 2,077 | 1,459 | 1,026 | 45.5% | 41,119 |
| Recruitee | 3,554 | 2,444 | 386 | 724 | 68.8% | 39,628 |
| Teamtailor | 3,175 | 2,334 | 506 | 335 | 73.5% | 33,539 |
| All six | 30,729 | 18,164 | 3,958 | 8,607 | 59.1% | 439,867 |
Live means the token resolved and the board had at least one posting open at the moment of reading. Empty means it resolved with zero postings — a real company with nothing open — and is counted as not live, which is the conservative choice. Gone is a 404, a 410, or a response that was not a job board at all.
A correction to our own number
Erratum — Lever, 2026-09-03
- We projected
- 1,721
- We measured
- 1,925
- Our stated 95% interval
- 1,591–1,852
Lever's API asks for one request per second, so a full pass takes about eighty minutes. Rather than wait, we probed a seeded random sample of 1,000 of the 4,961 tokens, got a 34.7% hit rate, scaled it, and published 1,721 live boards with a 95% confidence interval of 1,591–1,852.
The pass has since finished. The measured answer is 1,925 — -10.6% above the truth, and below the bottom of the interval we published with it. A 95% interval is allowed to miss one time in twenty; the honest reading is that our sample was not the uniform draw we treated it as, because liveness is not independent of where a token sits in the harvest.
We are leaving this here rather than quietly swapping the number, because a coverage figure you cannot audit is worth nothing, and the only way to show ours is auditable is to show it being corrected.
Sample the companies if you must. Never sample the postings.
Erratum — Breezy, 2026-09-03
- Boards projected from 250
- 1,770
- Boards measured in full
- 2,077
- Postings projected
- 29,653
- Postings measured
- 41,119
Breezy was added to this index on the strength of a 250-token sample, then probed in full before anything was published. The sample got the board count nearly right — 14.8% out — and got the posting count wrong by 27.9%.
The two errors are not the same size because they are not the same kind of quantity. Whether a board is live is roughly a coin flip you can sample. How many postings it carries is long-tailed: the median Breezy board has four open roles and the largest carry hundreds, so a small draw almost always misses the boards that hold most of the corpus, and it misses them downward every time.
So the rule we now follow, and the reason every number on this page says how it was obtained: a company count may be projected from a sample and labelled as such. A posting count may not. It is measured or it is not published.
A 403 is a fact about your client, not a verdict about the company
Greenhouse starts returning 403 once a client has asked too often. Our verifier
originally filed any non-429, non-5xx failure as dead — so throttling was being
recorded as "this company has no board."
The tell was arithmetic, not a stack trace. A run over 10,091 Greenhouse tokens
reported 4,391 live boards. An earlier run over 7,692 tokens — a strict subset of the same list — had
reported 4,932. Adding tokens cannot remove companies, so one of the two runs was not measuring what it claimed.
Splitting by status code found 2,107 403s in the new run against zero in the old, and 1,008 of those were
boards the earlier run had confirmed live hours before.
A refused request is now retried, and one still refused after its retries is recorded as
blocked — excluded from the denominator rather than counted as a miss. The general form of the
mistake: when a service rate-limits you, the failure describes your client, and folding it into a verdict about the
world biases the result downward while leaving it entirely believable.
Recruitee, 2026-09-03, is the same lesson costed. Its full pass at concurrency 8 left
7 tokens blocked or errored out of 3,554. Re-asked one at a
time, three seconds apart, 4 of them answered with live boards carrying
61 open postings; 2 still refuse at that pace and are counted as
neither live nor dead. Teamtailor's identical re-probe recovered nothing at all — which is what makes the
Recruitee correction a measurement rather than a number we went looking for. The cost of skipping this step is small
and one-directional: it only ever makes coverage look worse than it is, which is the flattering direction to be
wrong in and therefore the easy one to leave alone.
Guessing a company's token across vendors does not work
A company's board token is often the same slug whichever ATS it runs — stripe,
figma, ramp — so tokens harvested from one vendor look like free candidates for the
others. Tested on seeded random samples of up to 400 tokens per direction, it is worth 2–3% marginal coverage,
and 0.3–0.5% into Lever. Hypothesis rejected. Published so nobody has to rediscover it.
Measured against the largest free dataset, in both directions
kalil0321/ats-scrapers publishes a free
snapshot across 65 sources. On the six ATS providers we both cover we diffed their published manifest against this
roster, board token by board token. They had 2,035 boards we did not.
We probed every one of them on the same terms as everything else here — their snapshot supplied the candidate
list and nothing else, no status, count or posting figure of theirs was carried over — and
1,803 answered live with 40,469 open postings between them.
Nothing was blocked, so this is a measurement rather than a partial one.
Those boards are in the roster above. The script is scripts/diff-ats-scrapers.mjs and it reads their
manifest live, so this is reproducible in one command rather than on our word.
The diff pointed the other way too, and we published that half first: 5,929 of our boards are absent from their snapshot, carrying 81,785 open postings. The list is in net-new-vs-ats-scrapers.csv and was handed to them in an issue on their own repository, with the caveats stated against our own number. Their board counts are a lower bound — 11.3% of their rows on these six sit on a company careers domain and yield no token — which is exactly why the 2,035 above is a floor on what they have that we do not.
One thing the probe found that the diff could not: 152 of their 458 Lever boards answer 404 today. A published snapshot ages, and no amount of size substitutes for calling the endpoint. That is the argument for this roster and it applies to ours the same way, which is why every row here carries the URL that produced it.
What this measurement does not tell you
A board is findable here only if something on the public web linked to it while Common Crawl was looking. Companies that never linked their board anywhere crawlable are missing, and there is no way to count what you cannot see. This is a floor, not a census.
It also says nothing about freshness. The dead tokens measure the decay rate of an archive-aged list; they do not say how fast new companies appear, which is the number a re-sweep cadence should be tuned to. And the 3,958 boards that answered with nothing open are not classified — seasonally quiet and permanently dormant look identical from outside.
There is no Workday here, and no Taleo. Both are large, and both need a different method than a public unauthenticated board API.
Check it yourself
The 18,164 boards that answered are published in full, with the exact API URL that produced each row. Call any of them and you should get a similar count, drifting as they hire.
Public domain, no attribution required, no signup. The directory is here.