How this research was made

Who is in this data, and who is not

Respondents are people who found the free MyPassion.AI career quiz and chose to take it, unprompted, while looking for career direction. Nobody was recruited into a panel, quota-sampled, or paid to answer. There was no incentive to complete beyond receiving the quiz result, and respondents were not told a report would be produced.

That makes this a self-selected sample of a specific population: people actively questioning their career at the moment they answered. It is the right frame for questions about what career doubt looks like from the inside, and the wrong frame for questions about how common career doubt is. The distinction matters for every figure in the section, so it is stated on every page rather than only here.

This data describes the people inside the question. It does not measure how many people are in it.

Collection window and monthly volume

Collection opened on 29 January 2026 and the dataset was frozen on 28 July 2026, a window of 181 days. Volume was not even across it. January and July are both partial: January covers three days, and July is where the quiz reached its largest monthly audience so far. Any month-on-month reading of this dataset has to account for that, which is why the report does not publish a trend line over the window.

Why every statistic states a different base

The quiz grew over the collection window, from 20 questions to 26. Questions added partway through were never shown to earlier respondents, and several questions can be skipped. Coverage therefore varies by field, from 95.3% for life situation down to 34.1% for the flow trigger, which was among the last added.

Rather than impute missing answers or restrict everything to the smallest common sample, each figure is computed on the respondents who answered that specific question. That keeps the larger questions at full strength, at the cost of requiring every number to travel with its base. That is why no percentage anywhere in this section appears without one.

What was removed before publication

All free text. Job titles, background descriptions, dream-work answers and the optional write-in follow-ups were removed from the published dataset. Free text in a career quiz routinely contains employer names, job titles specific enough to identify one person, and personal circumstances. None of it is recoverable from the CSV. No identifiers of any kind are published: no user IDs, no session IDs, no user agents, no timestamps finer than the month.

Superseded question versions. Where a question's answer set was revised, responses recorded against the old set were excluded rather than mapped onto the new one. Remapping would attribute to a respondent a choice they were never shown. This is the single largest exclusion in the dataset and it is the reason the published bases are smaller than the raw response counts.

The small-cell rule

Two rules, both applied mechanically by the script that builds the figures rather than case by case, so neither can be relaxed for a number that would otherwise be interesting.

Row floor. In any cross-tab, a row whose base falls below 100 respondents is suppressed entirely: its cells publish as "n/a". Below that threshold a single respondent moves the figure by more than a whole point, which is a precision the base cannot carry.

Cell floor. An individual percentage backed by fewer than 10 people is suppressed even when its row base is large. This one is disclosure control rather than precision: a cell standing for a handful of people, crossed with the two attributes that define it, starts to describe individuals rather than a population. Because of this rule, published cells in a row do not always sum to 100.

Rows that clear both floors but sit close to them are flagged in the text where they appear, because clearing a threshold is not the same as being sturdy.

Why the row-level file is not released

The dataset carries no direct identifiers. No names, no email addresses, no user or session IDs, no free text, and no timestamp finer than the month. On that measure it is clean, and it was built that way from the first export.

That is not the same as being anonymous, and we would rather say so than imply otherwise. Each row carries nine attributes, and across 6,837 rows most combinations of those nine occur exactly once. Someone who already knew that a specific person had taken the quiz, and knew roughly where they live, what they studied and how long they have worked, could plausibly pick out that person's row and read what they answered. Generalising the coarsest fields does not fix it: the uniqueness comes from the number of attributes, not from any one of them.

So we do not hand out the response-level file. What we do hand out is the aggregate tables behind any published figure, and custom cuts of them, which is what a reporter checking a number actually needs. Those are governed by the two suppression rules above, so no published cell stands for fewer than 10 people.

What this data cannot show

It is not prevalence. The most common misreading of this dataset would be to treat a share of respondents as a share of workers. When 19.1% of respondents say they are well paid and want out, that describes the composition of people taking a career quiz, not 19.1% of any workforce.

It is self-reported and self-diagnosed. Every field is the respondent's own description of themselves, captured at one moment while they were actively unsettled about work. Someone who calls themselves stuck or bored is reporting a feeling, not a clinical or occupational assessment. The report's boredom figure in particular is a measure of disengagement among people questioning their career and is not a burnout prevalence estimate.

It is a snapshot, not a trend. One collection window, no repeat measurement of the same people, and uneven monthly volume mean nothing here supports a claim about change over time. Cross-sectional differences between groups are what the data can carry.

It is correlational. Where two fields move together, such as education level and naming boredom, the data shows the association and nothing about direction or cause. Respondents were not randomised into anything and there is no control group.

The audience is English-language and search-led. The quiz is in English and most respondents arrive through search, which shapes who is present. Country coverage spans 156 entries but is self-reported free choice, with a long tail of low-count values, so only top-line counts and the largest shares are published.

We collected it and we benefit from it. MyPassion.AI operates the quiz that produced this data and sells a paid career report. That is a reason to publish the method in full and to state every base, which is what we have done, and to hand out the aggregate tables behind any figure to anyone who wants to check it.

Re-deriving any published figure

The dataset has ten columns: month, country, education, years of experience, life situation, core struggle, work style, flow trigger, completion flag, and duration band. Nearly every figure in the section is one of two operations on it.

For a single distribution, filter to rows where that column is non-empty and take the share of each value. That base is the n printed beside the figure. For a cross-tab, filter to rows where both columns are non-empty, group by the row column, and take each value's share within that group. Cells read across a row and sum to 100 within it.

We share the response-level file with journalists and researchers on request. EmailEmail the research team with a line about what you are working on. Every chart, table and figure already published across the research section is free to reuse right now under CC BY 4.0 with a link back.