Count each posting — every job posting counts once. Amazon has 601 postings
with pay and a small lab has 15, so Amazon moves the numbers 40× more. This answers
“what share of openings ask for this?”
Count each employer — work out each company’s own rate, then average those
rates. Amazon and a 15-posting lab count equally. This answers
“what share of companies want this?”
It changes the picture where you would expect: C++ falls 35% → 23% and vague
“Machine Learning” 21% → 13% (Amazon boilerplate), while PyTorch rises
28% → 35% and Kubernetes 8% → 13% — skills most employers want but
the biggest board posts proportionally less.
Employers with fewer than 5 postings are left
out rather than counted noisily: one posting would give a 0% or 100% rate. 61 of 1,385
employers currently qualify, so this is “the market among employers we have read enough
of”, not the whole long tail.
Postings analysed
—
Employers
—
Distinct skills
—
Pay
Employer-stated bands only.
Distribution of stated pay
A median hides shape — two groups of roles at different levels read as one middling number.
Pay difference vs the slice median
Dot is the median difference; the bar is a 90% interval from resampling employers, so one company posting twenty similar roles counts once.
The interval covers sampling noise, not confounding.
Set Seniority above to hold job level constant — several buckets are
really measuring seniority: "Leadership & management" reads +13.7% across
all levels and 0.0% among senior roles alone.
Skills
What postings ask for, and how they frame it.
What employers ask for
Share of postings the extractor has read, split by how the posting framed it.
Who the role is for
Read from the posting text by the LLM pass.
Pay by skill
Median of employer-stated bands.
Sources
Abridged sources carry shorter text, so their skill rates are not comparable with the rest.
Postings
Every number above traces back to sentences in here.
Browse
Pick one to see what was extracted from it, and the sentence each skill came from.