What the score is trying to answer
The score answers one question: how healthy is the hiring market for this occupation right now, given what official sources have actually published. It is a market signal, not a prediction about any individual job search, and it is not a quality or desirability rating.
Three components combine into the composite. Each is scaled to a 0 to 100 range so that occupations of very different sizes remain comparable, then weighted as shown below.
Current Demand
BLS JOLTS and BLS CES
The near term hiring signal, and the largest single weight. It reads job openings, hires, and layoffs and discharges from JOLTS, together with employment and payroll momentum from CES.
Layoffs and discharges are inverted, so rising involuntary separations pull a score down rather than counting as activity. Because JOLTS and CES publish by industry rather than by occupation, the signal is weighted toward the industry mix that employs each occupation.
Occupation Fundamentals
Occupational employment trend and labor availability
The structural condition of the occupation itself: how its employment level has been trending, and how tight or available labor is for that kind of work.
This component is not a pay measure. Wage levels are shown on occupation pages as context and are not an input to any part of the score.
Long-Term Outlook
BLS Employment Projections
Projected employment growth over the official projection horizon, together with projected average annual openings, which combine growth with replacement of workers who leave the occupation.
It carries the smallest weight on purpose. Projections describe the published projection horizon, while a job search happens in months, so near term demand matters more to the composite.
What is context, not score
Occupation pages show more than the score. Median annual wages and metro employment come from BLS OEWS, and occupation descriptions, alternate titles and skills come from O*NET. Both are supporting context. Neither feeds the score.
Keeping pay out of the score is deliberate. A well paid occupation that is barely hiring is not a healthy market, and a moderately paid occupation adding people quickly is. Mixing the two would make the number harder to interpret rather than easier.
State pages also show a context-only statewide WARN panel where an official state source can be read reliably. The statewide notice count counts notices whose published notice dates fall within the displayed 90-calendar-day UTC period. The card shows the last successful source check and links to the agency information page, not a spreadsheet download. Company names, individual notices, individual worker counts and locations are not included in the public WARN response. Sources that also record smaller Rapid Response or courtesy filings below federal WARN thresholds are labeled as broader state tracking systems rather than federal-WARN-only lists. Oregon additionally shows statewide reported-position totals and a comparison of the latest two completed calendar months when those totals are complete. WARN is statewide context only and never changes the Job Market Health score.
The score is national. There is no state-level Job Market Health score, and the national score is never restated under a state name. State pages and the state comparison report published BLS QCEW industry conditions for the industries that employ an occupation: employment change, business locations and average weekly industry wages, each with its national benchmark in percentage points. Those figures are industry measures, not occupation counts, occupation pay or job openings, and a comparison sentence is only produced when every compared state passes the same identity, source-verification, coverage, quarter and industry-mix vintage checks.
Normalization and comparability
Raw source values are not comparable across occupations on their own. An occupation employing two million people and one employing forty thousand produce openings and hires counts of entirely different magnitudes. Each component is therefore scaled against a comparable baseline before weighting, so that the composite reflects relative condition rather than absolute size.
Beyond that, this page does not describe internal transformation details that the published implementation does not itself expose. Where the score reports a component value, that value is shown directly on the occupation page rather than being summarized here.
Score history, revisions and restatements
Federal data gets revised. When a source revises an earlier month, the affected scores are recalculated. That is why each month can carry two values: the score as first published, and the restated score reflecting later revisions. Occupation pages show the published score alongside a restated value when one exists.
Each score also carries a confidence rating and the reasons behind it, so a month built on a partial or stale source is not presented as though it were fully supported.
Versioning and provenance
This is methodology version 1.0. When weights or inputs change, the change is published as a new version rather than quietly rewriting history, and scores record the methodology version they were computed under. Every occupation page names the release behind each figure and the date it was imported.
More on release cadence, coverage and limitations is on the about the data page, and each dataset has its own page under data sources.
AI data: exposure, usage, adoption and outcomes
JobMarketHealth reads seven AI source families and keeps what each one measures separate. None of them is treated as a direct measure of AI-caused job loss, and none enters Job Market Health scores.
- Exposure (technical applicability): OpenAI's GPTs are GPTs occupation ratings, Microsoft's Working with AI applicability scores, the Felten, Raj & Seamans AI Occupational Exposure (AIOE) index and, when a manual official file is available, the OECD AI Exposure Measure. These describe how much of an occupation's work current AI could technically assist with.
- Observed usage: the Anthropic Economic Index, built from privacy-preserving aggregates of real Claude conversations mapped to O*NET tasks.
- Automation vs augmentation: from the same index, the share of observed task conversations that followed automation patterns (directive, feedback loop) or augmentation patterns (task iteration, validation, learning), weighted by task prevalence. Shown only when at least three tasks match.
- Employer adoption (Census BTOS, two separate series): the BTOS Core AI Use question, asked every two weeks, on whether a business used AI in the last two weeks, nationally and by state; a change from a year earlier is called clear only when it exceeds 1.96 combined standard errors. Separately, the one-time 2025-26 BTOS AI Supplement (collected November 17, 2025 to February 8, 2026) covers business functions using AI, how AI related to employee tasks, operational and workforce changes, employee use of generative AI, plans and barriers, nationally and by state, sector (NAICS) and firm size. The supplement is a structural snapshot, never blended with the core trend. Both are business-level estimates; Census did not measure occupations, so occupation pages show them only as all-industry context. Suppressed estimates are omitted. Sector results are not bridged to occupations, because no documented occupation-by-industry weighting is applied.
- Labor-market outcomes and projections: the employment, wage, hiring and postings data already on each page, plus BLS Employment Projections as official long-term structural context. BLS prepares annual 10-year projections covering about 800 detailed occupations and about 300 detailed industries and explicitly researches emerging technologies including AI. The projections are conditional on stated assumptions, carry substantial uncertainty, and are not interpreted here as estimates of jobs lost to AI. BLS AI impacts methodology.
How the exposure band is formed
Each exposure source is ranked within its own distribution of U.S. occupations (percentile 0 to 100; the OECD capability gap is inverted because a lower gap means closer to AI). The summary band is the simple mean of the available current percentiles, shown only when at least two sources cover the occupation, and split into lower, middle and higher thirds. Every page lists the component measures and how many were available. State workforce exposure weights these national ranks by OEWS state employment and is shown only when at least 75% of published state employment is covered.
Mapping limits
The OECD AI Exposure Measure publishes an AI Capability Gap for 879 O*NET occupations and nine domains (language, social interaction, problem solving, creativity, metacognition, knowledge/learning/memory, vision, manipulation, robotic intelligence). A lower gap means current AI is closer to what the occupation requires, so higher potential exposure; the raw gap is shown as published and only its rank is inverted. The OECD's hypothetical "+1 cognitive level" scenario is not used. OECD data: CC BY 4.0, OECD (2026), OECD Artificial Intelligence Papers No. 59, doi:10.1787/f3da0f0a-en.
AIOE (Felten, Raj & Seamans) also publishes industry exposure (AIIE, 4-digit NAICS), geographic exposure (AIGE, by state and county FIPS) and language-modeling and image-generation variants. Each is stored as its own measure and never merged with another. State pages show the published state AIGE, which reflects that state's occupation mix in the authors' vintage, not current AI use. The AIOE repository states no explicit data license; the authors request citation, so reuse terms are treated as unclear.
Occupations are keyed to 6-digit SOC codes. O*NET detailed occupations use the base ".00" row when one exists, otherwise the mean of the detailed rows, and that mapping is recorded. AIOE uses SOC codes as published and may reflect an older SOC vintage; unmatched codes are dropped, never fuzzy-matched. Task statements shared by several occupations count toward each of them.
Freshness
Research datasets are versioned snapshots from their official repositories, with file hashes recorded; they are treated as stale after about 13 months without a successful check. BTOS Core is stale 45 days after its latest reference period; the one-time AI Supplement ages from its retrieval. Stale or missing sources are dropped from every page rather than shown as current. The OECD file link is found on the official OECD page and the file is downloaded from oecd.org.
Limits worth stating plainly
- Coverage is national. There is no local or state level score today.
- Published data is lagged. The most current month available is weeks behind the calendar, and occupational sources are annual.
- Monthly demand data is collected by industry, so occupation level demand is weighted exposure rather than a measured occupational vacancy count.
- Nothing is estimated to fill a gap. When a source has not published a value, the page says so.