# The ATS Pipeline Decoder, From Upload to a Human Opening Your File

*You will be able to trace your resume through each ATS stage, name what happens at each one, and fix the failures that actually lose you a read.*

- Canonical URL: https://www.refolk.ai/candidates/guides/ats-pipeline-decoder
- Pillar: Positioning and materials
- Format: Reference
- Published: 2026-08-18
- Last reviewed: 2026-08-18
- Reading time: 15 min
- Keywords: applicant tracking system, ats resume, resume parsing, ats parse rate, ats knockout questions

## Key takeaways

- An applicant tracking system stores your original file first, then parses it into structured fields; even a bad parse leaves your resume viewable by recruiters, because every ATS in a 2025 practitioner review let recruiters open the original.
- The widely repeated claim that 75% of resumes are never seen traces to a defunct company called Preptel with no disclosed methodology, and in a 2025 study of 25 US recruiters only 8% configured any content-based auto-rejection.
- Parser accuracy averages near 95% only because name and email hit 0.99; skill extraction runs at 0.75 to 0.85, and skills are what your ranking depends on.
- File type still matters: DOCX averaged a 96.7% parse rate versus 91.3% for text-based PDF and 4.3% for image-based PDF, so a Canva image export scores near zero.
- There is no universal passing score; most systems rank you relative to other applicants for that specific role, so the real hurdle is volume and human-set criteria, not the software.
- In Refolk's index there are 112,728 US recruiters and talent-acquisition professionals versus 7,295 in the UK, a 15.5x gap that shapes how much human triage sits behind any single ATS.

Every job seeker has heard that a robot reads their resume first and throws most of them away. Some of that is true, most of the scary version is not, and the difference decides what you should actually fix. This guide traces your resume through each stage an applicant tracking system puts it through, from upload to a human opening the file, names what each stage does, and tells you what to check when a stage silently fails you. It is a lookup document: jump to the stage or the failure you care about and leave.

An applicant tracking system, or ATS, is the software an employer uses to receive, store, parse, filter, rank, and track candidates for a job. This guide is for anyone writing or rewriting a resume who wants to stop guessing about the black box between "submit" and "we reviewed your application."

## What actually happens between upload and a human reading it

On upload the ATS stores your original file and then parses it into structured fields: name, contact, work history, education, skills, and certifications, saved as a candidate profile rather than as the document. The document does not vanish. In a practitioner review of live systems, every ATS still let recruiters open and read the original file, so even an imperfect parse leaves your resume viewable.

The rest of the pipeline is a sequence, not one gate. One documented flow runs like this: the file is stored, knockout questions remove ineligible applicants, recruiters search keywords for core skills, applications are ranked by relevance and timing, the strongest advance while others may never be opened, and then humans assess, interview, or reject.

#### The ATS pipeline

1. **Upload** - ATS stores your original file against the job
2. **Parse** - Engine extracts text into structured fields
3. **Knockouts** - Yes/no eligibility questions filter the pool
4. **Rank** - Keyword search or scoring orders the shortlist
5. **Triage** - Recruiter reviews the ranked list top-down
6. **Decision** - Human moves you to interview or rejection

*Your file is stored before it is parsed, and a human decision sits at the end, not the start.*

The single most useful correction to the popular story: the machine mostly sorts and ranks, and a person mostly decides. Only 8% of recruiters in one 2025 study configured the software to reject on content automatically. The other 92% rely on human review guided by knockouts and optional scores.

**8% - Recruiters who configure content-based auto-rejection**

From a 2025 study of 25 US recruiters; the other 92% rely on human review guided by knockouts and scores.

## The pipeline, stage by stage

Here is the same pipeline as an ordered procedure, with what each stage does and what "done" looks like. Note on ordering: some guidance places knockouts before parsing conceptually, but most sources show file storage first, then parsing, then knockouts. Work from that order.

#### What the ATS does to your file, in order

1. **Upload and file ingestion** - You submit the resume and the ATS accepts and stores the original file against a requisition. This is instant, and the stored original stays viewable to recruiters regardless of parse quality. Done when the file is attached to the job.
2. **Parsing and text extraction** - The ATS or an embedded engine extracts text and maps it into structured fields: name, contact, work history, education, skills. This takes sub-second to a few seconds per resume. Done when a structured candidate profile exists alongside the stored file.
3. **Knockout questions** - You answer compliance and eligibility questions that can auto-filter ineligible applicants, such as work authorization or a minimum-years threshold. This is immediate. Done when ineligible applicants are flagged out of the active pool.
4. **Keyword search and ranking** - A recruiter searches the parsed database like Ctrl+F, or the system scores your profile against the job description, producing an ordered list. This happens minutes to hours after you apply. Done when a ranked shortlist exists for the role.
5. **Human triage and shortlisting** - A recruiter reviews the ranked applications and advances the strongest; some lower-ranked files may never be opened. This depends on volume. Done when a shortlist is passed to the hiring manager.
6. **Review and decision** - A hiring manager or recruiter reviews shortlisted candidates and moves them to interview or rejection, mostly by hand. This takes days to weeks. Done when your file has a decision recorded, not just a rank.

Parsing is fast. One parser reports a median parse time of 50 milliseconds per resume. Speed is not the problem; fidelity is, and I cover that below.

## How parsing works, and where it drops the field that matters

Parsing turns your uploaded document into structured data, and it is uneven: identity fields parse almost perfectly while the fields your ranking depends on parse much worse. That gap is the quiet failure most format advice ignores.

Whether you send PDF, DOCX, or plain text, the parser extracts text and maps it to fields. But field-level accuracy is not one number. Name and email hit 0.99 or higher. Skill extraction in production runs at 0.75 to 0.85. A parser can be "95% accurate" in aggregate and still miss one in four of your skills, because the near-perfect identity fields pull the average up.

```table

```

| Dimension | High | Low |
|---|---|---|
| File type parse rate | DOCX 96.7% | Image PDF 4.3% |
| Parser generation accuracy | LLM 95-97% | Rule-based ~65% |
| Field-level accuracy | Name/email 0.99 | Skills 0.75-0.85 |

Read the table as three separate risks. File type is the crudest: DOCX averaged a 96.7% parse rate versus 91.3% for text-based PDF and 4.3% for image-based PDF. Parser generation is the vendor's problem, not yours, but it explains why the same resume behaves differently across employers: an LLM-era parser reads 95 to 97% of a clean single-column document while a legacy rule-based tool reads closer to 65%. Field-level accuracy is the one to internalize: even a good parser reading a good file will drop skills at a higher rate than anything else.

> **Watch out:** "95% accurate" is a field-weighted average
>
> The 95% aggregate hides the ~0.75 accuracy on skills. Do not assume the terms you rely on for ranking parsed as reliably as your name did.

Field-level accuracy on messy real-world resumes measured near 87%, against roughly 96% for careful human data entry. So the parser is worse than a person typing your history in by hand, and the deficit lands hardest on skills. The practical response is not exotic: single-column layout, standard section headings, and skills written in plain running text where a parser expects them.

## Do ATS systems auto-reject you? The 75% myth, priced honestly

No, applicant tracking systems do not automatically bin 75% of resumes, and that specific figure has no established source. The stat traces to Preptel, a defunct recruiting-service company, with no disclosed methodology. Treat any repetition of it as folklore.

What the software actually does automatically is narrow. In a 2025 study of 25 US recruiters, 100% used knockout questions for compliance, but only 8% (2 of 25) configured content-based auto-rejection, and those two used strict thresholds like "match less than 75%" or "fewer than 7 of 10 required skills." One tool's real pipeline data reported a median ATS score of 48 out of 100 with 52% of keywords missing, and candidates still advanced, because there is no universal passing score. Most systems rank you relative to the other applicants for that specific role.

> The machine mostly sorts and ranks. A person mostly decides. The bottleneck is the criteria a recruiter set, not the software that executes them.

The real bottleneck is human-set criteria, executed literally by software. In Harvard Business School and Accenture's Hidden Workers report, which surveyed more than 8,000 workers and 2,250 executives across the US, UK, and Germany, 90%-plus of employers use a recruiting system to make a first cut, and 88% believed their own system weeded out high-skilled prospects. The exclusion is not clever AI overreaching; it is a rigid filter, such as a degree requirement or a hard gap rule, that a recruiter configured and the software applied without judgment. The report estimates about 27 million US workers are "hidden," effectively screened out by these processes.

> **Rule:** Silence is usually volume, not a verdict
>
> No response rarely means an algorithm rejected you. Before you rebuild your resume for "the ATS," check your knockout answers and the number of applicants for the role. Volume and eligibility explain far more silence than parsing does.

## "Most popular ATS" depends entirely on the sample

There is no single most-common applicant tracking system, because three credible methods produce three different leaders. Which one is "most popular" depends on whose hiring you measure, so match the claim to the sample before you act on it.

| Sample / method | Leader and share | Runner-up |
|---|---|---|
| Fortune 500 installs (Jobscan) | Workday 39%+ | SuccessFactors 13.2% |
| Global revenue (Apps Run The World) | iCIMS 10.7% | Oracle / Workday next |
| 3,222 top-rated employers crawl (ResumeGeni) | Greenhouse 49.0% | Workday 21.8% |

Among Fortune 500 companies, 97.8% run a detectable ATS, and Workday dominates that population. By global revenue across the whole market, iCIMS leads. A broad crawl of top-rated employers puts Greenhouse first at 49.0%. These are not contradictions; they are different populations. If you are applying to enterprise employers, expect Workday and SuccessFactors. If you are applying to venture-backed and mid-market companies, expect Greenhouse and Lever. The named systems you will meet most are Workday, Greenhouse, Lever, iCIMS, Taleo/Oracle, SAP SuccessFactors, and BambooHR, and some embed third-party parsing engines such as Textkernel or RChilli under the hood.

Do not over-optimize for a specific vendor's quirks. The behaviors that survive across all of them - single column, text-based file, contact details in the body, skills in plain text - matter more than any one system's rumored preferences.

## The scale of human triage behind the machine

Behind every ranked list is a person with a finite workday, and there are far more of those people in some markets than others. That ratio is a useful reality check on how much human judgment, versus pure automation, sits between your file and a decision.

| Market | Recruiters / TA professionals | Share of pair |
|---|---|---|
| United States | 112,728 | 93.9% |
| United Kingdom | 7,295 | 6.1% |
| US : UK ratio | 15.5x | - |

In Refolk's index of professional profiles, a search for recruiters, technical recruiters, and talent-acquisition professionals returns 112,728 in the US against 7,295 in the UK, a 15.5x gap. The counts are Refolk's; the percentages and ratio are derived from them. The practical read: in the US market there is a deep bench of humans doing triage, which is consistent with the finding that most rejection is human-mediated, not automated. It also tells you the person on the other side is often processing a high pile, which is why ranking and clean parsing help you surface, even though they do not decide for you.

**15.5x - US recruiters per UK recruiter in Refolk's index**

112,728 US versus 7,295 UK across the Recruiter, Technical Recruiter, and Talent Acquisition title set.

If you want to know exactly who screens for the systems you are applying through, you can search for the people rather than guess at the software. [Refolk](/candidates) writes your resume from your own history and tailors it to each posting, and its index lets you name the recruiters behind a given stack.

Ask me this: `Talent acquisition leaders at US companies that run Greenhouse for hiring.` - [run the search](https://www.refolk.ai/start?q=Talent%20acquisition%20leaders%20at%20US%20companies%20that%20run%20Greenhouse%20for%20hiring.).

*Returns named TA leaders at Greenhouse-using employers, so you can see who is on the human side of the pipeline you are applying into.*

## How this goes wrong: failure modes and false positives

Most resume-versus-ATS damage comes from a short list of specific, checkable mistakes. Here is each one, what it looks like, and the test that confirms it.

### The failures worth checking first

- **"The ATS auto-rejected me."** The false positive is reading silence as an algorithmic bin. Check your knockout answers and the applicant volume, not your formatting. Only 8% of systems enable content auto-rejection, so the true hurdle is usually volume.
- **Contact details in the header or footer.** Headers and footers may be ignored by parsers, so your name, phone, and email can silently vanish. Keep all contact details in the main body of the document.
- **Image-based or Canva PDF.** It looks perfect and extracts almost nothing; image PDFs scored 4.3% parse rate. Open the file in a browser and try to highlight the text. If the cursor selects rectangular regions or nothing, it is image-based and your score will be near zero.
- **Trusting a flat "95% accurate" claim.** The aggregate masks the ~0.75 accuracy on skills. Assume your skills parsed worst, and write them in plain text under a clearly labeled section.
- **Multi-column or table layouts.** These scramble the reading order in either PDF or DOCX. Stick to a single column so the parser reads top to bottom.
- **Keyword stuffing.** Padding can backfire; more keywords correlated slightly worse with outcomes, around -0.09. Write to the role, not the filter, and only include terms you can defend in an interview.
- **Ignoring a stated format instruction.** If the employer asks for PDF, send PDF. Ignoring an explicit instruction is a worse signal to a human than any parsing risk.
- **Assuming a universal passing score.** There is none. Most systems rank you relative to the other applicants for that specific role, so the same resume can rank differently for two jobs.

#### Where to spend your fix effort

Horizontal axis runs from Low effort to fix to High effort to fix. Vertical axis runs from Low impact on getting read to High impact on getting read.

| Quadrant | What it means |
| --- | --- |
| Keyword tone (low/low) | Nudge language toward the posting, do not obsess |
| Full portfolio rebuild (low impact/high effort) | Skip until basics are done |
| File type and header placement (high impact/low effort) | Fix first: single-column, contact in body, text-based file |
| Reframing weak experience (high impact/high effort) | Worth it, but after the cheap wins |

*Prioritize failures that are both high-impact and cheap to fix, such as moving contact details out of the header.*

There is one failure mode you cannot fix by formatting, and you should know it exists. Brookings research found that in widely used screening models, male-associated names were selected 51.9% of the time versus 11.1% for female-associated names. That is a bias in the tools, not in your document, and it is why regulation is arriving vendor-first: the EU AI Act's hiring provisions take effect August 2026, and the Mobley v. Workday class action targets the software vendor rather than only the employers. You cannot format your way past this. You can name it, and you can prefer employers and channels that involve a human earlier.

## A copy-and-check resume structure that survives parsing

The structure below is the one that behaves consistently across parsers. It is deliberately plain, because plain is what parses. Adapt the content, keep the shape.

**ATS-safe resume skeleton**

```
FULL NAME
City, State  |  phone  |  email  |  one profile link
(all of the above in the document BODY, never in a header or footer)

SUMMARY
Two lines naming your role and the results you produce, in plain text.

EXPERIENCE
Job Title, Company, City  |  Start - End
- Achievement with a number and the skill it used
- Achievement with a number and the skill it used

Job Title, Company, City  |  Start - End
- Achievement with a number and the skill it used

EDUCATION
Degree, Institution  |  Year

SKILLS
Plain comma-separated list of skills you can defend in an interview
```

*Single column, standard headings, contact in the body. Replace the bracketed content and delete these instructions.*

Two rules make this work. First, keep the layout single-column so reading order stays linear. Second, put every contact detail in the body. If you must brand the document, do it in the body text, not in a design header the parser may drop.

## Verify before you submit

Run this list against each version of your resume before you send it. It targets the failures above, in the order that costs you the least to fix.

#### Pre-submit ATS checklist

- [ ] The file is DOCX or a text-based PDF, not an image export.
- [ ] I can highlight and copy the text when the file is open in a browser.
- [ ] My name, phone, and email are in the document body, not a header or footer.
- [ ] The layout is a single column with no side-by-side tables.
- [ ] Section headings use standard words: Experience, Education, Skills.
- [ ] My skills appear in plain text, not only inside a graphic or chart.
- [ ] If the employer specified a file format, I sent exactly that format.
- [ ] Keywords reflect the posting only where they are genuinely true of me.
- [ ] I checked the knockout questions and answered every eligibility field accurately.

## What to do next, and how to keep this current

The durable move is to stop optimizing for a mythical robot and start optimizing for two real things: a parser that reads identity fields well and skills poorly, and a human triaging a ranked pile. Get the file readable, get the skills in plain text, answer knockouts honestly, and then spend your remaining effort on the content a person will judge.

Keep the guide current by re-checking the mechanisms, not the headline numbers, since the numbers move. Market share shifts, so re-run the "which sample" question rather than trusting a single leaderboard. Parser accuracy climbs as LLM-era engines spread, but the field-weighting problem, where skills parse worse than names, is structural and worth assuming until a vendor proves otherwise for your case. Regulation is the fastest-moving piece: the EU AI Act's hiring provisions and the Mobley v. Workday litigation both point at vendors, so watch for outcomes that change what employers are allowed to automate. If you tailor a resume per posting and want the skills section to match each role's language without stuffing, that is exactly the per-application work Refolk automates from your own history, so the version you submit is both readable and true.

## Frequently asked questions

### Do applicant tracking systems automatically reject 75% of resumes?

No, that figure is not established by research. It traces to a defunct recruiting company called Preptel with no disclosed methodology. In a 2025 study of 25 US recruiters, every one used knockout questions for compliance, but only 8% configured any content-based auto-rejection, and even then only for strict thresholds like matching fewer than 7 of 10 required skills. The real hurdle is application volume and human-set criteria, not silent algorithmic binning.

### Should I submit my resume as PDF or DOCX for ATS parsing?

DOCX is the safer default. In one 2026 test DOCX averaged a 96.7% parse rate versus 91.3% for text-based PDF, and image-based PDFs (including most Canva exports) scored 4.3% and should never be submitted. That said, if an employer explicitly asks for PDF, send PDF; ignoring a stated format instruction is a worse signal than any parsing risk. Structure matters more than extension: single-column beats multi-column in either format.

### How accurate is resume parsing really?

Vendor claims cluster near 95%, but that number is field-weighted. Name and email parse at 0.99 or higher, which pulls the average up, while skill extraction in production runs at 0.75 to 0.85. Parser generation matters too: LLM-era parsers hit 95 to 97% on clean single-column documents, while legacy rule-based tools sit near 65%. The gap between name accuracy and skill accuracy is where strong candidates quietly lose ranking.

### Which applicant tracking system is most common?

It depends entirely on whose hiring you measure. Among Fortune 500 installs, Workday leads at 39% or more with SAP SuccessFactors second at 13.2%. By global revenue, iCIMS leads at 10.7%. A 2026 crawl of 3,222 top-rated employers put Greenhouse first at 49.0%. Three credible methods produce three different leaders, so there is no single correct answer without naming the sample.

### Does keyword stuffing help me pass an ATS?

No, and it can backfire. Filters reward keyword density, which tempts candidates to write to the filter rather than the role. In production data, more keywords correlated slightly worse with outcomes, around -0.09. Match the language of the specific posting where it is genuinely true of your experience, but padding the document with terms you cannot back up in an interview trains the filter to reward the wrong behavior and weakens your actual answers later.

---

*From the Refolk guide library. I revise these guides rather than replacing them, so the current version is always at https://www.refolk.ai/candidates/guides/ats-pipeline-decoder*