Skill adjacency is a measure of how similar two roles' work content is — their tasks, their knowledge requirements, even their job titles. It ranks which roles sit close to which. It does not measure whether a particular person can do the second job. That remains an assessment question, and the distinction decides whether a redeployment plan works or quietly fails.
Adjacency scores are now embedded in talent marketplaces, internal mobility tools and skills-graph products, usually surfaced as "related roles" or a percentage match. They are genuinely useful. They are also routinely read as something they are not: evidence about a person.
What an adjacency score is actually built from
The clearest published example is the ONET Related Occupations matrix, because its method is documented in full. Each of the 923 data-level ONET-SOC occupations is mapped to 10 primary and 10 supplemental related occupations, and the research dataset scores every one of the 923 against the other 922.
The score is an equally weighted composite of three standardised variables, documented in the 2022 method report and refreshed in the November 2025 update:
- Work-based similarity — the average cosine similarity between the occupations' task statements and their Detailed Work Activities, computed from sentence embeddings of the text.
- Knowledge similarity — cosine similarity between the occupations' importance-rating profiles across O*NET's 33 knowledge domains.
- Job title similarity — similarity between the occupations' alternate titles.
O*NET's own rationale for that choice is "what people in the occupations do, what they know, and what they are called."
What is not in the composite matters just as much:
- Abilities
- Skills
- Work styles, interests and work values
- Salary
In the 2022 development work those domains were used as external criteria to check the composite, not as inputs to it. And the method is not fully mechanical: in the 2025 refresh, human reviewers replaced 190 primary relations across 149 of the 635 occupations they reviewed. O*NET describes the purpose of the product as career exploration, career guidance and job search — not personnel decisions.
So a related-role score is a statement about the similarity of two job descriptions. Two employees in the same current role have an identical adjacency score to the same target role, and very different readiness for it.
Why this matters more than it used to
The redeployment question is no longer occasional. The World Economic Forum's Future of Jobs Report 2025, based on over 1,000 employers representing more than 14 million workers across 55 economies, puts it in per-100-workers terms: 59 of 100 workers will need training by 2030, of whom 29 could be upskilled in their current roles and 19 "could be upskilled and redeployed elsewhere within their organization." Eleven are unlikely to receive the training at all.
The same report finds that 39% of employers' existing skill sets will be transformed or become outdated over 2025–2030, that 50% of employers plan to transition staff from declining to growing roles, and that 63% name skill gaps the single biggest barrier to business transformation.
Nineteen in a hundred is a lot of internal moves to get right, and adjacency tooling is what most organisations now use to choose them.
Direction matters more than distance
Here is the finding that should change how you read an adjacency number. In Research Policy, Neffke, Nedelkoska and Wiederhold (2024, 53(2)) tracked German work histories from 1975 to 2010 and found that displacement increases the probability of occupation change eleven-fold. Crucially, the earnings consequences depended on the type of mismatch the move created, not merely its size:
- Skill shortage — the new occupation demands skills the worker does not yet have. These workers returned relatively quickly to the earnings path they would have been on anyway.
- Skill redundancy — the new occupation leaves skills the worker already has unused. These workers were put on paths with permanently lower earnings.
An adjacency score is symmetric: role A to role B scores the same as B to A. The evidence says those are not the same move. A 0.82 similarity can be a stretch upward, which tends to resolve, or a write-down of accumulated capability, which tends not to. The score cannot tell you which, because level and difficulty are not in it.
What to do instead: an eight-step redeployment check
- Treat the top ten as hypotheses, not conclusions. Adjacency is excellent at generating a shortlist of target roles and useless at closing one.
- Classify the direction of every pair — stretch, genuinely lateral, or write-down — using level and importance data for the critical requirements, not the similarity score.
- Decompose the pair into the requirements that actually differ. Typically two to four. The overlap needs no evidence; you already have it in current performance.
- Measure only the gap. Re-testing the whole target role wastes candidate time and tells you things you know.
- Match method to requirement. A work sample for the technical gap; a scenario-based judgement measure for the behavioural gap. Running two or three targeted measures on pay-as-you-go credits at ₹30 / US$0.50 each — AssessAll gives 250 free credits on corporate signup — costs less than one week of a failed redeployment.
- Set the bar before you see the scores. A cut score chosen after the fact is a description of the candidate pool, not a standard.
- Baseline, then bridge, then re-measure. If you fund bridging training without a pre-measure, you cannot distinguish a worker who was ready from a programme that worked.
- Audit the tool against your own job analysis. Take twenty role pairs your managers know well, compare the tool's ranking with theirs, and find out where the two disagree before you trust the other nine hundred.
When the adjacency score alone is the right answer
Sometimes it is, and pretending otherwise wastes money:
- Career exploration and self-service browsing. This is what the O*NET matrix was built for, and for showing an employee what adjacent work exists it is better than anything a manager would produce from memory.
- Shortlist generation at scale. Narrowing 900 occupations to 10 is exactly what the score is good at.
- Moves within a job family where the assessment evidence already exists. If the same validated measure covers both roles and the requirement levels match, the adjacency score is simply confirming what your evidence already says.
What it cannot do is carry a decision about an individual. Under the Uniform Guidelines on Employee Selection Procedures, a "selection procedure" includes promotion, transfer and retention decisions — not just external hiring. If a similarity score is what moved one person forward and held another back, it is functioning as a selection procedure, and the job-relatedness evidence has to exist. The Standards for Educational and Psychological Testing make the same point from the measurement side: content similarity supports an inference about a job's requirements, not a substitute for evidence about a person's standing on them.
Takeaway
An adjacency score tells you which roles resemble each other and in what direction to look; it never tells you who is ready. Use it to generate the shortlist, then spend a small amount of measurement on the two or three requirements that actually differ — and record which direction each move runs in, because the evidence says that is where the cost hides.