How an ATS Works: Parse, Filter, Rank, and What Recruiters See

An ATS works by turning your resume into a database record and then letting a recruiter query that record. The five stages are parse (extract the text), structure (map the text to fields), search and filter (the recruiter narrows the pile), rank (a match layer sorts what is left) and the recruiter’s screen (a human reads the top). Every mainstream system, from Workday at an MNC to Naukri RMS at a 40-person company, runs some version of these five. Understanding how ATS works matters for one reason: each stage has its own way of losing you, and none of them is a rejection.

This page walks through the stages in order, says what the vendors document about each one and what is inferred, and maps every stage to the checks in our rubric on the How we score page. The definition of the term itself, and the list of systems Indian employers use, are on the parent hub, What Is ATS? Full Form, Meaning and How Applicant Tracking Systems Read Resumes.

Stage 1: the parser extracts text from your file

The parser’s first job is to pull the characters out of your PDF or .docx in reading order, and nothing else in the system can see what the parser could not extract. This is the stage where formatting matters, and it is the only stage where it matters.

What vendors document. Greenhouse publishes the clearest list of what breaks its parser: files over 2.5 MB, “spaces between the letters”, “graphics, photos, or word art”, a file uploaded as an image rather than a document, “complex resumes with tables, headers, and footers”, name and contact information “in the header, footer, or text box”, a columned layout, and resumes without clear sections. When the parse fails, Greenhouse attaches the file to the candidate without filling any fields, and a recruiter types the details by hand or does not.¹ Textkernel, the parsing engine under iCIMS and SAP SuccessFactors, explains why columns fail: a plain left-to-right, top-down read “mixes bits and pieces of each column together”, and its machine-learning fix for column separators moved well-rendered CVs from 62% to 90%.² Lever’s parser, per Jobscan’s guide (Lever’s own help centre is login-walled), “can’t parse images” and can read columns and tables but “the format can sometimes be affected”.³ Workday’s administrator guide says “resume parsing results can vary based on resume format and order of words” and recommends resumes “that don’t have images or image-based styles”.⁴

What is inferred. No vendor publishes its reading algorithm. That .docx headers are skipped by many extractors and that icon fonts come out as private-use characters are observations of extractor behaviour, not vendor statements.

Rubric checks at this stage: P1 (text layer), P6 (glyphs), P7 (letter spacing), P9 (single column), P10 (tables), P11 (text boxes), P12 (contact in the body), P13 (hidden text), P14 (graphics). Together they are most of the 35 parse-fidelity points.

The fastest way to see this stage happen to your own file is to test your resume against an ATS and open the What the ATS read tab, which shows the extracted text in the order a parser produced it.

Stage 2: the extracted text becomes structured fields

Once the text is out, the parser segments it into sections and maps each section to fields: name, email, phone, location, each job’s title, organisation and dates, each degree, and a skills list. These fields, not your PDF, are what the rest of the system works on.

What vendors document. Textkernel’s candidate data model lists what its parser returns: name, birth date, nationality and gender (normalised to the ISO 5218 standard), phone and email, personal URLs such as LinkedIn and GitHub, and for each job a title as stated, an organisation, start and end dates in YYYY-MM-DD form and computed years and months of experience, with job titles normalised to standard profession codes as an add-on. A profile picture is returned as an embedded image, which is to say the photo is stored, not read.⁵ Greenhouse’s Talent Matching documentation says it extracts “skills, years of experience, job titles, start/end dates of employment, company names from employment history”.⁶ Lever, per the same third-party guide, extracts name, work history, job title, address, email, phone, LinkedIn profile, organisation, education and skills.³ Workday’s administrator course manual says parsing on upload fills personal and contact information, work experience and education, and the administrator guide adds that Workday does not auto-fill the Skills field from the resume.⁴

The segmenting step depends on headings. RChilli’s Resume Quality output, which is a parser reporting on its own confidence, treats “Work experience section was not found” and “Education section was not found” as fatal problems, “Experience or education sections with no header have been found” as a major issue, and a position with no job title or employer as a data issue.⁷ Textkernel’s equivalent output carries the advice as a rule: “Every section should have a clear, unambiguous, commonly-used header on a separate line directly above the content.”⁸

What is inferred. Which heading words each parser recognises is not published. “Experience”, “Education”, “Skills”, “Projects” and “Summary” are safe because every parser we can read documentation for names those sections; “My Journey” is not.

Rubric checks at this stage: P17 (email and phone found), P18 (standard section headings), P19 (every role has a title, an organisation and dates), L4 (one date format). Textkernel computes years of experience from dates, so a role with no dates contributes zero tenure, and stage 3 filters on that number.

Stage 3: the recruiter searches and filters the fields

A recruiter with 400 applicants does not open 400 PDFs; they type a search and set filters, and both run on the fields from stage 2. This is where most resumes are lost, and it is invisible: nobody tells you a search did not return you.

What vendors document. Naukri’s Resdex support pages say the keyword box searches the profile, summary of experience, current organisation, designation and educational institute name, and that Advanced Search combines keywords with experience, annual salary, current location, functional area, industry and included or excluded employers.⁹ Naukri’s EZ Keywords feature splits the search into Any, All and Exclude boxes because complex Boolean queries were something only expert users could build in one box.¹⁰ Naukri also documents a notice-period search.¹¹ Lever’s search stems words, so “collaborating” finds “collaborated”, but “cannot identify abbreviations”: a search for “Search Engine Optimization” does not return a resume that says only “SEO”.³

What is inferred. Whether an ATS searches only structured fields or also the full text is rarely stated. Treat both as possible: put each term in the skills section and in a bullet.

Rubric checks at this stage: F1 (skills as atomic, canonical terms), F2 (abbreviations spelled out once), F4 (recognisable job title), F5 (total experience computable), F6 (location stated), F10 (CTC and notice period stated). The India-specific pair, F10, exists because Resdex filters on both and an empty field is an absent candidate.

Stage 4: a match layer ranks what is left

Newer systems sort the filtered list by a match grade, and every vendor that documents such a layer says it does not reject. The grade changes reading order, not outcome.

What vendors document. Greenhouse Talent Matching asks the hiring team to set a calibration first: the skills, job titles and years or industry of experience for the role, with four to six key skills each rated for importance. It then places each candidate in one of five categories: Strong match, Good match, Partial match, Limited match, or Needs manual review (used when resume processing fails, AI is switched off or the candidate opts out). Candidates appear in the Application Review table grouped by category and ordered by application date inside each group.¹² Greenhouse’s FAQ adds two details worth knowing: it “compares related terms on each resume to the calibrated skills you set, and multiple terms can map to the same calibrated skill, so a longer list of matched terms doesn’t always mean a higher match score”, and “Talent Matching does not auto-reject or auto-advance any candidate”.⁶ Workday’s HiredScore is described in Workday’s feature reference as “a means of prioritizing candidates based on comparison of job requirements and candidates resumes”, shown to recruiters on a Smart Candidate Profile alongside a parsed resume in a consistent format.¹³ Workday also documents that after you apply it “suggests skills” from your application and resume, which you can keep or remove, and that a configured Candidate Skills Cloud can score how your skills align with the requisition.⁴

What is inferred. Neither vendor publishes the weighting; both document the input, which is extracted skills, titles, dates and companies. A skill buried in prose, or a title that does not standardise, gives the model less to work with.

Rubric checks at this stage: F3 (listed skills evidenced in a bullet), F9 (core skills for the role named), I2 (bullets carry a number). None is a parse issue; they are what a match model and a recruiter both read.

Stage 5: the recruiter’s screen

What a human finally reads is the parsed profile beside the list, then the PDF for a few seconds, and the two must agree. This is the stage the impact and language checks are written for.

What vendors document. Greenhouse’s Application Review stage shows the reviewer a candidate panel (applied date, source, personal URLs), the documents, answers to the job’s custom questions and application history, with three actions: Advance, Reject with a reason and an email, or Leave Feedback on a scorecard.¹⁴ Greenhouse’s structured-hiring material describes scorecards with defined attributes so that every candidate is judged on the same criteria.¹⁵ Workday’s HiredScore profile shows a parsed CV “in the same format” for every candidate beside the job description and the original attachment.¹³

What is inferred. How long a recruiter spends on the PDF is survey evidence, not vendor documentation: the 2025 survey of 25 recruiters behind our impact checks put short, active bullets (72%) and measurable achievements (52%) at the top of what they skim for.¹⁶

Rubric checks at this stage: I1 to I8 (verbs, numbers, bullet length, openers, filler, summary, bullets per role, length) and the language checks L1 to L5.

How ATS works on one table: the five stages and the rubric

Each stage loses a different kind of resume, and the rubric’s four groups follow the stages.

Does an ATS reject your resume? What 25 recruiters said

No. In a 2025 survey of 25 recruiters, 92% said their system never auto-rejects on resume content or formatting, and the only automatic rejection any of them use is a knockout question on the application form. The figure usually quoted against this, that three-quarters of resumes are rejected before a human sees them, has no study behind it. What removes a resume is stage 1 and stage 3 above: a parse that fills the wrong fields, and a search that never returns you.

The survey

In September and October 2025 a resume company asked 25 recruiters working across eleven systems, Workday, iCIMS, Lever, Greenhouse, SuccessFactors and others, what their ATS actually does with a resume.¹⁶ The answers were consistent. 92% said the system never auto-rejects on resume content or formatting. 100% use knockout questions on the form: work authorisation, location, a required licence or degree. 44% have a match score of some kind available to them, and 8% use it to reject. Twenty-five recruiters is a small sample and the survey was run by a resume company, so read the percentages as a strong indication rather than a census. It is still the only recent survey that puts the question to recruiters directly, and every vendor document on this page points the same way.

Where the famous rejection figure came from

The sentence “75% of resumes are rejected by ATS before a human sees them” traces to a sales pitch by a vendor called Preptel around 2012. Preptel closed in 2013. No study, no dataset and no method were ever published. Because there was never an original to pin it to, the number drifted to 70% and 88% in later articles, and it still appears on checker landing pages in 2026. Two career writers, at Uncharted Career and Hiring Thing, traced the claim back to that pitch and found nothing beneath it.¹⁷

A second source gets folded into the same myth. The 2021 Harvard Business School and Accenture report “Hidden Workers” is often cited as proof that systems reject the qualified. What it actually reports is an opinion survey: 88% of the executives surveyed agreed that qualified candidates are vetted out because they do not match the criteria specified in the hiring system.¹⁸ That describes filters and knockout criteria that employers set, which is real, not a parser rejecting formatting.

What actually filters you: the parse and the search

The file is parsed into fields: name, contact, titles, companies, dates, skills, education. Recruiters then search and filter the pile. On Naukri Resdex that is a Boolean string plus filters on experience, CTC, notice period and location.⁹ On LinkedIn Recruiter it is the standardised title first. The newer AI layers, Workday’s HiredScore grades and Greenhouse’s Strong, Good, Partial and Limited categories, sort the list; their own documentation says they do not reject, and SmartRecruiters and Oracle Taleo document the same division between a prescreening question on the form and a ranking of what is left.^{6 19}

So there are two ways to lose, and neither is a rejection. Your resume parses into the wrong fields (a title in the wrong place, a role with no dates, a phone number in a header the extractor skipped), or it does not contain the words the recruiter searches for. Both are fixable in an evening, and both are silent: nobody sends an email saying the search did not return you. That distinction is the whole reason our rubric scores parse fidelity and findability separately from readability.

Bias, the myth’s honest cousin

Textkernel’s data model shows a parser stores gender and nationality if the resume states them, and the photo as an image.⁵ The system does not judge those fields; a human might. That is why the rubric flags photo, date of birth, father’s name and marital status for private-sector applications without counting them.

How ATS keyword matching works: exact, stemmed and weighted

Keyword matching in an ATS is three different things depending on the system: a literal search, a stemmed search, or a weighted comparison against calibrated skills. Knowing which one you face decides how you write the skills section.

• Exact. Naukri Resdex Boolean is literal and does no synonym expansion: “Amazon Web Services” and “AWS” are two different searches. Write both forms once (F2).
• Stemmed. Lever matches word forms (“collaborate”, “collaborated”) but not abbreviations.³ Stems help with verbs; they do nothing for “SEO” versus “Search Engine Optimization”.
• Weighted. Greenhouse Talent Matching maps multiple related terms to one calibrated skill and weights skills by the importance the recruiter set, so repeating a term does not raise a match.⁶ What raises it is the skill being present, in the skills list and in a dated role where it was used.

No candidate can see an “ATS keyword score” in any of these systems. The keyword tab on our checker compares your file to a pasted job description and names the terms present, implied and missing; it is a proxy for stages 3 and 4, not a number from inside any ATS. Choosing and placing terms is the resume keywords hub’s subject.

Why an ATS “rejects” a resume: the 12 reasons, ranked by cost

Only one item on this list is a rejection; the other eleven are ways a resume disappears from search or reads wrongly on screen. We will publish fail rates from our own engine once the volume is large enough to be honest; until then the order is by the points each check carries in the rubric, which is the vendors’ documented failures weighted by how much they break.

1. A knockout question answered against the requirement (work authorisation, location, licence, degree). The only automatic rejection, and it happens on the form, not in the resume.
2. No text layer: a scanned, photographed or image-exported PDF (P1, 6 points, fail). Nothing downstream can run.
3. Skills section missing or written as prose, so the search terms are not there as terms (F1, 5 points).
4. Core skills for the role absent from the file (F9, 5 points).
5. Two or more columns, interleaved on extraction (P9, 4 points).
6. Contact details in the header, footer or a text box, dropped by the parser (P12, 3 points, fail).
7. Non-standard section headings, so sections are not detected (P18, 3 points, fail).
8. A role with no dates, so tenure is zero and experience filters exclude it (P19, 3 points, fail; F5, 3 points).
9. No email or phone found (P17, 3 points, fail).
10. Tables around experience or education, merging or splitting entries (P10, 3 points).
11. Abbreviations without the long form (F2, 3 points).
12. No CTC, notice period or city, so Indian recruiter filters skip the profile (F10, 3 points; F6, 2 points).

Hidden white text (P13) is deliberately not on the list. It does not cause a lost search; it causes a recruiter who sees a wall of keywords in the parsed view to stop reading.

What to do with the five stages

Fix the stages in order, because a stage 1 failure makes every later fix pointless. Read the extracted text and correct structure before wording: one column, standard headings, contact details in the body, dates on every role. Then the findability terms, then the bullets. Re-check after each pass; the score is deterministic, so the delta is exactly what the change was worth. The numbered action list per stage is how to pass ATS screening, and the parent hub, What Is ATS? Full Form, Meaning and How Applicant Tracking Systems Read Resumes, has the definition and the list of systems.