We Traced 47 Hiring Statistics to Their Sources. 11 Survived.

Somewhere in the last decade, the career-advice industry agreed that recruiters spend about six seconds looking at your resume. Or seven point four. Or eleven. Or thirty. Or a full minute.

All five of those numbers are in print right now, on pages dated 2026, each stated as fact. We went looking for where they came from.

Then we kept going. We pulled every statistic off 27 of the best-ranking career-advice pages on the web — 3,269 numeric claims in total — picked 47 of the most repeated, and walked each one backwards, hop by hop, to whatever was underneath it.

Ten of the 37 claims we traced from other publishers survived unchanged. The other 27 had drifted: wrong population, wrong year, wrong unit, a sample size borrowed from a different study, or no original at all.

We also audited our own site by the same method, and it did worse. That section is further down, and it is not decorative.

Nobody can answer the genre’s simplest question

“How long do recruiters look at a resume?” is the single most-repeated fact in career advice. Across our corpus it has at least eleven mutually incompatible answers, spanning six seconds to three minutes — a thirty-fold range.

Two pages contradict themselves. One article states that 24% of hiring managers spend less than 30 seconds on a resume, and — further down the same page — that recruiters spend 60 seconds. Another says one in four hiring managers spend under 30 seconds, then says 57% spend one to three minutes. Neither page acknowledges its own discrepancy.

The famous numbers trace like this.

Six seconds comes from a 2012 eye-tracking study by the recruitment firm TheLadders. We obtained the original: a five-page document whose entire stated methodology is one sentence — “Thirty professional recruiters took part in the study during a 10-week period.” No equipment, no recruitment method, no number of resumes, no dispersion, no peer review. It is a sales document for the company’s resume-rewriting service.

7.4 seconds is the same firm’s 2018 follow-up, which supersedes it. We obtained that one too. Its methodology section is three sentences and it publishes no sample size at all.

Which brings us to the detail that explains a decade of confusion. The phrase attached to “7.4 seconds” almost everywhere is “a study of 30 recruiters.” That sample belongs to the 2012 study, not the 2018 one. Neither publication ever made that pairing.

We found out why. Both reports were published at the same filename — TheLadders-EyeTracking-StudyC2.pdf. Archive records show that URL serving the 2012 document continuously from April 2012 through January 2019. The 2018 press release links to no PDF at all. So every journalist covering the 7.4-second finding in November 2018 who clicked through to “the study” was handed the 2012 paper, with its 30-recruiter methodology printed on the page.

It was not carelessness. It was a link pointing at the wrong document for the entire period the story was being written, and everything downstream inherited it.

So how long is it, really?

The honest answer is a range with its instruments attached, not a number.

That last row is the most telling. The most rigorous instrument ever pointed at this question did not report the number everyone quotes — and the academic literature’s own background citation for screening duration is a self-report survey from 1991.

Nothing supports “six seconds”, “7.4 seconds”, or “under ten seconds” as a current figure. Two sound 2024 surveys put most initial reviews under two minutes and almost none under ten seconds — and they still disagree with each other by twenty points.

What we checked, and how

We took the top-ranking pages for six seed queries — resume statistics, job search statistics, applicant tracking system statistics, cover letter statistics, hiring statistics, and interview statistics — and extracted every sentence containing a number. That produced 3,269 raw claims across 27 pages belonging to 21 publishers.

From those we selected 47 claims to trace in full: the resume-screening-time family, the ATS rejection and adoption family, the claims carried by three or more independent publishers, and every numeric claim on our own site that fell in scope.

For each one we walked the citation chain backwards, recording every hop with a URL and a retrieval date, until we reached a primary source or ran out of chain. Then we checked four things: does the source state this number, for this population, in this unit, and when was the data actually collected.

Before any of that, though, there is a cruder measurement worth reporting, because it requires no judgment at all — only counting.

Of the 3,269 claims, 286 carry a link to a source. That is 8.7%. Seventeen of the 27 pages carry no source links anywhere, and those seventeen pages account for 75% of all the claims in the corpus.

Four of those seventeen do name sources in prose. One appends a seventeen-entry bibliography with study titles, sample sizes and dates — careful work — entirely unhyperlinked, so not one of its claims is one click from verification.

The results

We want to be careful about what that table does and does not say. It is not a finding about honesty. In every single case we traced, we could see how the error happened, and in most of them nobody did anything deliberate. The most common cause is not invention — it is a number passing through one summary too many.

How a number loses its provenance

The same handful of mechanisms turned up over and over. These are the ones worth naming, each with the clearest case we found.

One sub-sample becomes two “independent” sources

Two claims sit in the corpus looking like corroboration. One page says “24% of hiring managers spend less than 30 seconds reading each resume,” credited to CareerBuilder. Another says “1 in 4 hiring managers spend less than 30 seconds looking at a resume,” credited to Zippia, 2023. Different wording, different source, different year. A reader counting sources scores this as two findings that agree.

They are the same datum. The primary source prints both forms in a single sentence: “Nearly 1 in 4 (24 percent) spend less than 30 seconds.” It is a CareerBuilder/Harris Poll release from September 2016, fieldwork May–June 2016, n=2,153 — and the 24% is the roughly 251-person HR-manager sub-sample, which the release explicitly distinguishes from what it calls “the actual hiring managers.”

So a 251-person sub-sample from 2016 became two differently-worded, differently-attributed, differently-dated statistics that appear to confirm each other. The one heuristic every careful reader uses — if several independent sources agree, it is probably true — is defeated by the citation chain itself.

An indictment of employers becomes an instruction to candidates

This is the one with real consequences, and it implicates our own industry directly.

A widely-circulated claim holds that 88% of employers believe ATS systems are screening out highly qualified candidates whose resumes lack ATS-friendly formatting. It is carried by four independent publishers. Trace it and every trail converges on one aggregator page that cites nobody and ran no survey.

The real 88% is from Harvard Business School and Accenture’s Hidden Workers: Untapped Talent, September 2021 — Figure 10, page 25. The actual question asked was:

“Do you think your organization’s hiring system filters out… candidates who could successfully perform the job, but don’t fit the exact criteria in the job description?”

It was asked of 2,275 executives across Germany, the UK and the US, fielded in January–February 2020. The 88% is the high-skills half of an 88/94 pair — the middle-skills figure is higher still. The phrase “ATS-friendly” appears zero times in that report.

The mechanism was swapped. The finding was that employers write job descriptions rigid enough to screen out people who could do the work. The circulating version tells those people the problem is their resume file.

That inversion turns a criticism of hiring practice into a reason to buy resume-formatting help — which is a category we sell into. We benefit commercially from the misreading, and saying so is the price of publishing this.

A ratio assembled from two unrelated paragraphs

“On average, a corporate job attracts 250 applicants, and only 2% are invited to interview” may be the most-quoted figure in job-search advice. Every hop in its chain resolves. The chain ends nowhere.

It runs from today’s career blogs, to Inc., to a Glassdoor eBook from January 2015, to Glassdoor’s own now-dead citations page, which credits “ERE Survey, 2013.” That resolves to a May 2013 opinion column on ERE.net whose sentence carries no citation at all. There was no survey; an opinion column was relabelled as one.

And the ratio is manufactured. In that column, the “250” and the “4–6 interviews” are in separate paragraphs with separate sources, and the 4–6 figure’s denominator is 100 completed applications out of 1,000 job-post views — not 250 resumes. The two were welded into one sentence in 2015, producing a ratio that exists in neither source. The “2%” is arithmetic performed on that splice.

Glassdoor quietly dropped the figure from its own 2017 and 2018 editions while footnoting every other statistic in them. The originator walked away. The genre kept it.

A vendor survey acquires a university byline

One claim in our set is credited to a named university. Following it, the university’s careers page turns out to be a verbatim republication of a commercial survey’s article — methodology paragraph included, bylined to a university career counsellor. The giveaway is that the page’s hero image is still served from the original vendor’s CDN.

The survey underneath is n=200, and it does not state which country its respondents were in. One hop later it is being cited as academic.

A number gets revised, and the revision is not disclosed

One trend claim — “81% of US employers now use skills-based hiring, up from 73% in 2023 and 57% in 2022” — fails in four ways at once, three of them at the primary source.

The sample cited belongs to a different edition than the number. The word “US” is false: that year’s employer sample was 22% US. The trend stacks four annual surveys whose sample sizes ran 2,736 → 1,500 → 1,019 → 1,084 and whose country rosters changed every year. And the publisher’s 2022 report states that 76% of the businesses it surveyed used some form of skills-based hiring, while its own 2023 and 2024 reports retroactively describe that same survey as 56% — a twenty-point revision, in near-identical wording, disclosed nowhere.

All four carriers reproduced it faithfully. The defect is upstream of all of them.

What actually held up

A study that found nothing verifiable would be a study with a broken method, so it matters that plenty checked out.

The cleanest case was a labour-market claim: “6.9 million open jobs and 7.6 million unemployed workers, February 2026.” It names both Bureau of Labor Statistics release products — JOLTS and the Employment Situation — date-stamps the reference month, and matches. It also survives a revision check: the JOLTS figure was later revised from 6,882,000 to 6,922,000 and is still 6.9 million.

The best-documented source in the entire audit was Gallup: a Q4 2025 release with 22,368 respondents, a probability-based panel, a ±1.0 margin of error, and a methodology table covering sixteen quarters. It was also the only non-vendor primary source we found anywhere in the set. The page citing it used a numbered footnote that resolved, and the target contained the claim verbatim, including its internal breakdown.

We also found genuine independent corroboration once. The claim that 19% of resumes include a hobbies and interests section is supported by two unconnected censuses an order of magnitude apart — one of 31,000 resume-builder users, one of 133,289 resumes — reaching the same integer. That is what real agreement looks like, and it looks nothing like the manufactured kind.

And in several cases a publisher we expected to find careless turned out to be the careful one. One of the highest-ranking resume-statistics pages in our corpus debunks both the 75%-ATS-rejection myth and the six-second figure on its own page, dating them correctly and writing plainly that neither number is current.

The pattern is about page type, not publisher

The most useful thing we learned is that the failures do not sort by company. They sort by format.

One publisher in our corpus operates both the best-cited page we traced — correct sample, correct figure, fieldwork month stated — and the worst rendering in the entire corpus of the same underlying claim, with an invented country, no attribution, and the source absent from its own 40-entry reference list.

Variance inside a single publisher exceeded variance between publishers.

The difference is what kind of page it is. A dedicated research page publishes its instrument, its sample and its field dates. An “N statistics for [year]” listicle strips all three from figures that were fine at source. Same organisation, same week, opposite standards.

That reframes the whole problem. It is not that the industry lies. It is that one very popular format systematically discards the metadata that makes a number checkable — and once discarded, nothing downstream can tell a number that had provenance from a number that never did.

We audited ourselves, and we failed too

Auditing other people’s sourcing while leaving your own unchecked would make this piece worthless, so our own site went through the identical protocol.

We found 114 numeric claims across wensaresume.com. Of the 10 that fell inside the traced set, exactly one verified as published. Thirty-eight corrections went live across 19 URLs before this study was published.

The worst of them was ours alone. One of our pages stated four times, including in an FAQ, that “a 2026 benchmark that tested how eight major ATS platforms parse real resumes found that two-column layouts failed to parse correctly in 7 out of 8 systems.”

No such benchmark exists. Exactly one page on the open web carries that figure: a vendor post with an organisational byline and no named author, which declines to name the eight systems, tested demo environments rather than the platforms, used a single fictional resume, and never defines what “failed” means. Our version was stronger than its source — “eight major ATS platforms” and “real resumes” were our additions — and we stated it as settled fact.

A second page of ours cited a real test by name and attributed two figures to it that do not appear in it, and reported the direction of its finding backwards. That page now states what the test actually measured.

We also carried the 7.4-second claim, unlinked. We described an adoption statistic as screening behaviour on five surfaces — the exact drift documented above. We attributed a vendor’s figures to a professional association. And we asserted that most companies only review candidates scoring above 70–80% on an ATS match score, a mechanism that does not exist as described: the major enterprise systems output named categories, and one states in its own documentation that it does not automatically advance or reject candidates.

One correction is worth singling out because nobody was careless. We dated a survey to 2023 on three pages. Its fieldwork was 2022 — but the publisher itself labelled it 2023 from 2024 until February 2026, then corrected it. Our pages copied a source that was wrong at the moment we copied it and has since been fixed. The citation simply froze.

What we got wrong

Twelve errors were found in our own research process while it was running, all by the verification step catching the summarisation step. They are worth listing because they are the same errors this study documents.

We manufactured a false consensus. Our clustering tool grouped claims by their number, and merged “52% of job seekers were ghosted after an interview” with “52% of companies have an interview process that is too long.” We briefly believed, and said internally, that the ghosting figure was carried by six publishers. It is carried by two. A study about numbers that look corroborated and are not, produced one — by the same mechanism.

We fixed the tool, then tested the fix against seven claims whose real carrier counts had been established by reading live pages. It got four of seven, up from one of seven. Wrong in both directions. No threshold fixes it, because the failure is semantic. So no cross-publisher count in this piece comes from the tool — every one was established by hand, and the tool is used only to decide what to look at.

We introduced three errors while writing our own research briefs, by restating a summary instead of re-reading the source. One changed a file format. One changed a unit of measurement. One dropped an attribution the publisher had actually provided. All three were caught only because the people checking were instructed to verify against the source rather than trust the brief.

We inherited errors from a secondary source. Our first pass on the 75%-ATS claim relied on an existing published debunk. Three of its details were wrong, and we had repeated all three. We also cited a university-hosted copy of the 2018 eye-tracking report as a mirror — the file is byte-identical to the 2012 report, misfiled under the later name. We fell into the same trap we were documenting, while documenting it.

And one scope claim was unsupported. We described the welded “30 recruiters + 7.4 seconds” version as appearing across a large share of the career-advice web. Measured, it appears on zero of our 27 corpus pages. It is real, but it lives in the layer below the one we audited.

Other limits worth stating: we had no backlink data, so “most-repeated” means frequency across independent publishers, which is a proxy. We excluded the thin, recently-built sites that fill these search results, because auditing them proves nothing — but they are how an unsourced number becomes ubiquitous, and they are not in our counts.

How to check a statistic in about two minutes

Most of what we did is not specialised. It is four questions, and they catch the overwhelming majority of bad numbers.

And one shortcut: if a number appears on many pages in identical wording with no common source, that is not corroboration. It is one page, copied.

FAQ

How long do recruiters actually spend looking at a resume?

The defensible answer is a range, not a number. Two 2024 surveys of US hiring professionals — ResumeGo (n=418) and Zety (n=753) — both put most initial reviews under two minutes, and ResumeGo found only 1% spend under ten seconds. They disagree with each other by about twenty points on “under two minutes,” and no peer-reviewed study publishes a duration at all. Anyone quoting a single precise figure is quoting something that was never measured that precisely.

Is the “75% of resumes are rejected by an ATS” claim true?

No, and it is more interesting than simply false. The earliest datable instance is a March 2012 article that attributes it to a vendor called Preptel, and the original assertion was that applicant tracking systems “kill 75% of candidates’ chances of landing an interview” — a statement about probability, not a share of resumes discarded by a machine. Preptel ceased trading in August 2013 having published no study. The number in circulation is a unit transformation of a claim that was never about resumes being thrown away.

Does that mean ATS software doesn’t matter?

It matters, but for a different reason than the myth suggests. Applicant tracking software is close to universal among large employers, and it parses your resume into structured fields that recruiters then search and filter. If a skill never parses out of your document, nobody can find you by it. That is a real mechanism and a good reason to keep formatting simple. It is not the same as a machine rejecting you outright.

Which publishers were worst?

That framing does not survive the data. The clearest pattern we found is that the same publisher can produce the best-sourced page in our corpus and the worst, in the same period. The variable is the format — a dedicated research page versus a statistics listicle — not the organisation. We have not published a league table because the evidence does not support one.

Can I use this research?

Yes. Everything here is free to quote and cite. The per-claim dataset — all 47 claims with their carriers, verdicts and our confidence in each, our own claims included and flagged as ours — downloads from the link above. The full hop-by-hop chains, with every URL and retrieval date, sit behind those verdicts and are available on request; we will publish them in full alongside the next study. If you find an error in our work, tell us: we log corrections rather than quietly editing them.

Sources

Everything below was retrieved and read directly during September 2026. Where a page no longer resolves at its original address, the link goes to an archived copy and says so. Sources we found to be unsound are listed too, under their own heading — leaving them out would make the audit impossible to check.

Primary research and official statistics

• Bureau of Labor Statistics — Job Openings and Labor Turnover Survey, February 2026 reference month, released 31 March 2026.
• Bureau of Labor Statistics — The Employment Situation, February 2026, released 6 March 2026.
• Gallup — Worker Thriving Declines as Job Market Pessimism Grows, Q4 2025. n=22,368, probability-based panel, ±1.0pp, with a sixteen-quarter methodology table. The best-documented source in this audit and the only non-vendor one.
• Pina, Petersheim, Cherian, Lahey, Alexander & Hammond — Using Machine Learning with Eye-Tracking Data to Predict if a Recruiter Will Approve a Resume, Machine Learning and Knowledge Extraction 5(3):38, 2023. 221 recruiters, 2,043 resume views, Tobii Spectrum. Peer reviewed, and publishes no screening duration.
• Frieder, Van Iddekinge & Raymark — “How quickly do interviewers reach decisions? An examination of interviewers’ decision-making time across applicants,” Journal of Occupational and Organizational Psychology, 2016. 166 interviewers, 691 real interviews.
• Harvard Business School & Accenture — Hidden Workers: Untapped Talent, September 2021. 2,275 executives in Germany, the UK and the US; fieldwork January–February 2020. The 88% figure and its question wording are Figure 10, page 25.

Vendor and publisher research we relied on

• ResumeGo — How Much Time Do Recruiters Spend on a Resume?, 2024. n=418 US hiring professionals, SurveyMonkey, fielded 8 January – 1 March 2024. The only source in the whole audit that publishes its questionnaire as text.
• Zety — Recruiting Preferences Report, 2024. n=753, fielded 4–5 June 2024.
• Zety — Resume Statistics, a census of 133,289 resumes, 2022.
• Enhancv — What Happens to a Two-Column Resume When a Machine Reads It, 2026. 357 extraction runs across 17 templates and seven extraction tools.
• Jan Tegze — How Long Do Recruiters Look at a Resume?, 2023. 114 recruiters, three resumes each, timed without participants’ knowledge.
• interviewing.io — Are recruiters better than a coin flip at judging resumes? 76 recruiters, roughly 2,200 evaluations.
• Ashby — Recruiter Productivity, Talent Trends Report. Over 109 million applications across 247,000 jobs, January 2021 – March 2026. The largest dataset in this audit.
• CareerPlug — 2024 Recruiting Metrics Report. 60,000+ US small businesses, 10M+ applications.
• Greenhouse — Candidate Experience Report, 2024. 2,900 respondents across six countries; fieldwork April 2024. Superseded twice by Greenhouse itself since.
• Jobscan — ATS Usage Report, 2026. All 500 Fortune 500 careers pages reviewed; an ATS detected at 487.
• TestGorilla — The State of Skills-Based Hiring, 2024 edition.
• CareerBuilder / Harris Poll — annual survey release, September 2016. n=2,153, fieldwork 11 May – 7 June 2016.
• Greenhouse — Talent Matching documentation, on what its matching feature does and does not do.

Sources we traced and found unsound

These are cited because the audit’s conclusions depend on them being checkable, not because we recommend them.

• TheLadders — Eye-Tracking Study, 2012 (archived). Five pages; the entire methodology is one sentence. Source of the “six seconds” figure.
• TheLadders — Eye-Tracking Study, 2018 (archived; the live URL returns 403). Publishes no sample size. Source of “7.4 seconds”. Note both reports were served at the same filename, which is how the 2012 sample became attached to the 2018 number.
• Meridith Levinson — 5 insider secrets for beating applicant tracking systems, 4 March 2012. The earliest datable appearance of the 75% claim, attributed to Preptel and to no study.
• Preptel (archived). The vendor the 75% claim terminates at. Ceased trading August 2013 having published no research.
• John Sullivan — Why You Can’t Get A Job… Recruiting Explained By the Numbers, ERE, 20 May 2013. An opinion column, later relabelled as a survey. Origin of the “250 applicants” figure.
• Glassdoor — 50 HR and Recruiting Stats. The 2015 eBook edition welded the 250 figure to an unrelated interview ratio; Glassdoor dropped it from its own 2017 and 2018 editions.
• QuickCV — I Tested 8 ATS Systems, 2026. The only page carrying the “7 out of 8” claim we repeated. Organisational byline, unnamed systems, one fictional resume, no citations.
• Christine Assaf — Your job application was rejected by a human, not a computer, 2020. The existing debunk of the 75% claim. We relied on it in our first pass, then found three of its details wrong — which is in the limitations above.

The corpus

The 27 pages whose claims were extracted are not listed individually here. They are named per claim in the downloadable dataset, which is the form in which that list is actually useful.