Home/What an ATS checks
What an ATS actually checks on your resume
An applicant tracking system does not reject your resume. It reads it — and if it reads it wrong, a recruiter never finds you. Those are different failures with different fixes, and almost every piece of resume advice blurs them together. Here is what actually happens, and the 21 mechanical checks we run.
Published 25 October 2026, last updated 27 October 2026. Every rule below is the one this site actually runs — if a rule changes, this page changes with it.
Read this first: this page documents the rule set inside ATS Score, which is our own product. We publish it because you should be able to check our work — and because the useful part (how a parser mangles a two-column layout) is true whether or not you ever use our tool. Where a rule depends on local hiring convention rather than mechanics, we say so.
How an ATS reads a resume
Four things happen between you clicking submit and a recruiter seeing your name. Most resume problems are born in the first one.
1. Extraction — the file becomes text
A PDF is a page-description format, not a document format. It contains instructions like "draw this glyph at x=412, y=96". There is no object in the file that says "this is a line" or "this is a column". So before anything else can happen, the parser has to guess lines back out of coordinates — and that guess is where most of the damage occurs.
Two PDFs that look identical on screen can extract in completely different orders depending on which tool exported them. The classic case is a two-column layout: the exported text frequently comes out interleaved, one line from the left column, then one from the right, all the way down the page.
2. Segmentation — the text becomes fields
Once there is text, the parser has to decide what each chunk is: this is the contact block, this is a job, this line is the employer, that line is the title, these are the dates. It does that by matching patterns it already knows — section headings, date formats, and position on the page.
Anything it does not recognise does not politely disappear. It lands in the wrong field, or in a catch-all bucket. A section titled "Where I've Been" has a real chance of not being read as work history at all; its contents get folded into whatever section came before it.
3. Indexing — the fields become a searchable record
The segmented fields become one record per candidate: name, employers, titles, dates, skills, education. From here on, what exists is not your resume. It is that record.
4. Search — not rejection
This is the part most resume advice gets wrong. In the systems we looked at, the common pattern is not "the ATS rejects 75% of resumes". It is:
- The recruiter searches the index for keywords — often with boolean operators — and gets a list.
- Candidates who are not in that list were never rejected. They were never seen.
- Separately, some postings have knockout questions on the application form: work authorisation, years of experience, salary expectations. Those do reject you — but they are form fields, not parsing. A perfect resume does not save you from answering "no".
Two different failure modes, two different fixes. A parsing failure is a file problem: fix the file. A knockout failure is a fit problem: no resume edit changes it. Most "the ATS rejected me" stories are the first kind — but not all, and no scanner can tell you which, because a scanner cannot see the employer's form.
1–4 · Can it read the file at all?
These four are upstream of everything else. If a file fails here, the quality of the writing inside it is irrelevant — nothing downstream ever sees it.
1. Is there any text in the file?
If you paste text, this passes automatically. If you upload a PDF, we count extractable characters per page. A scan, a photo, or a "flattened" export has almost none.
No text layer means nothing to segment and nothing to index. The resume can look immaculate and still be empty to every system that reads it.
Fix: open the PDF and try to select a sentence with your cursor. If you cannot highlight it, no parser can read it. Re-export from the original document with text enabled, or send a DOCX.
2. Two-column layout, tables, and text boxes
We score the layout rather than the design: a ratio of tabular-looking lines for columns, plus flags for tables and text boxes. Tables and text boxes weigh heaviest, because their content is positioned absolutely and lands wherever the extractor happens to put it.
A three-cell table row becomes one run-on line, or three lines scattered far apart. A sidebar built from a text box usually turns up at the end of the page — after your work history, where nobody expects to find your skills.
Fix: one column, real bullets, no table cells. If you want dates aligned on the right, use tab stops — a right-aligned date at the end of its own line extracts cleanly, which is why we treat those lines as safe rather than risky.
3. Are the section headings the boring ones?
We look for the standard set — contact, summary, experience, education, skills, projects — and subtract for custom headings that have real content underneath them.
Headings are the anchors segmentation uses. "Professional Experience" is recognised. "The Story So Far" is a coin flip. When an anchor is missed, the block under it does not vanish; it gets absorbed into the previous section, which is how a work history ends up filed as part of a summary.
Fix: Experience, Education, Skills, Projects, Certifications. Put the personality in the sentences, not the headings.
4. Contact details buried in a header or footer
Header and footer regions are extracted separately from the body, and a number of parsers drop them outright. Your email address is then missing from the very record the recruiter searches — you become a candidate who cannot be contacted.
Fix: name, email, phone, city and LinkedIn in the first body block, above everything else. Not in a Word header, not in a page-footer, not in a text box pinned to the top margin.
5–9 · Structure
These five are mostly about the reader rather than the parser. They still cost you, just later in the process.
5. Page count versus years of experience
One page is the target below ten years of experience, two between ten and twenty, three beyond that. Two pages early in a career is penalised hard — harder than being one page short.
This is not a parsing rule. It is attention: a recruiter working through eighty applications spends seconds on the first pass, and page one is most of what gets read.
Fix: cut the oldest or least relevant role rather than shrinking the font. If you have publications or patents, the extra page is defensible; otherwise it rarely is.
6. Are the contact details complete and parseable?
We weight email at 40%, phone at 35%, city at 15%, LinkedIn at 10%.
Email and phone are how you get contacted. If either fails to parse, the application is dead no matter how good the resume is. City matters because it feeds location filters. LinkedIn rarely decides anything, but it costs one line.
Fix: plain text — not an icon, not an image, not a hyperlink whose visible text reads "Email me". One phone number, with the country code if you are applying across borders.
7. Are date formats consistent?
We count the distinct date formats in the document. One is clean; two costs you; three or more fails.
Dates are what the parser uses to build employment ranges. "03/2021 – Present" and "Mar 2021 – Present" are two different patterns; mixed in one document, a single role can split into two entries, or a duration comes out years wrong.
Fix: pick one format and use it everywhere. Month plus year is the safest across locales — "03/2021" reads as March in the US and as invalid or as the third of an unknown month elsewhere.
8. Unexplained employment gaps
We find the intervals between one role ending and the next beginning, and check whether anything in the text explains them.
Again, not parsing — a gap is a question in the reader's head, and an unanswered question during a twenty-second scan tends to resolve to "no".
Fix: put a line on the resume: "Career break — full-time caregiving, 2022–2023". It takes one line and closes the question. You do not owe anyone an apology for it.
9. Photo, age, marital status, nationality and similar fields
We flag photographs, dates of birth, marital status, gender, nationality and national ID numbers.
This one is genuinely regional. In the US and UK these create equal-opportunity exposure, and recruiters are trained to set such resumes aside rather than risk the appearance of bias. In Germany, France and much of East Asia, a photo is expected — omitting it makes the CV look unfinished. The rule is not "delete the photo"; it is "know which market you are applying to".
Fix: for US, UK and Canadian applications, remove them. For DACH and East Asian markets, follow local convention. Applying to both means keeping two versions of the file — which you should be doing anyway.
10–13 · What the bullets actually say
This is where a resume stops being a list of jobs and starts being evidence.
10. Does every bullet start with a verb?
We measure the share of bullets opening with an action verb, and set the target at 80% rather than 100% — a bullet starting with a noun phrase is not automatically wrong.
Bullets beginning with "Responsible for" describe a job description, not a person. They also waste the first two words of every line, and the first two words are the part that gets read.
Fix: "Responsible for managing the onboarding pipeline" becomes "Managed a 40-person onboarding pipeline". Start with what you did.
11. Quantified results
The heaviest single rule in the set. We look for numbers, percentages and amounts inside bullets, and target half of them — not all of them, because a number invented to fill a slot is worse than no number.
A number is the difference between a claim and evidence. "Improved performance" cannot be checked; "cut p95 latency from 800ms to 120ms" can be sized up in one second.
Fix: at least two numbers per role. If you have no business metrics, use scale — team size, dataset size, budget, uptime, ticket volume, number of users served.
12. Vague filler phrases
We count occurrences of phrases like "team player", "hard-working", "results-driven", "detail-oriented" and "go-getter".
They are unfalsifiable, every applicant uses them, and so they carry no information at all. Worse, they occupy space a number could have used.
Fix: delete them. Do not replace them with synonyms — replace them with the evidence they were standing in for.
13. Are the skills concrete tools?
We check whether the skills section names things — languages, frameworks, platforms, certifications — rather than qualities. Three or more generic entries caps the score regardless of how good the rest is.
Keyword search matches strings. "Python" is a string a recruiter searches for. "Strong analytical skills" is not.
Fix: list tools you have actually used. If the list runs long, group it: Languages / Frameworks / Infrastructure / Tools.
14–17 · Matching the job
These four need a job description. Without one they are skipped entirely and the score is normalised over the other seventeen, so the number stays comparable either way — it just measures less.
14. Hard-skill keyword coverage
The heaviest rule overall. We extract the concrete skills a posting asks for and check whether they appear in your resume, matching on word boundaries so that "Java" does not score a hit for "JavaScript".
This is the search the recruiter actually runs. Below roughly 70% coverage is where a resume stops surfacing in that search.
Fix: use the posting's exact term for anything you have genuinely done. If they say "Kubernetes" and you wrote "container orchestration", you are invisible to that query — same skill, different string.
15. Job-title alignment
We compare the words in the posting's title against your resume, and separately against your first few lines, since a headline title carries more weight than one buried in a role.
Recruiters filter by title. A title matching the posting is the shortest possible signal that you are the same kind of thing they are looking for.
Fix: if your internal title is idiosyncratic — "Member of Technical Staff" — put the market term next to it: "Member of Technical Staff (Senior Software Engineer)".
16. Industry terminology hit rate
We check for domain vocabulary drawn from the posting's industry.
It is how a reader decides whether you understand the context, and how search surfaces candidates from adjacent industries.
Fix: if you have real exposure to the domain, name it explicitly — the compliance regime, the customer segment, the product category. If you do not, leave it out. Faking domain knowledge is what interviews are designed to catch.
17. Soft-skill keyword match
The lightest of the four. Some postings are keyword-matched on phrases like "stakeholder management" or "cross-functional", and having the phrase where it is true costs nothing.
Fix: only where it describes something real you did — and ideally inside a bullet that also contains a number.
18–21 · Detail and risk
18. Concrete outcomes and scale
We look for bullets that combine a metric, an action verb and at least eight words — and give full marks when at least two of those metrics are hard ones: percentages, money, user counts, records, transactions.
"Worked on the migration" is a task. "Migrated 40M order records to Postgres with zero downtime" is an outcome with a size attached.
Fix: for each role, write one bullet that answers two questions: how big was this, and what changed because of it.
19. Certifications and licences
We look for a certifications section first, then for recognisable credential names.
In some fields the licence is a hard gate — nursing, accounting, teaching, cloud architecture. In others it is a tiebreaker. Either way, an unlisted credential cannot help you.
Fix: give it its own section, with the full name and the year. Remove or mark anything expired — a lapsed licence reads worse than no licence.
20. Spelling
We check against a dictionary plus repeated-word detection.
It is the cheapest possible reason to be screened out. A spell-checker will not catch "manger" when you meant "manager", because "manger" is a word.
Fix: read it backwards, bottom bullet to top. Your eye stops autocorrecting when the sentences no longer flow.
21. Keyword stuffing, including hidden text
We detect white-on-white text, zero-width characters, and text set at a size too small to see.
These are detectable, and the consequence is not a lower score — it is a resume a recruiter can no longer trust. Some systems strip hidden text and the trick simply fails; others keep it, and your record reads as gibberish to the person searching.
Fix: never. Cover the keyword honestly or leave it out.
What none of this can tell you
We think the limits matter more than the feature list, so here they are plainly.
- Your score is not the employer's score. Nobody outside a vendor knows how that vendor ranks candidates. Ours is 21 deterministic rules that approximate the failure modes common to most parsers. Treat it as a mechanical pre-flight check, not a verdict — and read how to read that number before you act on it.
- Whether a human will like it. These checks find mechanical problems. They cannot tell you whether your experience is relevant to this role.
- Whether you will get an interview. A clean parse gets you into the search results. It does not get you chosen, and it does not make a mismatched application fit.
- Whether the job is real. Some postings are already filled, or were opened to satisfy a process. Nothing on your side of the desk changes that.
And one more, because it is the most repeated number in this industry: the claim that 75% of resumes are rejected by ATS has no study behind it. It traces to a 2012 press release from Preptel, a resume-optimisation vendor, which shut down in August 2013 without ever publishing a sample, a method or a dataset. The number still circulates as 70% or 88% depending on who repeats it — which is what happens to a figure that never had an original to check against.
What the available evidence actually shows: when Enhancv asked 25 US recruiters across nine applicant tracking systems in 2025, 23 of them — 92% — said their system does not auto-reject on resume content at all, while all 25 used knockout questions. Harvard Business School and Accenture's Hidden Workers report (2021; 2,275 employers across the US, UK and Germany) found that 88% of employers agreed their own process screens out qualified candidates who do not match a job description's exact wording. That is a different claim, and a more useful one: the failure is ranking and wording, not a robot deleting you. Both are mechanical, so both are testable before you apply.
The one-page checklist
If you only do ten things before submitting, do these. The arithmetic behind the score — every rule's weight, the normalisation, and a worked example you can check by hand — is on how the score is calculated.
- Select text in your PDF. If nothing highlights, re-export it.
- One column. No tables, no text boxes, no sidebar.
- Contact block in the body, not in a header or footer.
- Standard headings: Experience, Education, Skills, Projects.
- One date format, used everywhere, month and year.
- Email and phone as plain text with a country code.
- Every bullet starts with a verb.
- At least two numbers per role.
- Skills are tools, not adjectives.
- Paste the job description and read the missing keywords before you send it.
Questions people ask
- Does my file get uploaded?
- Not for the free scan. Parsing and scoring both run in your browser and the file never reaches a server. The paid rewrite is the one exception: the text you submit for it is sent to our server and to DeepSeek (processed in China) and is not kept afterwards.
- Is this the same score the company's ATS will produce?
- No. It measures mechanical failure modes that are common across parsers — it is not a vendor's ranking, and no external tool can reproduce one. Treated as a pre-flight check it is useful; treated as a verdict it is misleading.
- Do I have to paste a job description?
- No. Seventeen of the 21 checks run without one, and the score is normalised over those seventeen so the number stays comparable. You just get less information — the four job-match checks are usually where the actionable gaps are.
- I'm applying in Germany and my CV has a photo. Is that wrong?
- Not there. The photo rule is about US, UK and Canadian hiring practice, where it creates bias exposure. Follow local convention — and if you are applying to both markets, keep two files. Back to rule 9.
- Will the AI rewrite invent experience for me?
- No. It reorganises and rewords what you already wrote and flags the gaps it cannot fill. Anything you have to supply yourself comes back in square brackets for you to replace.
Comparing tools rather than rules? There is a price and feature comparison with Jobscan and Resume Worded at our Jobscan alternative page, including the cases where they are the better choice.
See which of the 21 you fail
Upload or paste, get the score in about ten seconds. No account, no card, and the file stays on your machine.