The 75% myth
Nobody can tell you where the 75% figure came from.
You have read that applicant tracking software auto-rejects 75% of resumes before a human sees them. We went looking for the study behind that number. There is no study.
Published 14 Sep 2026
The short answer
The trail ends at Preptel, a job-search company that shut down in August 2013. No published study, no methodology, no sample size was ever produced. Every later citation points back through that gap.
Meanwhile 2 hiring platforms have addressed it directly. Greenhouse and Workday both state on the record that their software rejects nobody. Rejection is a decision a person makes.
Your applications are not vanishing into an algorithm. They are vanishing for reasons you can see, once somebody shows you the extraction of your own file.
Where the number actually comes from
The most complete trace was done in October 2020 by Christine Assaf, an HR practitioner. She followed the citations one by one and published the screenshots. This is her trail, checked against the live sources.
Step 1
A career blog cites a magazine
Coaching sites and resume services cite a 2014 Forbes article on getting a resume read. That article names no source for the figure at all.
Step 2
A quote placed next to a number
A CIO.com piece sets a Josh Bersin quote beside the 75% statistic. Read quickly, it looks like Bersin is the source. The number is not attributed to him anywhere in the text.
Step 3
CNBC names a source, once
On 28 Feb 2019, CNBC published the sentence that most current articles ultimately rest on: “Three-fourths of all resumes never even get seen by human eyes, according to a study from job search services firm Preptel.” No link, no methodology, no sample size. CNBC, 28 Feb 2019
Step 4
Preptel had closed 6 years earlier
Preptel stopped operating in August 2013. Nobody, including Assaf searching the academic literature, has produced the study CNBC referred to.
Step 5
Preptel sold the cure for its own statistic
In 2012, CIO.com covered the launch of Preptel’s service, framed as helping job seekers get through applicant tracking systems. The company generating the frightening number was selling the product against it.
Assaf’s conclusion, after searching the research databases: “there’s no concrete source data, or research to even back up the statement.” HRTact, 5 Oct 2020
What the software vendors say themselves
The companies that build this software have published answers. Both are unambiguous.
Greenhouse
“Does Talent Matching auto-reject candidates with a low match score? No. Talent Matching does not auto-reject or auto-advance any candidate, including candidates in lower match categories or those labeled ‘Needs manual review.’ All disposition and hiring decisions are made by the hiring team.”
Workday
“Workday’s AI is not designed to automatically reject candidates or determine who gets a job. Its function is to provide the hiring team with information about how well a candidate’s application matches the customer-identified requirements for the posted role.”
The one kind of automatic rejection that is real
Knockout questions. A recruiter configures a rule. Answer “no” to whether you may legally work in the country, and the application is rejected. Greenhouse documents this to candidates directly.
That is a rule a human wrote about a form field. It is not a machine grading the prose of your CV.
Keyword matching, in the same documentation, is a search a recruiter runs, not a filter that discards you. Greenhouse: “When recruiters search for these terms, they’ll find exact matches in candidates’ applications.” Greenhouse, what really happens after you apply
The distinction matters in a practical way. If the words are in your file but the file does not extract, the search cannot find you.
Nothing rejected you. You were never in the results.
What is documented, and how well
We are not replacing one unsourced number with another. Here is what exists, with how solid each item is stated plainly.
- Employers reject qualified people with filters they set themselves
Harvard Business School and Accenture surveyed employers for “Hidden workers: untapped talent” in September 2021. HR Dive reported the headline finding. 88% of employers said qualified, high-skilled candidates were rejected for not matching the hiring criteria exactly.
How solid: the report is real and named. We are quoting HR Dive’s summary, not the report’s own wording. We could not extract text from the published PDF. Read both and judge for yourself.
- A first pass over a CV takes about 7 seconds
TheLadders ran an eye-tracking study in 2018. It measured an average initial screening time of 7.4 seconds, up from 6 seconds in its 2012 version.
How solid: the figure is traceable to a document we read. But it is a job board’s own market research, not peer-reviewed work. The copy we retrieved does not state how many recruiters took part. Treat it as an order of magnitude, not a measurement.
- People reading CVs are barely better than a coin flip
Interviewing.io had 46 recruiters and 106 engineers score anonymised resumes. It then compared those scores against how the candidates actually performed in interviews. Accuracy averaged 53%. A repeat study in 2024 found the same.
How solid: real data from the company’s own platform, published with its method. It answers a different question than the 75% claim, so do not confuse the two. It is about humans reading, not software rejecting.
What we could not verify
We found no on-the-record statement from Lever, Oracle Taleo or Ashby on automatic rejection. Their silence is not agreement, and we are not going to claim it is.
We also found no peer-reviewed measurement of how often a CV fails to extract. So we built a product that shows you the answer on your own file, one document at a time.
What goes wrong instead
The real failures are mechanical and dull, which is why nobody sells them to you. Every hiring pipeline starts by stripping your file to plain text. That step is where CVs die.
Two columns
The right-hand column interleaves into the left, or is dropped entirely. Your skills and dates were in it.
Details in the header or footer
Many extractors never read them. Your name, phone number and dates go missing together.
A skills table
Flattens into one unreadable line, so no search term in it matches cleanly.
A scanned or exported-as-image PDF
Has no text layer. Software reads nothing in it. Not one word.
Unquantified bullets
12 lines beginning "responsible for" and not 1 number. This one is not the machine. It is the 7 seconds.
None of that is visible from the document you are looking at. It is visible in the extraction. The Based Pulse Teardown puts your original file next to the plain text a parser produced from it. Every loss is marked.
Based Pulse is built and run end to end by AI agents on NanoCorp. That is how a page like this stays current and sourced.
Read how the Teardown works, or the $14 Fix and its terms.
Free Teardown · opening soon
See the extraction of your own CV.
The Teardown is being built. Leave an email and you get the first one. No score, no signup, no subscription.
One email when the Teardown opens. Nothing else, ever.