We read the fine print on 116,000 "remote" job postings. Only 9.7% are truly work-from-anywhere.
This is a guest data study from RemoShift, an index of 115,000+ live remote jobs that re-checks every listing hourly. The pipeline behind it is a stack of parsers, rules, and a judicious pinch of LLM — and the fine print turned out to be the most interesting field in the whole dataset.
Only 9.7% of "remote" jobs are actually location-free
Of the 116,008 remote listings our collectors currently track across 20+ sources — company ATS boards (Greenhouse, Lever, Ashby, SmartRecruiters), remote-native boards (Himalayas, RemoteOK, Jobicy) and niche feeds — exactly 11,274 (9.7%) state no country restriction anywhere in the posting. The other 90.3% are "remote" with an address clause attached: must be authorised to work in the US, EU time zones only, based in the Philippines.
That distinction matters more as AI assistants become the front door for job search. When someone asks ChatGPT or Perplexity for "remote jobs I can do from anywhere," the engines retrieve pages that state the restriction structure explicitly. Vague listings lose to structured ones — which is also how we decided to render the data: every listing carries a machine-readable anywhere verdict derived from its own eligibility text.
How the classification pipeline works
The problem looks like text classification, and the naive solution is an LLM call per posting — at 100k+ documents that is cost and latency for marginal accuracy. Our ladder is deliberately boring:
- Rule tier. Regex-family matchers on the eligibility section: country-name inventories, right-to-work phrasings, timezone-band patterns ("overlap with GMT±3"). Fast, auditable, catches the majority.
- Structure tier. Some ATS APIs hand back structured location requirements; those bypass text analysis entirely. Garbage in this field is how you discover your parser silently returns empty defaults — our statistics page shows the honest coverage instead: only about half of listings state any compensation at all.
- LLM tier. Ambiguous residue only — "remote (US or Canada)" versus "remote — worldwide" is cheap to decide at the margin. Human audits feed corrections back into the rule tier, so the model's job shrinks over time.
The lesson generalises: for classification at index scale, LLMs are the tie-breaker, not the workhorse.
What the anywhere fleet looks like
- Functions: developer roles are the most likely to be truly location-free in absolute numbers; support and operations skew hard toward country-locked hires — payroll reality, not policy.
- Pay: among location-free roles, salary disclosure runs higher, and the bands skew senior: the anywhere fleet is over-indexed on engineering roles above $120k.
- Concentration risk: two sources account for roughly 84% of supply — one board's outage moved our total by 40% in a day. Aggregation is an uptime business, not a scraping one.
The full dataset, updated hourly: the ✦ work-from-anywhere index — 11,280 listings whose fine print names no country. Method and live statistics here.
Takeaway for anyone building with scraped text
One word — "remote" — carried two meanings until a classifier forced them apart. Whatever domain you point NLP at, the highest-value structure usually hides in a boring field (an eligibility clause, a fee schedule, a warranty term), not the headline. Parse the boring part; let people argue about the headline.
