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 a quarter century of them (two dozen tech companies, 1999–2026, 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 selection under leverage: a careers page is an audience-selection device, and its language records who the company needed to persuade — or was willing to lose — at each moment. When workers had somewhere else to go, the pages filled with belonging, care, and diversity commitments; as that leverage receded, the concessions deflated, the surviving care individualized, and an openly exclusionary register ("we're not for everyone") spread. Underneath both swings, the language that serves the employer — performance, merit, the unmeasured "high bar" — holds steady in every market. 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.
- DEI Language Industry-wide adoption, retraction, and counter-programming on careers pages. →
- A Team, Not a Family Netflix's 2009 culture deck, how narrowly it spread, and the scoreboard that isn't there. →
- Bring Your Own Resilience Care talk rose and fell with worker leverage; the care that survived is the kind you bring yourself. →
- Masculine Energy Gender-coded language across every careers page in the corpus. →
- The Non-Political Workplace A founder's blog, scored with the same instruments as the careers corpus. →
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 = 0.90 for the well-being locus codebook). Classifier agreement is measured the same way, against blind hand-labeled samples spanning every company: α = 0.81 for page-type classification as used in analysis (0.76 over the full seven-way taxonomy, n = 440), α = 0.80 for the finer-grained DEI registers (n = 201), and α = 0.98 for DEI stance (n = 100; partly in-sample — the stance prompt was refined against this sample's errors). The register number sits right at the conventional 0.80 reliability bar — a caveat the DEI story carries.
The embedding rankings are checked against instruments that fail differently. An LLM judge re-ranks the same years by reading the actual quotes — pairwise, order-randomized, with the concept paraphrased so the judge never sees the axis's own pole phrases — and an axis publishes only where the two rankings agree (Spearman ≥ 0.6; Netflix's performance axis passes at 0.62 by year and 0.77 by sentence, while the craft axis fails the same bar and stays unpublished). The gender-coding axis carries its own battery: known-answer occupation anchors, split-half reliability across its sixteen word pairs, agreement with the published Gaucher gendered-word dictionary (Spearman 0.66), and a masking test showing the company ranking survives removal of every sentence containing gendered terms or personal names (rank correlations 0.99 and 0.94).
Where the data can't carry a claim, the story says so rather than reaching — the honest nulls matter as much as the findings. Story prose is an AI-assisted synthesis of my drafts and the data, reviewed and edited by me before publishing. 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.
About
I'm Becca Bailey, a software engineer. This project started as curiosity about why job listings all sound the same, and grew into the corpus and pipeline behind these stories. I'm not a trained researcher, so the validation above is how I keep myself honest — and the code is public so you can check. More of my work is at becca.is.