How this is built

The rule

Every relationship in our data carries five things: what it connects, where it came from, which version of that source, when it was observed, and how confident we are. A fact that cannot carry those does not go in.

Where the numbers come from

What a job involves — tasks, skills, tools, software — comes from O*NET, the U.S. Department of Labor's occupational database. We hold 43 consecutive releases going back to October 2015, which is what lets a page say what changed rather than only what is.

What it pays comes from the Bureau of Labor Statistics Occupational Employment and Wage Statistics, 2003 to 2024, as a full distribution by occupation, place and industry — never a single average.

Where it is heading comes from the BLS Employment Projections 2024–34.

The same job elsewhere comes from ESCO, the European classification, linked through the crosswalk published jointly by the ESCO Secretariat and the U.S. Department of Labor.

What we do to it

Sources rename things. Across these releases O*NET renamed its technology files three times, merged two occupations into one — “Accountants” and “Auditors” became “Accountants and Auditors” in 2020 — and changed the identifier scheme for work activities entirely. Handled naively, each of those quietly breaks a decade of history. We keep our own identity for every occupation, task, skill and tool, record every source identifier against it, and follow occupations through merges and retirements, so a comparison across ten years compares the same thing.

How the charts are coloured

The brand palette is bright by design. Bright is not the same as readable: the brand yellow sits at 1.41:1 contrast against a light page, so it is used as a surface and never as a data mark. Chart colours are darker steps of the same hues, checked with a validator for lightness, chroma, colour-blind separation and contrast — in both light and dark mode, each against its own background rather than flipped from the other. Where two categories appear, signs and labels carry the meaning as well as colour.

What we do not do

We do not tell you a job is “safe” or “at risk” from artificial intelligence. Doing that honestly means separating four different questions — whether a machine can do a task, whether it is permitted to, how much human judgement it needs, and who stays accountable. Those pages will appear when they can cite a standard or a regulation rather than a model's opinion.

Where we are thin

Wage geography before 2011 is national only, because that is all that was published. Occupations are re-surveyed on a rolling cycle, so an unchanged page can mean an unchanged job or an unvisited one — the survey date is shown so you can tell. Live employer demand is currently limited to U.S. federal vacancies, which is a real corpus but not the whole labour market.