Tony Wang9 min read47,807 Remote Jobs Is a Floor — Not Remote Market Share
Crawlora's Jobs remote_open rose from ~27k to 47,807 after free-text inference. Ashby flags ~59% remote; Workday ~2%. That gap is data quality, not the market.
Every “remote work is dead / remote work is back” take needs a data source. Career-board crawls look objective until you open the remote field. On hosted ATS boards, that boolean is optional, inconsistently populated, and historically missing on the providers that carry the most volume.
Crawlora’s Jobs dataset now reports 47,807 open postings with remote=true (remote_open on GET /datasets/jobs/facets). A naïve reader will divide by ~525,000 open postings and say “about 9% of jobs are remote.” That sentence is false even when the arithmetic is correct. This post is the methodology note the series was missing — with the before/after of the 2026-07-19 parser fix, not a labor-economics forecast.
The headline number and why it still lies
| Metric | Value | Use |
|---|---|---|
remote_open | 47,807 | Count of open postings with remote=true |
| Provider-sum open postings | ~524,976 | Best denominator for this crawl |
| Naïve share | ~9.1% | Do not publish as market share |
Pre-inference remote_open | ~26,926 | Historical under-count floor |
The jump from ~27k to ~48k is real: enterprise boards encode remote in free text (US-Remote, Remote in the US, Canada Remote, office names like Remote - France) and used to leave the boolean empty. Filling that gap improves the floor. It does not turn the floor into a census.
Why not market share?
- Provider mix. Workday alone is ~49% of open postings in this crawl and still only ~2%
remote=true. Startup-leaning Ashby is ~3.5% of volume and ~59% remote-flagged. Change the vendor mix and the global % moves without the labor market moving. - Encoding is dual. Remote appears in booleans, location strings, office names, and workplace_type. We union several of those signals; we still do not read every PDF or every “fully flexible” synonym in body HTML.
- Coverage frame. Hosted ATS boards from Common Crawl / fingerprint discovery — not LinkedIn, Indeed-only, or internal ATS. Same series caveat as Workday share and Bengaluru geography.
- Hybrid is not remote. Inference deliberately does not treat bare “hybrid” as
remote=true. Hybrid stays on workplace_type when present.
Remote rate by ATS provider (the actual story)
If you only remember one chart from this post, remember this: remote-flag density is a property of the ATS integration and employer mix, not a single national remote rate.
| Provider | Open | remote=true | Rate | Text remote, flag false |
|---|---|---|---|---|
| Workday | 255,630 | 5,016 | 2.0% | 34 |
| Greenhouse | 115,575 | 15,283 | 13.2% | 3 |
| Workable | 38,882 | 9,742 | 25.1% | 0 |
| Oracle | 21,535 | 194 | 0.9% | 0 |
| SmartRecruiters | 19,440 | 2,748 | 14.1% | 15 |
| Ashby | 18,211 | 10,766 | 59.1% | 0 |
| Teamtailor | 13,503 | 168 | 1.2% | 0 |
| Recruitee | 10,897 | 1,481 | 13.6% | 1 |
| Personio | 9,931 | 584 | 5.9% | 0 |
| Rippling | 8,034 | 1,691 | 21.0% | 0 |
| iCIMS | 7,447 | 3 | ~0% | 0 |
| UKG | 3,542 | 54 | 1.5% | 0 |
| Eightfold | 2,236 | 46 | 2.1% | 0 |
| Lever | 113 | 31 | 27.4% | 0 |
| All (sum) | 524,976 | 47,807 | 9.1% | 53 |
Read the table left to right. Ashby’s 59% does not mean “Ashby companies are 30× more remote than Workday companies.” It means Ashby’s public API exposes a structured remote signal (and startup boards use it), while Workday’s public job boards mostly put workplace type in free text that we only partially recover. Greenhouse sits in the middle after location/office inference (~13%).
Volume-weighted truth: Workday + Greenhouse together are ~71% of open postings. Their combined remote flags dominate the global floor even when their rates look “low.” Startup-ATS percentages are interesting for product analytics; they are poison for “the market is X% remote” headlines.
What broke, what we fixed, what remains
Before
Workday list/detail and Greenhouse board normalizers often left remote unset. Structured flags worked on Ashby, Workable, and others. Result: remote_open was a floor on providers we already parsed well, not on the market. Series posts correctly refused remote share (roles post, Bengaluru).
After (2026-07-19)
- Shared
inferRemote/applyInferredRemoteacross ATS normalizers (location, title, workplace, multi-location, office names). - Topic false-positive guards for “Remote Monitoring / Sensing / Operations / …” when location is not remote.
- One-shot ES worker
jobs-dataset-backfill-remote-flagsrewrote open docs that already had remote-ish free text. - Quality gauges:
remote_trueandlocation_says_remote_flag_falseper provider, scraped every ~15 minutes. - Greenhouse board list always loads office names (upstream needs
content=trueforoffices) so Datadog-styleRemote - Franceoffices set the flag even when city sits inlocation.
Residual (honest leftovers)
- Topic titles: “Remote Monitoring Technician” can match a coarse text search without being workplace-remote. Guards and residual gauges catch most of this; expect a small nonzero residual forever.
- Description-only remote: body HTML that says “this role is fully remote” without location language may still miss if no structured flag.
- Hybrid ≠ remote by design.
- Board re-crawl lag: new postings get inference on write; historical closed docs were not the backfill target.
Random sample of remote=true postings after backfill showed high agreement with free-text remote language on title/location/workplace. Agreement on the positive class does not prove completeness of the negative class.
How to use the flag without lying
Do
- Filter
remote=truewhen you want roles that are labeled remote somewhere in structured or free-text fields we parse. - Compare provider-scoped remote rates (Ashby vs Workday) for integration quality and employer mix.
- Re-pull
remote_openon draft day — the crawl moves.
Do not
- Publish
remote_open / provider_sumas “percent of jobs that are remote.” - Compare our floor to BLS / household surveys without a methods paragraph (different universe).
- Rank cities by “most remote jobs” from this boolean alone (compound of two incomplete fields).
Same series, different unit
| Post | Unit |
|---|---|
| Workday market share | ATS vendor volume |
| Hiring power law | Employer open-req concentration |
| Roles / employment form | Department + employment_type labels |
| Bengaluru geography | City location strings |
| This post | Remote boolean quality + provider rates |
If you are building a remote-job product on ATS JSON, treat remote like employment_type: a dirty, high-value signal that needs inference, residual monitoring, and a refusal to over-claim.
What this is not
- Not a remote-work trend study. One crawl snapshot, one boolean after inference.
- Not “enterprise hates remote.” Workday’s 2% is a field-population rate, not a policy census of Workday customers.
- Not complete. LinkedIn-only and internal boards are out of frame.
- Not hybrid measurement. Use workplace_type and free-text hybrid language separately.
Everything above is re-queryable: facets for the global floor, quality gauges for per-provider rates, and the free job search tool for remote=true samples. When the next re-crawl moves remote_open, update the floor — do not pretend the percentage was ever the market.
Filter live postings with remote=true — knowing the bias
Search open roles across 14 ATS providers, facet remote_open, and inspect provider-level remote density. 2,000 free credits a month, no card.
Frequently asked questions
How many remote jobs does Crawlora's Jobs dataset report?
As of the 2026-07-19 facet pull, remote_open is 47,807 open postings with remote=true — about 9.1% of roughly 525,000 open postings by provider sum. That is a floor on the boolean after free-text inference and ES backfill, not a remote-work market share.
Why did remote_open jump from about 27,000 to 47,807?
Workday and Greenhouse public boards often leave remote unset and encode workplace in location or office names (US-Remote, Remote - France). Shared free-text inference plus a one-shot ES backfill of open postings recovered those cases. Pre-fix remote_open was about 26,926 on the same crawl era.
Is about 9% of jobs remote?
No. Dividing remote_open by the provider-sum denominator is arithmetic on a biased field. Workday is about half of open volume and only about 2.0% remote-flagged; Ashby is about 59% remote-flagged on a much smaller slice. Provider mix moves the global percentage without the labor market moving.
Which ATS providers flag remote most and least?
In the 2026-07-19 quality profile, Ashby leads among larger providers at about 59% remote=true (10,766 of 18,211). Workable about 25%, Rippling about 21%, Greenhouse about 13% after inference, Workday about 2.0% (5,016 of 255,630). Rates reflect API fields and employer mix, not a pure remote policy census.
What residual gaps remain after the remote inference fix?
Quality residual location_says_remote_flag_false is small (dozens of open rows, including topic titles like Remote Monitoring). Description-only remote, hybrid vs remote, and non-hosted boards (LinkedIn-only, internal ATS) are still out of scope. Hybrid is not treated as remote by design.
How should product builders use the remote filter?
Use remote=true as a biased filter for roles labeled remote in structured or free-text fields Crawlora parses. Compare provider-scoped rates for integration quality. Do not publish remote_open as percent of all jobs, and do not rank cities by remote from this boolean alone.