Does an ATS read PDF or Word?

Send either one. An applicant tracking system reads both PDF and Word (.docx) as long as your resume is single-column and made of real, selectable text. In our own June 2026 test, a clean single-column PDF lost none of its fields and every Word file read cleanly, so the file type is rarely what breaks. What breaks a machine read is the layout: text saved as an image, side-by-side columns, contact details stranded in a page header or footer, and unusual embedded fonts. Pick the format the job posting asks for. If it says nothing, either is safe when the file is single-column and text-based, and never a picture of a resume.

By UnchartedCareer. Last updated: July 2026.

Every measured figure below comes from the same source: a test of 48 synthetic resumes read by three local open-source text extractors (pdfplumber, pdfminer.six, python-docx), June 2026. We contacted no commercial product. These extractors run the same kind of text step a real system runs first, so read the numbers as a floor on what can go wrong in the reading step, not a verdict on any named ATS.

Does it matter whether you send a PDF or a Word file?#

Less than the internet tells you. We wanted a real answer instead of a repeated rule, so we built the test described above: 48 synthetic resumes across nine layouts, each saved as both PDF and DOCX, each read by three open-source text extractors, for 600 scored field reads. A field passed only when its content came out of the extracted text in the right order.

Here is the headline that looks like it settles the format war, and does the opposite. Across every layout, Word files lost 0 percent of their field reads and PDF files lost 19.9 percent (48 synthetic resumes, three local open-source extractors, June 2026, a floor on what can go wrong and not a verdict on any product). Read fast, that says never send a PDF. Read honestly, it says something narrower.

The loss was not spread across PDFs evenly. A clean single-column PDF lost 0 percent of its 48 field reads, exactly as clean as the Word version. The 19.9 percent came almost entirely from a handful of designed layouts, and two of the worst of those existed only as PDFs in our test. Set the layout right and the format question dissolves.

Why did PDF lose more than Word in our test?#

Two reasons, and both are about mechanics rather than magic.

First, how each file stores your words. A .docx keeps your text as a structured stream a reader can lift straight out. A PDF stores glyphs at fixed positions on a page, and the extractor has to reassemble them into reading order. That reassembly is where a busy layout trips, and a Word file skips the step entirely.

Second, the test's own shape. Two of our nine layouts, text rendered as an image and three side-by-side columns, existed only as PDFs, because they are PDF-native problems. Every failure they produced counted against PDF and none against Word. Text-in-image alone lost 37.5 percent of its fields and three columns lost 14.6 percent (June 2026 reading test, a floor and not a verdict), and because these extractors run no OCR while some commercial systems do, treat the image figure as a worst case. Those two layouts contributed 25 of the 86 total PDF failures. Strip them out and PDF still trailed Word, but the gap narrows and the reason is always the layout underneath, not the three letters after the dot.

One more honest caveat. We rendered our PDFs from HTML through a headless browser. A resume exported straight from Word or Google Docs may embed its fonts differently, so treat the unusual-font result as directional, not a promise about your exact file.

So which should you send?#

Here is the tension worth seeing. In our reading test, Word edged out PDF: every .docx read cleanly, and even for the same layout, PDFs lost a little more. But Greenhouse, one of the larger applicant tracking systems, tells applicants the reverse in its own support documentation (fetched June 2026). It recommends PDF for best results, treats Word as equally acceptable, and singles out the image-based upload as the file that breaks.

Both are right about the part that decides the outcome. A clean, single-column, text-based file reads fine either way, and a resume saved as an image is the one that fails. So the verdict is boring on purpose:

• Send exactly what the posting asks for. If it names a format, that is your answer.
• If it says nothing, a single-column text-based PDF is safe and preserves what the human sees after the parse. Word is safe too, and had the marginally more forgiving machine read in our test.
• Do not agonize between the two. Spend that energy on the layout, which is where the real losses live.

What actually breaks a machine read?#

The layout does. Here is the field-loss by layout from the June 2026 reading test, worst to best. Read every number as a floor on what can go wrong, not a verdict on any named product.

| Layout | Field-read loss | |—-|—-| | Text saved as an image | 37.5% | | Contact details in a page header | 25% | | Contact details in a page footer | 25% | | Unusual embedded font | 20.8% | | Three side-by-side columns | 14.6% | | Work history inside a bordered table | 12.5% | | Two columns | 1.4% | | Nonstandard section headings | 0% | | Single column | 0% |

The image layout is the one to fear. On those resumes the extractors dropped the candidate's name, email, and phone 100 percent of the time, because a picture of text gives a text reader nothing to read (June 2026 test, a floor and not a verdict). These extractors run no OCR and some commercial systems do, so treat that 100 percent as a worst case. No format choice saves an image-based resume, and no format choice is needed for a clean single-column one.

Two results are worth pinning up. Nonstandard section headings lost 0 percent, so calling a section "Where I've worked" instead of "Experience" cost nothing in the reading step, because the parser reads the content underneath the label. And a two-column resume came through at 1.4 percent, close enough that a single clean column is the only safe default worth insisting on. Greenhouse's own guidance names the same set of culprits: multi-column layouts, tables, headers and footers, contact information in a text box, graphics, and image-based files (support documentation, June 2026).

If you came here half-afraid that a machine auto-rejects most resumes on sight, that is a separate and older myth, and we trace it to its actual source in the "75 percent auto-rejected" story.

How do I check my own resume in ten minutes?#

Three checks, no tool required.

The parse test. Select all the text in your resume, copy it, and paste it into the plainest editor you own (TextEdit in plain-text mode, or Notepad). Read what lands. If your name, email, phone, and job titles all come through in the right order, a real parser will manage the same. If the contact line vanishes, or two columns fuse into word soup where "Senior Nurse" sits three words away from "Mercy General," a machine sees that scramble too.

The image check. Try to select your name inside your PDF with the cursor. If it highlights as text, good. If it will not select, or highlights as one solid block, your resume is a picture and a parser reads nothing on it. This is the single failure that actually makes a resume disappear.

The single-column check. Look at your page. If any two blocks of text sit side by side, read straight across one line the way a machine does. If the words from the left column collide with the words from the right, that is the interleaving that dropped fields in our test.

You can run all three by hand, and it is worth doing once because it shows you exactly what the machine is blind to. When you want the machine's own read of your file, UnchartedCareer's free ATS resume scan scores it from 0 to 100 and flags the sections a parser cannot read, with no sign-up. Same logic as the manual test, faster.

How we tested this#

The ATS Autopsy is our own test. We generated 48 synthetic resumes from three fabricated personas (no real people, no personal data, reserved example.com addresses) across nine layouts, saved each as PDF and DOCX, and read every file with three local open-source text extractors: pdfplumber 0.11.4, pdfminer.six 20231228, and python-docx 1.1.2, with the PDFs rendered through a headless Chromium shell. That produced 600 scored field reads, each field passing only when its content was recoverable in the right order, run on 2026-06-24 and reproducible exactly because the generator and scorer are deterministic.

We tested no named commercial ATS and make no claim that any product rejects or discriminates against a candidate. These extractors are the same kind of text step those systems run first, so every figure here is a floor on what can go wrong in the reading step, not a verdict on any named ATS. The vendor-behavior details come from Greenhouse's own support documentation, fetched June 2026. Where our measured result and a vendor's guidance point different ways, we show both.

Frequently asked questions#

Is PDF or Word better for an ATS?#

Neither wins by much. A single-column, text-based file reads cleanly in both. In our June 2026 reading test Word lost 0 percent of fields and a clean single-column PDF also lost 0 percent, while PDFs with hard layouts lost more. Send what the posting asks for, and keep the layout simple.

Will a PDF get my resume auto-rejected?#

No. A clean text-based PDF is fully readable, and Greenhouse recommends PDF for best results (June 2026). The idea that a file type triggers an automatic rejection is part of a separate myth about ATS auto-rejection that has no credible source.

Does saving as PDF turn my resume into an image?#

Usually no. A PDF exported from Word or Google Docs keeps real, selectable text. It becomes an image only if you scan a printout or export it as a picture. Check by trying to select your name with the cursor. If it will not highlight as text, rebuild the file from a document, not an image.

Is a two-column resume safe?#

It is close. Two columns lost 1.4 percent of field reads in our June 2026 test, a floor and not a verdict on any product, but a single clean column lost nothing and removes the risk entirely. If you can drop to one column, do.