The Language of Work
How companies talk about work
Careers pages are corporate self-presentation at its most deliberate — every word chosen to attract workers. This is a computational study of two decades of them (more than a dozen tech companies, 2005–2026, reconstructed from the Wayback Machine), read as a longitudinal record: what the language does over time, when it shifts, and why. Movement is tracked as position along embedding-based semantic axes.
The thread running through the studies is a counterforces idea: a workplace culture stays good not because leadership cares about it, but while something specific holds ordinary decay back — mostly, workers having somewhere else to go. When that leverage erodes, the language of values lingers as liturgy while the practice underneath reverts. The 2010s bet that a new generation would permanently fix work wasn't wrong so much as rented: the gains tracked the labor market, and receded the moment it inverted.
Careers-Page Archaeology
How companies describe themselves as employers over time, measured along embedding-based semantic axes built from archived careers pages.
- Culture is downstream of power Worker-serving language (DEI) surged when workers had leverage and was cut once they lost it; the management-serving substrate (performance) never moves. →
- DEI Language Industry-wide adoption, retraction, and counter-programming on careers pages. →
- Changing the World When did idealistic "change the world" copy peak — and who still sounds that way? →
- A Team, Not a Family Netflix's 2009 culture deck, the model it spread (narrowly), and the scoreboard that isn't there. →
- The Care That Survived Care talk spiked with the labor market and receded — and the care that endured individualized. →
How it's built
A shared pipeline chunks archived pages, uses an LLM to classify them into registers, and scores them on embedding-based contrast axes — each paired with a neutral control and a circularity check, so the measure is stance rather than mere topical proximity. Structured extraction pulls benefits into taxonomies validated against hand-coded samples (Krippendorff's α ≈ 0.9).
Every claim is stress-tested with coverage-controlled statistics, and where the data can't carry a claim, the story says so rather than reaching — the honest nulls matter as much as the findings. Published as interactive data stories built with Astro, React, and visx; the pipeline is Python (embeddings, LLM APIs, pandas/scipy) over a content-hash DAG.
Companies
A per-company overview of everything the careers pages reveal.
Related writing
Companion essays on Substack, and the source behind the studies.