Where people actually go: what happened when we added real career-change data to the graph

PivotHop’s graph has always been built from live postings: it reads what employers ask for and scores how much of a destination role you already cover. That number is useful and it has a blind spot you could drive a truck through. It knows what the market wants. It does not know what people do.

This month we wired in the missing half: three datasets of observed career transitions, actual humans counted moving from one occupation to another. An Oxford-built mobility network derived from the Current Population Survey (2010–2017, 464 occupations, published CC BY (an open license permitting commercial reuse) 4.0), a Department of Labor public-use file of 43,350 survey-weighted person transitions from Current Population Survey (CPS) and Survey of Income and Program Participation (SIPP), and, for finer European resolution, 355,315 anonymized career trajectories from the JobHop resume dataset. Career-tech is full of similarity models dressed up as mobility data. These three are the real thing: someone was an electrician in one interview and something else in the next.

Skills say no. People go anyway.

Across the 888 routes on our graph that now carry an observed signal, 224 are moves the skill math would have waved off, under 35 percent posted-skill overlap, that rank at or near the top of where people from that origin actually land.

The flow score is origin-relative: 100 means this is the single most common destination we can resolve for people leaving that occupation, not that everyone goes there. Read the electrician row as: of the places departing electricians turn up, construction manager leads. The posting-skill overlap between the two jobs is 13 percent, because postings for construction managers ask for scheduling, budgeting, and stakeholder wrangling, none of which an electrician’s posting mentions. The market learns those on the job. The similarity model can’t see it. The survey can.

Skills say yes. People stay home.

The reverse list is shorter and sharper: routes with 55 percent overlap or better where the observed flow is close to zero. Sixteen pairs qualify. Eleven of the twelve strongest point at a destination that requires a license.

This is worth sitting with if you are healthcare-adjacent. The skills genuinely transfer; the surveys show the moves genuinely not happening. Between those two facts sits two to six years of school and an exam. Career advice that only reads skill overlap will keep recommending these routes. The data on actual behavior prices them properly.

Why three sources, not one

Each dataset fails somewhere specific, which is the reason we run them together. The census occupation codes behind the Oxford network throw every designer, interior, graphic, industrial,UX (user-experience design), into one bucket, so “architects become designers” is as precise as that source can get. The European resume data is coded at European occupation classification (ESCO) leaf level, roughly 3,000 occupations, which separates a product designer from a signage designer, but it describes the Belgian labor market, so we display it as its own labeled signal and never blend it into US magnitudes. The Department of Labor (DOL) file isStandard Occupational Classification (SOC)-coded and fresher (roughly 2020) but covers only mid-level origin occupations. Where a pair falls into one source’s blind spot, the chain falls through to the next, and when none can resolve it, the route says so instead of inventing a number.

One caveat to carry into every table above: a flow of 100 is relative to the origin, and popular destinations are popular for everyone. Project manager and business analyst absorb leavers from half the economy, which says less about your specific adjacency than interior design’s pull on architects does. The curated relatedness prior we still keep as a fallback is damped for exactly this, and the flow scores inherit the same skepticism in ranking.

What changed on the instrument

Ranking on the graph now blends three signals, each shown separately in the route panel: skill readiness (weight 0.55, still the number printed on every node), shared work abilities fromOccupational Information Network (O*NET) (0.2), and observed mobility (0.25). For an architect, that pulled industrial designer onto the first ring at an unglamorous 15 percent readiness, with the panel stating why: people who leave architecture demonstrably go there. Structural engineer stays high too, but for the opposite reason, high overlap, modest flow. The two kinds of adjacency finally read differently.

One production note. The Department of Labor file arrived the hard way: the agency’s public-use listing for the study was removed from its site sometime between late 2022 and January 2026, and both dol.gov and bls.gov now refuse automated clients outright. We recovered the files byte-for-byte from Internet Archive snapshots and verified them against the study documentation. Public data has a shelf life. Mirror what you rely on.

Where this leaves you

When you evaluate a move, ask two questions instead of one. What share of the destination’s asks do I already cover, and do people from my occupation actually arrive there? High overlap with no flow usually means a licensing gate; budget years, not months. Strong flow with low overlap means the market routinely retrains people like you on the job, and the gap is more affordable than it looks. The instrument now shows both numbers on every route, with the source named. Run your own origin and read the disagreements first.