How an ATS actually reads your resume (2026 edition)

Most resume advice repeats the same myths: "use plain fonts", "avoid tables", "don't include images". All true. But why? Understanding the parser changes how you write — and it also tells you which of the scarier claims about applicant tracking systems are simply not true.

An ATS is a database, not a bouncer#

Start here, because almost everything else follows from it. An applicant tracking system is software a company uses to store and organize applications: a database with a search box, filters, and a workflow attached. It is not an AI sitting at the door deciding your fate.

You have probably seen the claim that some large percentage of resumes are "rejected by the ATS before a human sees them". It gets repeated everywhere and sourced nowhere. What actually happens is more mundane and more fixable: a recruiter searches and filters the pile, works down the results in order, and stops when they have enough people to interview. Nobody rejected you. They never got to you, or your record looked empty when they did.

There are two real exceptions worth knowing. Many systems support knockout questions — the short application form asking whether you're authorized to work, whether you have a required license, whether you'll relocate — and those answers genuinely can auto-disqualify. And some high-volume employers configure automatic filters on hard requirements. Both are configured by the employer, both act on form fields rather than on your prose, and neither is the mythical resume-reading robot.

The pipeline#

Every myth on the list above is a failure at step 2 or 3: if extraction scrambles your text or entity recognition can't find your dates, the match step runs against garbage — and no amount of qualification survives that.

Step 1 — Ingest

The file type decides the code path, which is why the same content can behave differently as a PDF and as a DOCX. A DOCX is a ZIP archive of XML: the parser unzips it and walks a document tree that already knows what a paragraph, a table cell and a heading are. A PDF is a page-description format that knows about glyphs at coordinates and, natively, nothing about reading order at all.

Neither is universally better — see PDF vs DOCX for which to send — but they fail in different ways, and knowing which failure you're exposed to is the point.

Step 2 — Text extraction

This is where most resumes quietly die. A PDF extractor reconstructs reading order from glyph positions, and it uses heuristics to do it. Put two columns on a page and a naive extractor can read straight across both, interleaving your skills sidebar into the middle of your job history one line at a time. The result is grammatically shredded text that still contains all your keywords — which is why it often scores as "fine" on a keyword count and reads as gibberish to a human.

Image-based PDFs are the harder failure: a resume exported as a picture, scanned, or produced by a design tool that outlines its text contains no selectable characters. Some systems run OCR over it, many don't, and the ones that do introduce their own errors.

Step 3 — Structuring

Extraction gives raw text. Structuring turns that into fields: name, contact details, a list of jobs each with a company, title and date range, education, skills. This step is why standard section headings matter far more than the font you set them in. A parser looking for a work-history section recognizes "Experience", "Work Experience", "Professional Experience". It does not reliably recognize "Where I've Made An Impact".

Dates are the other common casualty. A parser needs to read a range to compute duration — and duration is what feeds years-of-experience filters. Ranges written as Jan 2021 – Mar 2024 parse almost everywhere. Ranges written vertically beside a graphic timeline frequently don't parse at all.

Step 4 — Search, filter, rank

Now the recruiter arrives. In practice they do some combination of: run a boolean keyword search over the pool, filter by knockout answers and required fields, sort by whatever ranking the system offers, and read down the list. Your resume competes at this step as a set of parsed fields — and whatever didn't survive steps 2 and 3 simply isn't there to compete with.

The classic failure modes#

For the full prescriptive checklist — fonts, dates, file export — see the ATS-friendly resume format, and for which file type to send, PDF vs DOCX.

Why the same resume parses differently everywhere#

"The ATS" isn't one product. Workday, Greenhouse, Lever, Taleo, iCIMS, SuccessFactors, Ashby and a long tail of others each ship their own parsing stack, and they disagree — about columns, about text boxes, about headers and footers, about which date formats they'll read.

This is the single most under-appreciated fact in resume advice. A layout that survives the system you tested against can lose an entire section in the next one, and you never find out, because nothing tells you. It's the reason MeritSlate simulates nine vendor parsers against your actual file rather than declaring your resume "ATS-friendly" on the strength of one.

What actually filters you out#

In rough order of how often we see it:

1. Targeting. You don't cover the posting's stated requirements, so you rank below people who do. No formatting fix touches this one.
2. Evidence. You cover the requirements but assert them without outcomes, so a human comparing you against someone specific picks the specific one.
3. Parse damage. A section, your dates, or your contact details didn't survive extraction, so you compete with a partial record.
4. Knockout answers. A form field disqualified you before your resume was read at all.

Notice that the formatting problem everyone worries about is third. It's real, it's worth fixing, and it is not usually the thing costing you interviews — which is why a silent job search is best diagnosed from the top of that list down.

What to do next#

Run the free MeritSlate ATS checker to see exactly which of these issues your current resume hits — it simulates nine vendor parsers against your actual file, and shows you the extracted text each one produced.