Data sources
Almost every source below is public and free to access. One important nuance: the Stanford "Canaries" figures are published aggregates — the underlying payroll microdata is ADP's and is not public. The table is generated automatically from the site's source registry (each entry carries its URL, access method, and license note), so it stays in sync with what's actually powering the dashboard, and it shows which metrics each source feeds, how much of the source the dashboard currently covers, and when it was last retrieved.
Occupation employment counts are national BLS OEWS May 2025 figures (data_vintage = 'BLS OEWS May 2025'; suppressed estimates are shown as null, never zero). Where an occupation has no OEWS row, no employment figure is shown — pre-2025 imputed placeholders have been removed.
Federal Workforce Breakout
The "Government & Public Administration" row on the For Employers panel is really three different workforce segments folded into one number, and each carries a different AI exposure profile. The breakout below is based on BLS QCEW ownership classes and NAICS industry codes:
- Federal Civilian (NAICS 92, Public Administration): Lower AI exposure due to security clearance requirements, procurement buffers, and constituent-facing services. Key occupations: Compliance Officers, Budget Analysts, Program Managers.
- Federal Contractor (NAICS 541, Professional, Scientific, and Technical Services): Higher AI exposure due to knowledge-work content. Key occupations: Management Analysts, Software Developers, Market Research Analysts.
- 501(c)(3) Policy & Advocacy (NAICS 813, Religious, Grantmaking, Civic, Professional, and Similar Organizations): Mid-range AI exposure. Key occupations: Social Science Research Assistants, Political Scientists, Grant Writers.
Source: O*NET industry-occupation matrix (for AI exposure scores) and BLS QCEW (for ownership-class employment counts).
Policy Interventions Evidence Rubric
The evidence tags on the For Policymakers page — Strong, Emerging, Moderate, Mixed — follow the rubric below, and they're reviewed quarterly by the policy team.
- Strong: ≥3 peer-reviewed studies or RCTs in OECD jurisdictions with positive effect sizes (e.g., 5-15% reemployment rate improvement).
- Emerging: 1-2 peer-reviewed studies or pilot programs in early-stage jurisdictions. Promising but not yet replicated.
- Moderate: Several peer-reviewed studies with small-to-moderate effect sizes, or a single large study with methodological caveats.
- Mixed: Multiple peer-reviewed studies with conflicting effect sizes. Outcomes vary by jurisdiction, design, and target population.
Last reviewed: July 1, 2026. Review cadence: quarterly.
Geography & ZIP resolution
Labor-market data is published at different geographic levels, each with its own granularity and freshness. The dashboard organizes it into three tiers:
Handling ZIP codes
ZIP codes are postal delivery routes, not statistical areas — worth stating plainly, since it trips people up. When you enter a ZIP code, the dashboard resolves it into the Census geographies that contain it — the CBSA (metro area), county, and PUMA. The statistics shown are native to those areas, not to the ZIP code itself. A ZIP code spanning multiple counties resolves to the larger one, and the split is noted.
Confidence levels
Every number in the underlying database carries its provenance with it — two fields that say where it came from and what was done to it:
- Source class — where the value ultimately comes from: official (a government statistical agency), private (a company's published data), research (a peer-reviewed study), or computed (derived by this site from the sources above). The class is derived from the value's source ID through a committed, reviewable mapping — not hand-tagged by whichever job wrote the row.
- Transform — what happened to the value after collection: raw (as published), derived, interpolated, imputed, or modeled.
- Vintage — when the value was published. A value re-published on a new day is stored as a new vintage, never an overwrite, so the history is preserved exactly as it was known at the time.
Rows that fail validation are quarantined with a recorded reason — never silently dropped and never re-labeled with a fabricated source.
The API also publishes a simple five-label confidence legend — measured, extracted, estimated, computed, and proxy — rendered below as a compatibility layer for existing readers:
Update cadence
Different data sources update at different frequencies. Here's what to expect:
- Daily: Indeed Hiring Lab postings-by-sector index and AI-mention share
- Weekly: Indeed Hiring Lab postings index (national and metro) and GenAI postings share
- Every 6 hours: state WARN notice scraping, with per-state freshness checks that alert when a state that normally files returns zero rows for 30 days
- Monthly: BLS employment and unemployment (payroll survey, LAUS, CPS), JOLTS, Challenger layoff data, Stanford Canaries slope, Yale Budget Lab occupational-mix shift, Sahm Rule, metro stress index (MS1), and Zillow home values and rents (mid-month)
- Quarterly: NY Fed college labor market report (Feb/May/Aug/Nov) and BLS QCEW county industry data (published with roughly a 5-month lag)
- Every two weeks: Census BTOS AI-use survey waves (the adoption series)
- Annual: Census ACS PUMA data (occupation mix, education, income, broadband, industry concentration)
- Event-driven: Anthropic Economic Index releases (polled monthly; ingested when a new release appears)
The "as of" date on each card shows when its underlying data was last retrieved. Some indicators may lag by 1–2 months because the source publishes with a delay.