The 75% ATS rejection statistic is not real

The most repeated statistic in career advice says that 75% of resumes are rejected by an applicant tracking system before a human ever sees them. There is no study behind it. The interesting part is not that someone made it up. It is that a real number, said about something real, turned into a claim about something else, and nobody noticed for fifteen years.

The number has consequences. People pay for keyword tools, rewrite their resumes around phrases, and conclude that a machine decided. None of that follows from the evidence. What follows is what is documented, and where the documentation stops.

Where the number came from, and what it was actually about

On 19 October 2011 the president of Preptel, a company selling resume optimisation, wrote that resumes "have to match at least 75% of the keywords in the job description" to reach a human at all, and that a 90% match is typically needed to be interviewed. That is the ancestor of the statistic, and it is preserved in the Internet Archive capture of that post.

Read the two claims against each other:

Three things changed, and none of them was a study. The unit went from keywords to resumes. A recommendation became a measurement. And a number attached to one vendor's sales pitch became a general fact about applicant tracking systems.

The date is worth correcting too, because it is the detail most often repeated: the widely quoted version dates the claim to 2012, and Preptel's 2012 page carries an 80% ATS-adoption figure (archived). The 75% belongs to October 2011. Preptel no longer exists, which is why this has to be read in an archive rather than cited live.

And then there is the repetition, which is the part that made the mutation permanent. Thousands of articles carrying the same figure is not corroboration. It is one transformed sentence, copied, and each copy made the next one look more confirmed. The argument that a claim must be true because a great many people say it is the exact error this post is about, which is why nothing above is stated without a link.

One limit, stated plainly, and it is the reason this post needed a second pass. The trace above was made from the Internet Archive, reading Preptel's own words rather than a summary of them, and it changed the story: an earlier version of this post said the figure had been invented, dated it to 2012, and cited a vendor's provenance write-up as its authority. The number was not invented. It was misattributed, which is a harder error to spot and a more common one, because the figure is real and a reader checking it finds a real number. Workday's marketing page meanwhile asserts that more than 98% of Fortune 500 companies use an applicant tracking system, citing a blog post from another vendor that sells resume tools. That figure is unsourced too. The systems really are everywhere. The rejection rate is the part nobody can produce.

What an applicant tracking system actually does

In the vendor's own words. Workday describes an applicant tracking system as software that "acts as a central database for candidate information", and describes five operations:

• Parse. The "parsing engine", which "reads and interprets document text", extracts education, skills and work history and organises them into structured profiles. Workday says this uses natural language processing that recognises patterns, so a good system reads "managed team of five" and "led five-person team" as similar.
• Store. Candidates sit in a database until the requisition closes, with workflow steps and status tracking attached.
• Search. Recruiters search profiles by hand using "filters such as years of experience or required certifications".
• Filter. Hiring managers use the system to "filter out candidates based on specific requirements".
• Rank. "Many platforms rank candidates by match percentage based on the job description, though they don't just count matching words."

Two things in that list are worth stopping on. The ranking is explicitly described as something other than a word count. And neither Workday page contains a rejection rate: no threshold, no percentage, no published default. The vendor that sells these systems does not publish one, and PassATS states that no applicant tracking vendor has.

Workday's own implementation guidance is more revealing than its marketing page. It tells buyers to "put guardrails around automated screening", to decide "which decisions stay in human hands", and to ask "how often will you sample rejected applications to catch false negatives". A vendor selling a system that quietly discarded three quarters of all applicants would not be telling customers to audit it for false negatives.

The parsing caveat is the part that actually bites

Workday's administrator guide on resume parsing, last updated 2023-06-23, is blunt: "Resume parsing results can vary based on resume format and order of words. For best results, use resumes that don't have images or image-based styles."

The same page notes that Workday does not auto-fill the Languages and Skills fields from a resume, and that employers can hide resume parsing on job application templates for external candidates.

So the durable, checkable fact here is a formatting fact. A resume whose header is a graphic, or whose body is an image, can produce an unreadable profile. That is not a rejection rate. It is a file the parser cannot read, and it is fixed by exporting a plain document with ordinary section headings. Workday tells employers the same thing from the other side: publish guidance for applicants on simple section headings, limited graphics, standard fonts and industry-standard job titles.

The seven-second scan, and what it does not say

The Ladders eye-tracking study from 2018 reports an average initial screen of 7.4 seconds, up from 6 seconds in the previous study, which was run during the depths of a recession.

Read the methodology before the headline. It ran in two stages. In the first, recruiters who "were not aware they were being timed" worked through a stack of resumes looking for candidates to progress. In the second, eye-tracking followed recruiters "under lab conditions". So 7.4 seconds is a duration for a first pass. It is not a measure of accuracy and it is not a decision. Nobody finishes a judgement of a person in seven seconds, and the study does not claim they do. It is also eight years old, produced by a job board with a commercial interest in recruiting services rather than by a regulator.

What the study did measure is where the eyes went, and that is the useful part. Recruiters spent more time on job titles than on any other element. The best-performing resumes shared bold job titles supported by short bulleted accomplishments, clear fonts, and enough structure to produce a consistent scanning pattern down the page. The worst performers shared a cluttered layout, multiple columns, very little white space, and visible evidence of keyword stuffing.

The finding about page two is the most actionable. An engaged recruiter spends as much time on a second page as on the first, but time on the second page is strongly predicted by how compelling the first page is.

The useful reading is not "you have seven seconds". It is that the first screen is where the reading happens, and legibility is what buys the rest of it.

The tailoring evidence, at its real size

The number usually offered in this category is a multiplier: tailor your resume and get two or three times the interviews. No source I could find supports those multipliers, and none supports an exact keyword-match threshold. They come from vendor pages that document no control.

What has been measured is narrower and still worth having. A field experiment by ResumeGo submitted 7,287 fictitious applications to live postings on ZipRecruiter, Glassdoor and Indeed between 15 July 2019 and 10 January 2020. Callbacks within 30 days: 10.7% with no cover letter, 12.5% with a generic one, 16.4% with a tailored one, a 53% relative lift over sending nothing. Three qualifications matter here. These were cover letters and not resumes. The applicants were fictitious. The study was run by a company that sells resume writing.

On resumes specifically, a 2025 field experiment using 4,035 fictitious resumes found that listing advanced data skills produced a 2.5 percentage point gain in callbacks, 19.2% in relative terms. That is a real, measured effect on resumes. It is also a rounding error next to a 2x headline.

For scale, here is a Canadian effect on callbacks that is larger than the one resume-level result above, and it has nothing to do with formatting. A field experiment by Philip Oreopoulos, published in 2011, sent 12,910 randomly varied resumes to 3,225 postings in the Greater Toronto Area between April 2008 and September 2009, with a second wave in Toronto and Montreal. Resumes with an English-sounding name, Canadian education and Canadian experience drew callbacks 15.7% of the time. Change only the name to Chinese, Indian or Pakistani and keep the Canadian education and experience, and the rate falls to 11.3%. Adding language fluency, multinational firm experience, education from a highly selective school or active extracurricular activities did not close that gap.

The point is not the discrimination finding itself. The point is the magnitude: 4.4 percentage points, which is nearly twice the effect of adding advanced data skills to a resume. The levers that move callbacks in this market are not the ones the resume-tool industry sells.

What to do instead

• Make the file parseable. Plain section headings, "Experience" and "Education" as words rather than graphics, standard fonts, one column. This is the vendor's own advice and it costs nothing.
• Write the first screen for a reader in a hurry. Job title, employer, dates, and the two or three facts that make you obviously the person. That is where the eye-tracking found the attention going.
• Tailor to the posting's stated requirements. Eligibility and retrieval are the only mechanisms here with any documented effect. A posting is a list of the questions you will be asked, which is why the posting is the most useful preparation document you have.
• Ignore the match score. The company whose provenance work this post relies on says outright that applicant tracking checkers, including its own, have no access to employer systems, and that a score from one is a heuristic rather than a simulation of anything.

The fear is expensive and it produces worse documents. The two documented problems are a file a parser cannot read, and a person in a hurry. Both are cheap to fix.