Methodology

The three tiers of scientific backing

Every element of our recommendation system falls into one of three tiers, and we label each explicitly:

1. Published instruments, and adaptations we label — two published instruments with reliability data (the O*NET Interest Profiler Short Form for Holland RIASEC, and the IPIP 50-item Big-5 markers). We also run short adaptations of SRQ-A, the Passion Scale and Schwartz's values, and a VARK-style learning-preference quiz; these are shortened or reworded versions, so we do not claim they are validated. An ICAR-style reasoning screener's items are our own and are labelled as practice estimates. Our 16-type personality quiz is our own and is labelled as a quiz, not a validated instrument.
2. Research-informed direction — the direction each instrument contributes to each axis is defensible from published literature. Effect sizes come from Schmidt & Hunter (1998), Sackett et al. (2017), and instrument-specific meta-analyses.
3. Pragmatic engineering weights — the exact numeric weights (e.g. RIASEC 0.30, Big-5 Agreeableness 0.08) are our engineering defaults, not derived from a meta-analysis. We calibrate them internally and publish the values here so the choice is transparent.

Instruments we use (Tier 1 — adaptations marked)

• Holland RIASEC — Holland (1997), measured with the O*NET Interest Profiler Short Form (Rounds, Su, Lewis & Rivkin 2010; USDOL/ETA, CC BY-ND 4.0). Mean Cronbach's α ≈ 0.81 across the six scales (O*NET Center, 2010).
• Big-5 (IPIP 50-item markers) — Goldberg (1992), public domain (ipip.ori.org). Personality dimensions, shown as percent of the scale maximum; no norms are used.
• SRQ-A (adapted) — 16 items adapted from the Academic Self-Regulation Questionnaire (Ryan & Connell, 1989): Self-Determination Theory motivation quality. Our wording and length differ from the published scale, so it is not claimed as validated.
• Passion Scale (adapted) — 12 items adapted from Vallerand et al. (2003): harmonious vs obsessive passion. A shortened version; not claimed as validated.
• Prosocial values (adapted from Schwartz) — 6 items on benevolence and universalism, adapted from the Portrait Values Questionnaire (Schwartz, 1992 value theory). A short subset, not the full validated questionnaire.
• ICAR-style reasoning screener — item formats from Condon & Revelle (2014). Our items are written for this site and are not calibrated ICAR items, so results are raw scores with practice-estimate bands, not percentiles.
• Learning-preference quiz (VARK-style) — 12 items written for this site around Fleming's four modes (visual, aural, read/write, kinaesthetic). VARK® is Fleming's copyrighted questionnaire; ours is not the official VARK and is not validated. Research gives little support to matching teaching to a 'learning style', so we use it only for study tips.

Personality type quiz (16 types) — our own 20-item forced-choice quiz, not a validated instrument, so it sits outside Tier 1 and is used for self-reflection only. Inspired by Jungian type theory. Not the official MBTI® assessment.

The four Ikigai axes

❤️ Love — what you enjoy

Weights: interest_content 0.70 · personality_fit 0.30 · RIASEC 0.30 · Vallerand passion 0.20. Includes SRQ-A multiplier (0.6–1.4 based on Relative Autonomy Index).

💪 Good at — where you excel

Weights: subject_aptitude 0.25 · strength_match 0.15 · follow_through 0.05 · CHC 0.30 · behavioural taste-test 0.25. Weights rebalance when instruments are missing.

💰 Paid for — economic viability

Weights: salary_level 0.50 · salary_consistency 0.20 · education_ROI 0.30. Big-5 Neuroticism tilts consistency weight by ±0.10.

🌍 World needs — contribution

Weights: mission 0.20–0.30 · Schwartz prosocial 0.15 · market_demand 0.22–0.25 · AI_durability 0.22–0.25 · sustainability 0.20–0.21 · Big-5 Agreeableness 0.08.

The composite

The four axes converge through a harmonic mean (deliberately punishing to any single weak axis, faithful to the Ikigai model). Then a piecewise linear calibration maps raw scores 0–10 to a displayed 0–10 range where accurate matches read as 7+.

Calibration formula:

• raw ≤ 4 → displayed = raw × 1.75
• raw > 4 → displayed = 7 + (raw − 4) × 0.5

So raw 4 = displayed 7, raw 5 = displayed 7.5, raw 6 = displayed 8, raw 10 = displayed 10. Purely a UX choice, not a psychometric one.

Two ranker adjustments before calibration

• Authenticity multiplier — decision drivers from the retrospective intake produce a −1..+1 score, applied as 1 + 0.30 × score on Ikigai-seed component weights. Grounded in Sheldon & Elliot (1999) self-concordance model.
• Regret penalty — past-decision regrets keyword-matched to SOC prefixes; matching careers get raw × (1 − penalty). Pragmatic.

What's proprietary vs validated

• The Career DNA quiz, Career Taste Test, and the personalisation questionnaire (Ikigai seed) are proprietary discovery tools — not validated psychometrics. They contribute signal but shouldn't replace Holland / Big-5.
• The Ikigai four-circle Venn diagram is a 2014 Western adaptation by Marc Winn, not authentic traditional Japanese ikigai. Useful framework nonetheless.

Data sources

• Occupation profiles — O*NET 30.0 (tasks, skills, knowledge, entry education) and U.S. Bureau of Labor Statistics Occupational Employment and Wage Statistics (US pay) and Employment Projections
• University data — College Scorecard (US Dept. of Education) + NIRF (India) + UCAS (UK) + public university websites
• Pay outside the US — India: Government of India, MoSPI Periodic Labour Force Survey (PLFS) Annual Report 2025, Table 51: average monthly earnings of salaried employees for each occupation group (NCO-2015 division), urban and rural. It is one average per group, not a per-career figure; careers are matched to groups with the official BLS SOC↔ISCO-08 crosswalk (ESCO careers by their ISCO-08 code). UK: ONS Annual Survey of Hours and Earnings (ASHE) 2025, by occupation. Indian institutions' graduate salaries are the NIRF 2025 medians each institute reported; the few institution figures on Rankings that are our own estimates are labelled as estimates. US pay figures are shown in US dollars and are never converted into rupees as Indian pay. We don't use salary websites such as AmbitionBox or Glassdoor.
• AI-exposure scores — our own rubric (v2), not any one paper's numbers. Every career is scored on seven dimensions grounded in four published sources: Eloundou et al. (2023) on LLM task exposure, Frey & Osborne (2013) on automation susceptibility, Felten, Raj & Seamans (2021) on AI occupational exposure (AIOE), and Nedelkoska & Quintini (2018, OECD) on interpersonal and regulatory barriers. Each O*NET career has an anchor scored from its O*NET task statements, with a review date (every six months); anchors added in September 2026 were drafted with an LLM and are awaiting human review. Careers from ESCO without an O*NET match keep a stored score from the nearest anchored career.

Datasets and how current they are

Built from the files the site actually loads. "As of" is the date in the file's own record (publication, release or the day we read the source); a blank means the file doesn't record one.

Third-party data attributions

WhatTNext AI stands on the shoulders of several publicly-licensed datasets. We disclose each below, per each publisher's licence terms.

• O*NET-SOC (US Department of Labor) — occupation descriptions, skill/knowledge lists, and AI-exposure baselines are sourced from O*NET Online, a service of the US Department of Labor / Employment & Training Administration. Used under the O*NET Data Licence.
• ESCO (European Commission) — modern occupation titles (Data Engineer, ML Engineer, DevOps, etc.) not covered by O*NET are sourced from ESCO v1.2+. This service uses the ESCO classification of the European Commission. ESCO occupations are cross-referenced with O*NET where a crosswalk exists, and inserted as supplementary rows (identified by source="esco") where no O*NET analog exists. We have modified ESCO by generating pseudo-SOC codes (ESC-XXXX) for internal keying and by merging with O*NET; these modifications are indicated in the data itself via the source and esco_uri fields. The Commission does not guarantee the accuracy of ESCO data and shall not be liable for consequences of use.
• Lightcast Open Skills — skill category and taxonomy metadata is sourced from Lightcast Open Skills under CC-BY-4.0. Used for canonical skill naming and category rollups.

McKinsey durability framework

Several surfaces on this platform — the Skills Browser, My Report's Durable Skills Portfolio panel, the Pivot Planner adjacency count, and the AI Fluency vs Technical AI tagging — are framework informed by McKinsey Global Institute's Agents, robots, and us: Skill partnerships in the age of AI (November 2025). The report is the source of the eight high-prevalence "durable" skills (communication, management, operations, problem-solving, leadership, detail orientation, customer relations, writing), the AI Fluency vs Technical AI distinction, and the skill-adjacency-count mechanic.

Skill outlook (🔻 declining, 🔄 evolving, 🏛️ enduring): this is our own estimate, not McKinsey's Skill Change Index and not a demand forecast. For each ESCO skill we take the occupations that need it (ESCO occupation–skill links, placed in ISCO-08 groups) and average our AI Exposure Rubric v2 scores for them (carried from US SOC with the BLS crosswalk). A skill is declining only when that work is both highly exposed (6/10 or more) and routine (rubric dimension D2, 6/10 or more) — and never when it is an AI Fluency or Technical AI skill; enduring when exposure is under 4/10; otherwise evolving. The data carries sci_source: "rubric_v2_esco". Prevalence percentages we cite for the 8 durable skills (99% communication, 94% management, etc.) are drawn directly from the McKinsey report and used descriptively.

Selected citations

• Schmidt & Hunter (1998). "The validity and utility of selection methods in personnel psychology." Psychological Bulletin.
• Sackett, Zhang, Berry & Lievens (2017). "Revisiting meta-analytic estimates of validity in personnel selection." Journal of Applied Psychology.
• Ryan & Connell (1989). "Perceived locus of causality and internalization." Journal of Personality and Social Psychology.
• Vallerand et al. (2003). "Les passions de l'âme: On obsessive and harmonious passion." J. Pers. Soc. Psychol.
• John, Naumann & Soto (2008). "Paradigm shift to the integrative Big Five trait taxonomy."
• Holland (1997). Making vocational choices: A theory of vocational personalities and work environments.
• Condon & Revelle (2014). "The International Cognitive Ability Resource." Intelligence.
• Schwartz (1992). "Universals in the content and structure of values." Advances in Experimental Social Psychology.
• Sheldon & Elliot (1999). "Goal striving, need satisfaction, and longitudinal well-being."
• Eloundou, Manning, Mishkin & Rock (2023). "GPTs are GPTs: An early look at the labor market impact potential of large language models." arXiv:2303.10130
• Frey & Osborne (2013). "The future of employment: How susceptible are jobs to computerisation?" Oxford Martin School.
• Felten, Raj & Seamans (2021). "Occupational, industry, and geographic exposure to artificial intelligence: A novel dataset and its potential uses." Strategic Management Journal 42(12). Dataset. One of the four sources behind our AI-exposure rubric (the physical/dexterity dimension); we don't use its scores directly.
• Nedelkoska & Quintini (2018). "Automation, skills use and training." OECD Social, Employment and Migration Working Papers, No. 202.
• Fleming & Mills (1992). "Not another inventory, rather a catalyst for reflection." To Improve the Academy 11.
• Pashler, McDaniel, Rohrer & Bjork (2008). "Learning styles: Concepts and evidence." Psychological Science in the Public Interest 9(3).