Work
Looking for work online
The most repeated statistics about online job hunting have no primary source, while the ones that do exist measure discrimination and referral value rather than software.
The three statistics that dominate online job-search advice — that 75% of CVs are rejected by applicant tracking software before a human sees them, that 70% to 85% of jobs are filled through networking, and that recruiters spend six seconds on a CV — do not rest on published research. The first two have no locatable primary source at all; the third traces to small eye-tracking studies run by a CV vendor rather than to peer-reviewed work. What has been measured well is different: employers report that their own software filters out candidates they wanted, referred applicants are worth more to firms mainly because they quit less, and identical applications with different names still receive different callback rates. None of it identifies whether a machine or a person did the rejecting. For a practical commercial comparison with this kind of workplace measurement, see top remote companies.
The 75% rejection figure has no primary study
No research source for it exists. It travels through resume-optimisation vendors and career-advice sites, and it misdescribes the software: mainstream applicant tracking systems rank and search candidates, they do not auto-reject on a missing keyword. The vendor most responsible for the anxiety publishes adoption figures instead.
Jobscan states that over 98% of Fortune 500 companies use an applicant tracking system and that 99.7% of recruiters use keyword filters within it.
Direction: Increase. Strength of evidence: Weak.
Caveat Commercially interested — Jobscan sells CV-scanning subscriptions premised on fear of these systems — and gives no primary source, sample or date for either figure.
Note what the page does not say. It claims near-universal adoption, never a 75% rejection rate, and cites no source for one. The number is attributed to vendors who do not make it.
What employers believe their own software does
The closest defensible evidence on over-filtering is an employer survey, measuring what employers think rather than what happened to candidates.
88% of employers said qualified high-skill candidates are filtered out by their own recruitment software, 94% said the same of middle-skill candidates, and 48% said an employment gap over six months triggers automatic disqualification.
Direction: Increase. Strength of evidence: Mixed.
Caveat Employer self-perception from a survey co-produced with Accenture, which sells talent-technology consulting, and no count of candidates wrongly rejected.
The six-month gap rule is the most actionable item there, and it is still a configuration choice reported by employers, not an audit of outcomes.
The hidden job market rests on one 1969 survey
The claim that most jobs are never advertised traces back to Granovetter's argument that job information travels disproportionately through weak social ties rather than close ones (Granovetter, American Journal of Sociology, 1973). The theory held up far better than its evidence base, a small survey of professional, technical and managerial men in one Boston suburb in 1969, and the percentages now attached to it were never national statistics. It has since been tested at scale, on a platform able to randomise who was introduced to whom.
Weak ties increased job transmission, but only up to a point, after which returns diminished, and the size of the effect varied by industry.
Direction: Increase. Strength of evidence: Mixed.
Caveat Two authors were LinkedIn employees and the platform experimented on users without specific consent, drawing research-ethics criticism; the result qualifies Granovetter rather than confirming him.
Weak ties help, with diminishing returns, in some industries more than others — a much smaller claim than careers advice makes, and silent on how many vacancies go unadvertised. Whether an online presence changes this is examined in what a personal brand does to hiring.
Referrals have a measured value, mainly to the employer
The evidence on referrals is unusually good, because large firms shared hiring and performance records.
Referred applicants were more likely to be hired and to accept offers, much less likely to quit, and generated higher profits per worker, mainly through lower turnover and recruiting costs rather than measured productivity.
Direction: Increase. Strength of evidence: Strong.
Caveat Nine firms in three industries with non-random referral status, measuring referrals' value to employers rather than the share of hires coming through them.
Referrals pay off because referred workers stay, not because they perform better once hired. That is why firms run bonus schemes for them, and it is not evidence about how many jobs require contacts.
Discrimination is measured at the firm, not attributed to the software
The rigorous work here predates online applications. In the original US correspondence experiment, resumes with white-sounding names received 50% more callbacks than identical resumes with African-American-sounding names (Bertrand & Mullainathan, American Economic Review, 2004). It answered newspaper advertisements in 2001-02, before online application systems existed, so it says nothing about applicant tracking software despite being cited as though it did.
The modern replication is much larger and locates the problem in specific employers rather than the labour market as a whole.
Distinctively Black names received contact 2.1 percentage points less often than distinctively white names, with the top quintile of firms accounting for nearly half of all lost contacts and 23 companies identifiable as discriminating.
Direction: Decrease. Strength of evidence: Strong.
Caveat Entry-level high-volume roles at Fortune 500 employers only, measuring callback at the first screening stage, and unable to separate algorithmic screening from human recruiters.
That limitation is the one most often ignored. These studies measure a callback gap at firm level: they identify which companies discriminate, not which step does it, so citing them as proof that algorithms discriminate claims more than they establish. What automated systems can and cannot be shown to do is covered in the evidence on automation and jobs.
The short version
- No primary study supports the claim that 75% of CVs are rejected by applicant tracking software, and the vendor most associated with the fear publishes only adoption figures.
- 88% of employers surveyed by Harvard Business School and Accenture in 2021 said their own software filters out qualified high-skill candidates, a belief and not an audit.
- Randomised experiments covering over 20 million LinkedIn users found weak ties do transmit jobs, with diminishing returns and wide variation by industry.
- Referred applicants generate higher profits per worker mainly by quitting less, not by performing better once hired.
- The evidence has a hard limit: correspondence studies find a 2.1 percentage point callback gap by name, concentrated in a fifth of firms, but cannot say what caused it.