<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[CVPosh]]></title><description><![CDATA[CVPosh]]></description><link>https://cvposh.hashnode.dev</link><generator>RSS for Node</generator><lastBuildDate>Sat, 21 Nov 2026 17:16:38 GMT</lastBuildDate><atom:link href="https://cvposh.hashnode.dev/rss.xml" rel="self" type="application/rss+xml"/><language><![CDATA[en]]></language><ttl>60</ttl><item><title><![CDATA[We ran one resume through 20 layouts and a text extractor: here is what got lost]]></title><description><![CDATA[One resume, 20 layouts, the same text extraction our ATS checker uses. Contact details, headings and reading order failed repeatably.]]></description><link>https://cvposh.hashnode.dev/we-ran-one-resume-through-20-layouts-and-a-text-extractor-here-is-what-got-lost</link><guid isPermaLink="true">https://cvposh.hashnode.dev/we-ran-one-resume-through-20-layouts-and-a-text-extractor-here-is-what-got-lost</guid><category><![CDATA[resume]]></category><dc:creator><![CDATA[Kien Phan]]></dc:creator><pubDate>Wed, 18 Nov 2026 14:57:26 GMT</pubDate><content:encoded><![CDATA[<h2>TL;DR</h2>
<ul>
<li>We rendered one synthetic resume 20 ways (14 PDFs, 6 DOCX) and ran each through the text extraction libraries behind CVPosh's free ATS checker, counting how many of 37 known content items survived.</li>
<li>The worst losses came from where text lives, not how it looks. Contact details in a DOCX header or footer, drawn as an image, or hidden behind "Email me" link labels were lost in 4 of 20 files. A resume saved as one page image lost everything (0 of 37).</li>
<li>Columns and positioned blocks kept every word but changed reading order in 4 layouts, and the checker still scored two of them 100. A high checker score does not prove a clean parse.</li>
</ul>
<p>This tests text extraction, not any employer's applicant tracking system. We did not run Workday, Greenhouse, Lever or any other ATS and cannot say how they behave.</p>
<h2>What we tested</h2>
<p>We wrote one fictional resume ("Jordan Sample": 228 words, four sections, nine bullets, ten skills) and generated 20 versions of it. Only the layout changed: single column, two-column sidebar, table layout, different bullet styles, emoji icons, letter-spaced headings, unusual heading names, a ligature-heavy serif font, absolutely positioned blocks, contact details as an image or as link labels, headings as images, and a full-page image.</p>
<p>For each file we extracted text and looked for 37 items: name, email, phone and link; role, company and dates for three jobs; the education line; ten skills; a certification; the summary; and all nine bullets as unbroken sentences. We then ran the CVPosh ATS checker's scoring code on the extracted text.</p>
<h2>Results</h2>
<p>PDFs were extracted with <code>pdf-parse</code> 1.1.1 and DOCX files with <code>mammoth</code> 1.11.0, the libraries the checker uses. "Order" counts how many of 10 landmarks in the resume (summary, employers, bullets, education, skills) appear in the expected sequence.</p>
<table>
<thead>
<tr>
<th>Layout</th>
<th>Contact found</th>
<th>Sections found</th>
<th>Items found</th>
<th>Order</th>
<th>Checker score</th>
</tr>
</thead>
<tbody><tr>
<td>PDF baseline, single column</td>
<td>all</td>
<td>4/4</td>
<td>37/37</td>
<td>10/10</td>
<td>100</td>
</tr>
<tr>
<td>PDF two-column sidebar</td>
<td>all</td>
<td>4/4</td>
<td>36/37</td>
<td>8/10</td>
<td>100</td>
</tr>
<tr>
<td>PDF table layout</td>
<td>all</td>
<td>4/4</td>
<td>37/37</td>
<td>10/10</td>
<td>100</td>
</tr>
<tr>
<td>PDF arrow-glyph bullets (➤)</td>
<td>all</td>
<td>4/4</td>
<td>37/37</td>
<td>10/10</td>
<td>90</td>
</tr>
<tr>
<td>PDF CSS-drawn bullets</td>
<td>all</td>
<td>4/4</td>
<td>37/37</td>
<td>7/10</td>
<td>90</td>
</tr>
<tr>
<td>PDF emoji icons</td>
<td>all</td>
<td>4/4</td>
<td>36/37</td>
<td>10/10</td>
<td>100</td>
</tr>
<tr>
<td>PDF contact as image</td>
<td>none</td>
<td>4/4</td>
<td>34/37</td>
<td>10/10</td>
<td>80</td>
</tr>
<tr>
<td>PDF letter-spaced headings</td>
<td>all</td>
<td>1/4</td>
<td>37/37</td>
<td>10/10</td>
<td>80</td>
</tr>
<tr>
<td>PDF creative heading names</td>
<td>all</td>
<td>1/4</td>
<td>37/37</td>
<td>10/10</td>
<td>80</td>
</tr>
<tr>
<td>PDF ligature serif font</td>
<td>all</td>
<td>4/4</td>
<td>37/37</td>
<td>10/10</td>
<td>100</td>
</tr>
<tr>
<td>PDF positioned blocks, shuffled order</td>
<td>all</td>
<td>4/4</td>
<td>37/37</td>
<td>7/10</td>
<td>100</td>
</tr>
<tr>
<td>PDF contact behind link labels</td>
<td>none</td>
<td>4/4</td>
<td>34/37</td>
<td>10/10</td>
<td>80</td>
</tr>
<tr>
<td>PDF headings as images</td>
<td>all</td>
<td>1/4</td>
<td>37/37</td>
<td>10/10</td>
<td>80</td>
</tr>
<tr>
<td>PDF full page as image</td>
<td>none</td>
<td>0/4</td>
<td>0/37</td>
<td>n/a</td>
<td>30</td>
</tr>
<tr>
<td>DOCX baseline (Word list bullets)</td>
<td>all</td>
<td>4/4</td>
<td>37/37</td>
<td>10/10</td>
<td>90</td>
</tr>
<tr>
<td>DOCX baseline (typed "•" bullets)</td>
<td>all</td>
<td>4/4</td>
<td>37/37</td>
<td>10/10</td>
<td>100</td>
</tr>
<tr>
<td>DOCX contact in header</td>
<td>none</td>
<td>4/4</td>
<td>33/37</td>
<td>10/10</td>
<td>70</td>
</tr>
<tr>
<td>DOCX contact in footer</td>
<td>none</td>
<td>4/4</td>
<td>34/37</td>
<td>10/10</td>
<td>70</td>
</tr>
<tr>
<td>DOCX table layout</td>
<td>all</td>
<td>4/4</td>
<td>37/37</td>
<td>10/10</td>
<td>90</td>
</tr>
<tr>
<td>DOCX two-column table</td>
<td>all</td>
<td>4/4</td>
<td>37/37</td>
<td>8/10</td>
<td>90</td>
</tr>
</tbody></table>
<p>Raw numbers for both extractors are in the <a href="https://cvposh.com/research/resume-layout-test/results.csv">results CSV</a>.</p>
<h2>Finding 1: contact details are the most fragile field</h2>
<p>Four files lost email, phone and link together: the contact line as an image, the contact line as link labels ("Email me", "Call me") with the real addresses only in link targets, and contact details in a DOCX header or in a DOCX footer.</p>
<p>In the link-label PDF the address exists only as a link annotation, and neither <code>pdf-parse</code> nor poppler's <code>pdftotext</code> returned it. For the DOCX files, <code>mammoth</code> returned the body only; in the header file even the name was gone because we had put it in the header too. In the checker this cost 20 points each (15 for no email, 5 for no phone). A contact line in the page body, as plain text, was extracted by every method we tried.</p>
<p>Also: in the two-column PDF the LinkedIn URL wrapped at its hyphen inside the narrow sidebar. <code>pdf-parse</code> returned <code>jordan-</code> and <code>sample</code> on separate lines and <code>pdftotext</code> dropped the hyphen (<code>jordansample</code>). The checker still saw a link, but not a working URL.</p>
<h2>Finding 2: headings need words the parser recognises</h2>
<p>The checker looks for experience, education, skills and summary or profile. Three PDFs fell to 1 of 4 sections:</p>
<ul>
<li><strong>Headings as images.</strong> The text was not in the text layer.</li>
<li><strong>Creative names</strong> ("About me", "Where I've worked", "Studies", "Toolbox"). All content was extracted, but the headings were not words a checklist knows. The one section still found came from "years of experience" in the summary.</li>
<li><strong>Letter-spaced uppercase headings.</strong> <code>pdf-parse</code> returned <code>E X P E R I E N C E</code>; <code>pdftotext</code> returned <code>EXPERIENCE</code>. The same file scored 80 or 100 depending on the extractor.</li>
</ul>
<p>Each cost 20 points. The bigger risk is any system that groups your history by section: it has to recognise the header first.</p>
<h2>Finding 3: columns and positioned blocks keep the words, not the order</h2>
<p>The two-column sidebar, the DOCX two-column table, the positioned-blocks PDF and the CSS-drawn-bullets PDF kept every word, but <code>pdf-parse</code> changed the order:</p>
<ul>
<li>In both column layouts, skills and education came before summary and experience, because the left column is read first.</li>
<li>In the CSS-bullets PDF, all nine bullets came after the three job headings, detached from their employers. Our dots were small positioned shapes, which changed the order the text was written to the page.</li>
<li>In the positioned-blocks PDF, which looks like a normal one-page resume, skills came first and the name last.</li>
</ul>
<p><code>pdftotext</code> re-sorted the CSS-bullet and positioned-block files into correct visual order (10 of 10). So the problem is partly the file and partly the extractor, and we cannot tell you which one an employer's system resembles. The score did not move: the checker has no reading-order rule, so the two-column PDF and the positioned-blocks PDF both scored 100. A checklist can show that data is present, not that it is in the right place.</p>
<h2>What did not break anything</h2>
<p>A table layout (dates in one column, roles in the other) extracted in order as PDF and DOCX. A Palatino font with ligatures enabled returned normal text, including "Airflow" and "Certified". Emoji next to contact details left the values intact, although they came out as private-use characters in <code>pdf-parse</code>. A ➤ bullet was extracted as a character.</p>
<h2>Our own checker undercounts bullets</h2>
<p>We found this while testing. The checker counts a bullet only when a line starts with <code>•</code>, <code>-</code> or <code>*</code>. A DOCX with real Word list bullets scored 90 and the same file with typed "•" scored 100, because <code>mammoth</code> does not put list markers in its raw text. The ➤ PDF also scored 90. So the checker deducts 10 points from some resumes whose bullets read fine, which is one more reason the score and a real parse can disagree.</p>
<h2>What to do with this</h2>
<ol>
<li>Put your name and contact line in the document body as plain text, not in a Word header or footer, an image, or link labels.</li>
<li>Use the section names a reader expects: Summary, Experience, Education, Skills.</li>
<li>Keep headings as real text, without letter-spacing.</li>
<li>If you use columns, paste the PDF into a text editor, or run it through <a href="https://cvposh.com/tools/analyzers/ats-checker">the checker</a>, and read the order that comes out.</li>
<li>Keep long URLs out of narrow columns.</li>
<li>Export a text-based PDF. If you can't select the text, nothing else works.</li>
</ol>
<h2>FAQ</h2>
<p><strong>Does an ATS reject resumes with columns?</strong>
We did not test any employer ATS, so we can't say. In our extraction every word of the two-column files survived, but the reading order changed.</p>
<p><strong>Is DOCX safer than PDF?</strong>
Not in our test. DOCX avoided the ordering problems but was the only format where header and footer content disappeared.</p>
<p><strong>Do icons hurt?</strong>
Icons next to text did not remove the text. Text drawn as an image did.</p>
<p><strong>Does a high checker score mean my resume will parse correctly?</strong>
No. The checker starts at 100 and subtracts points when a rule fires, and it has no reading-order rule. Two layouts with changed order still scored 100.</p>
<p><strong>Is this representative of real templates?</strong>
No. It is one synthetic resume laid out by us in Chrome and a DOCX library. Word, Google Docs, Canva or LaTeX output can behave differently. Treat the results as a list of things to check, not as rates.</p>
<h2>Methodology and data</h2>
<ul>
<li><strong>Date run:</strong> 2026-11-17. One fictional resume, no real person.</li>
<li><strong>Variants:</strong> 14 PDFs from Google Chrome 157.0.8103.79 (headless print-to-PDF, Letter) and 6 DOCX files from <code>docx</code> 8.6.0. Image variants are screenshots of the same HTML.</li>
<li><strong>Extraction:</strong> <code>pdf-parse</code> 1.1.1 and <code>mammoth</code> 1.11.0, as in the checker. PDFs were also run through poppler <code>pdftotext</code> 26.07.0 (default mode) as a cross-check; both sets of numbers are in the CSV.</li>
<li><strong>Scoring:</strong> <code>ats-scorer.ts</code> from the CVPosh repository at <code>origin/main</code> commit <code>184d4b4</code>. We also uploaded five of the files (PDF baseline, contact as image, letter-spaced headings; DOCX baseline and header) to the live checker API at cvposh.com; all five returned the same score as our local run. The other 15 were not uploaded.</li>
<li><strong>Items:</strong> 37; an item is found when its text appears after whitespace and case normalisation, bullets as complete sentences.</li>
<li><strong>Limits:</strong> one resume, one run (the pipeline is deterministic), no employer ATS, no OCR. Chrome's native list markers were not extracted by either extractor, so the baseline draws bullets as text ("•"). "Order" is a landmark test, not a full reading-order evaluation.</li>
<li><strong>Reproduce:</strong> scripts and CSV at <a href="https://cvposh.com/research/resume-layout-test/results.csv">cvposh.com/research/resume-layout-test</a>.</li>
</ul>
<p><em>CVPosh is our product. The checker is free and needs no sign-up. <a href="https://cvposh.com/tools/how-the-ats-checker-scores">How its score is calculated</a>.</em></p>
<hr />
<p>Originally published at <a href="https://cvposh.com/blog/resume-layout-text-extraction-test">cvposh.com/blog/resume-layout-text-extraction-test</a>.</p>
<p>Disclosure: we build CVPosh; this article was written with AI assistance and the data was checked against our own systems.</p>
]]></content:encoded></item></channel></rss>