<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Kien Phan</title>
    <description>The latest articles on DEV Community by Kien Phan (@kienpt91).</description>
    <link>https://dev.to/kienpt91</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3966089%2F0f4dfb81-74d4-4372-8484-74e69985c36b.png</url>
      <title>DEV Community: Kien Phan</title>
      <link>https://dev.to/kienpt91</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/kienpt91"/>
    <language>en</language>
    <item>
      <title>We ran one resume through 20 layouts and a text extractor: here is what got lost</title>
      <dc:creator>Kien Phan</dc:creator>
      <pubDate>Fri, 20 Nov 2026 09:30:00 +0000</pubDate>
      <link>https://dev.to/kienpt91/we-ran-one-resume-through-20-layouts-and-a-text-extractor-here-is-what-got-lost-20o7</link>
      <guid>https://dev.to/kienpt91/we-ran-one-resume-through-20-layouts-and-a-text-extractor-here-is-what-got-lost-20o7</guid>
      <description>&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;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.&lt;/li&gt;
&lt;li&gt;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).&lt;/li&gt;
&lt;li&gt;Columns and positioned blocks kept every word but changed reading order in 4 layouts, and the checker still scores three of them 100. A high checker score does not prove a clean parse.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;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.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Update 2026-11-19:&lt;/strong&gt; after this test we fixed two things in our own checker, so its scores moved. It now counts Word list items, numbered lists and ➤ or dash markers as bullets (pull requests #123 and #118 in our repository). We re-scored the same files: the DOCX baseline, DOCX table layout and DOCX two-column table go from 90 to 100, the DOCX header and footer files from 70 to 80, and the ➤ PDF from 90 to 100. The table shows the 2026-11-17 score and the score now. What each extractor lost did not change, and we re-ran six of the files through the live checker on 2026-11-19 to confirm.&lt;/p&gt;

&lt;h2&gt;
  
  
  What we tested
&lt;/h2&gt;

&lt;p&gt;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.&lt;/p&gt;

&lt;p&gt;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.&lt;/p&gt;

&lt;h2&gt;
  
  
  Results
&lt;/h2&gt;

&lt;p&gt;PDFs were extracted with &lt;code&gt;pdf-parse&lt;/code&gt; 1.1.1 and DOCX files with &lt;code&gt;mammoth&lt;/code&gt; 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.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Layout&lt;/th&gt;
&lt;th&gt;Contact found&lt;/th&gt;
&lt;th&gt;Sections found&lt;/th&gt;
&lt;th&gt;Items found&lt;/th&gt;
&lt;th&gt;Order&lt;/th&gt;
&lt;th&gt;Score 11-17 / now&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;PDF baseline, single column&lt;/td&gt;
&lt;td&gt;all&lt;/td&gt;
&lt;td&gt;4/4&lt;/td&gt;
&lt;td&gt;37/37&lt;/td&gt;
&lt;td&gt;10/10&lt;/td&gt;
&lt;td&gt;100 / 100&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;PDF two-column sidebar&lt;/td&gt;
&lt;td&gt;all&lt;/td&gt;
&lt;td&gt;4/4&lt;/td&gt;
&lt;td&gt;36/37&lt;/td&gt;
&lt;td&gt;8/10&lt;/td&gt;
&lt;td&gt;100 / 100&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;PDF table layout&lt;/td&gt;
&lt;td&gt;all&lt;/td&gt;
&lt;td&gt;4/4&lt;/td&gt;
&lt;td&gt;37/37&lt;/td&gt;
&lt;td&gt;10/10&lt;/td&gt;
&lt;td&gt;100 / 100&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;PDF arrow-glyph bullets (➤)&lt;/td&gt;
&lt;td&gt;all&lt;/td&gt;
&lt;td&gt;4/4&lt;/td&gt;
&lt;td&gt;37/37&lt;/td&gt;
&lt;td&gt;10/10&lt;/td&gt;
&lt;td&gt;90 / 100&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;PDF CSS-drawn bullets&lt;/td&gt;
&lt;td&gt;all&lt;/td&gt;
&lt;td&gt;4/4&lt;/td&gt;
&lt;td&gt;37/37&lt;/td&gt;
&lt;td&gt;7/10&lt;/td&gt;
&lt;td&gt;90 / 90&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;PDF emoji icons&lt;/td&gt;
&lt;td&gt;all&lt;/td&gt;
&lt;td&gt;4/4&lt;/td&gt;
&lt;td&gt;36/37&lt;/td&gt;
&lt;td&gt;10/10&lt;/td&gt;
&lt;td&gt;100 / 100&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;PDF contact as image&lt;/td&gt;
&lt;td&gt;none&lt;/td&gt;
&lt;td&gt;4/4&lt;/td&gt;
&lt;td&gt;34/37&lt;/td&gt;
&lt;td&gt;10/10&lt;/td&gt;
&lt;td&gt;80 / 80&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;PDF letter-spaced headings&lt;/td&gt;
&lt;td&gt;all&lt;/td&gt;
&lt;td&gt;1/4&lt;/td&gt;
&lt;td&gt;37/37&lt;/td&gt;
&lt;td&gt;10/10&lt;/td&gt;
&lt;td&gt;80 / 80&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;PDF creative heading names&lt;/td&gt;
&lt;td&gt;all&lt;/td&gt;
&lt;td&gt;1/4&lt;/td&gt;
&lt;td&gt;37/37&lt;/td&gt;
&lt;td&gt;10/10&lt;/td&gt;
&lt;td&gt;80 / 80&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;PDF ligature serif font&lt;/td&gt;
&lt;td&gt;all&lt;/td&gt;
&lt;td&gt;4/4&lt;/td&gt;
&lt;td&gt;37/37&lt;/td&gt;
&lt;td&gt;10/10&lt;/td&gt;
&lt;td&gt;100 / 100&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;PDF positioned blocks, shuffled order&lt;/td&gt;
&lt;td&gt;all&lt;/td&gt;
&lt;td&gt;4/4&lt;/td&gt;
&lt;td&gt;37/37&lt;/td&gt;
&lt;td&gt;7/10&lt;/td&gt;
&lt;td&gt;100 / 100&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;PDF contact behind link labels&lt;/td&gt;
&lt;td&gt;none&lt;/td&gt;
&lt;td&gt;4/4&lt;/td&gt;
&lt;td&gt;34/37&lt;/td&gt;
&lt;td&gt;10/10&lt;/td&gt;
&lt;td&gt;80 / 80&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;PDF headings as images&lt;/td&gt;
&lt;td&gt;all&lt;/td&gt;
&lt;td&gt;1/4&lt;/td&gt;
&lt;td&gt;37/37&lt;/td&gt;
&lt;td&gt;10/10&lt;/td&gt;
&lt;td&gt;80 / 80&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;PDF full page as image&lt;/td&gt;
&lt;td&gt;none&lt;/td&gt;
&lt;td&gt;0/4&lt;/td&gt;
&lt;td&gt;0/37&lt;/td&gt;
&lt;td&gt;n/a&lt;/td&gt;
&lt;td&gt;30 / 30&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;DOCX baseline (Word list bullets)&lt;/td&gt;
&lt;td&gt;all&lt;/td&gt;
&lt;td&gt;4/4&lt;/td&gt;
&lt;td&gt;37/37&lt;/td&gt;
&lt;td&gt;10/10&lt;/td&gt;
&lt;td&gt;90 / 100&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;DOCX baseline (typed "•" bullets)&lt;/td&gt;
&lt;td&gt;all&lt;/td&gt;
&lt;td&gt;4/4&lt;/td&gt;
&lt;td&gt;37/37&lt;/td&gt;
&lt;td&gt;10/10&lt;/td&gt;
&lt;td&gt;100 / 100&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;DOCX contact in header&lt;/td&gt;
&lt;td&gt;none&lt;/td&gt;
&lt;td&gt;4/4&lt;/td&gt;
&lt;td&gt;33/37&lt;/td&gt;
&lt;td&gt;10/10&lt;/td&gt;
&lt;td&gt;70 / 80&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;DOCX contact in footer&lt;/td&gt;
&lt;td&gt;none&lt;/td&gt;
&lt;td&gt;4/4&lt;/td&gt;
&lt;td&gt;34/37&lt;/td&gt;
&lt;td&gt;10/10&lt;/td&gt;
&lt;td&gt;70 / 80&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;DOCX table layout&lt;/td&gt;
&lt;td&gt;all&lt;/td&gt;
&lt;td&gt;4/4&lt;/td&gt;
&lt;td&gt;37/37&lt;/td&gt;
&lt;td&gt;10/10&lt;/td&gt;
&lt;td&gt;90 / 100&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;DOCX two-column table&lt;/td&gt;
&lt;td&gt;all&lt;/td&gt;
&lt;td&gt;4/4&lt;/td&gt;
&lt;td&gt;37/37&lt;/td&gt;
&lt;td&gt;8/10&lt;/td&gt;
&lt;td&gt;90 / 100&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Raw numbers for both extractors are in the &lt;a href="https://cvposh.com/research/resume-layout-test/results.csv" rel="noopener noreferrer"&gt;results CSV&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Finding 1: contact details are the most fragile field
&lt;/h2&gt;

&lt;p&gt;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.&lt;/p&gt;

&lt;p&gt;In the link-label PDF the address exists only as a link annotation, and neither &lt;code&gt;pdf-parse&lt;/code&gt; nor poppler's &lt;code&gt;pdftotext&lt;/code&gt; returned it. For the DOCX files, &lt;code&gt;mammoth&lt;/code&gt; 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.&lt;/p&gt;

&lt;p&gt;Also: in the two-column PDF the LinkedIn URL wrapped at its hyphen inside the narrow sidebar. &lt;code&gt;pdf-parse&lt;/code&gt; returned &lt;code&gt;jordan-&lt;/code&gt; and &lt;code&gt;sample&lt;/code&gt; on separate lines and &lt;code&gt;pdftotext&lt;/code&gt; dropped the hyphen (&lt;code&gt;jordansample&lt;/code&gt;). The checker still saw a link, but not a working URL.&lt;/p&gt;

&lt;h2&gt;
  
  
  Finding 2: headings need words the parser recognises
&lt;/h2&gt;

&lt;p&gt;The checker looks for experience, education, skills and summary or profile. Three PDFs fell to 1 of 4 sections:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Headings as images.&lt;/strong&gt; The text was not in the text layer.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Creative names&lt;/strong&gt; ("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.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Letter-spaced uppercase headings.&lt;/strong&gt; &lt;code&gt;pdf-parse&lt;/code&gt; returned &lt;code&gt;E X P E R I E N C E&lt;/code&gt;; &lt;code&gt;pdftotext&lt;/code&gt; returned &lt;code&gt;EXPERIENCE&lt;/code&gt;. The same file scored 80 or 100 depending on the extractor.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Each cost 20 points. The bigger risk is any system that groups your history by section: it has to recognise the header first.&lt;/p&gt;

&lt;h2&gt;
  
  
  Finding 3: columns and positioned blocks keep the words, not the order
&lt;/h2&gt;

&lt;p&gt;The two-column sidebar, the DOCX two-column table, the positioned-blocks PDF and the CSS-drawn-bullets PDF kept every word, but &lt;code&gt;pdf-parse&lt;/code&gt; changed the order:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;In both column layouts, skills and education came before summary and experience, because the left column is read first.&lt;/li&gt;
&lt;li&gt;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.&lt;/li&gt;
&lt;li&gt;In the positioned-blocks PDF, which looks like a normal one-page resume, skills came first and the name last.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;code&gt;pdftotext&lt;/code&gt; 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.&lt;/p&gt;

&lt;h2&gt;
  
  
  What did not break anything
&lt;/h2&gt;

&lt;p&gt;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 &lt;code&gt;pdf-parse&lt;/code&gt;. A ➤ bullet was extracted as a character.&lt;/p&gt;

&lt;h2&gt;
  
  
  Our own checker undercounted bullets (fixed after the test)
&lt;/h2&gt;

&lt;p&gt;We found this while testing. The checker counted a bullet only when a line started with &lt;code&gt;•&lt;/code&gt;, &lt;code&gt;-&lt;/code&gt; or &lt;code&gt;*&lt;/code&gt;. A DOCX with real Word list bullets scored 90 and the same file with typed "•" scored 100, because &lt;code&gt;mammoth&lt;/code&gt; does not put list markers in its raw text. The ➤ PDF also scored 90. So the checker deducted 10 points from resumes whose bullets read fine. We fixed both after the test: the same files now count 9 bullets and score 100. Two things remain. The CSS-drawn-dots PDF still scores 90 because its bullets are shapes, not characters, so the checker sees none. And the score can still disagree with a real parse in the other direction: contact details in a Word header or footer still cost 20 points (80 against 100 for the body version), and the checker still has no reading-order rule.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to do with this
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;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.&lt;/li&gt;
&lt;li&gt;Use the section names a reader expects: Summary, Experience, Education, Skills.&lt;/li&gt;
&lt;li&gt;Keep headings as real text, without letter-spacing.&lt;/li&gt;
&lt;li&gt;If you use columns, paste the PDF into a text editor, or run it through &lt;a href="https://cvposh.com/tools/analyzers/ats-checker" rel="noopener noreferrer"&gt;the checker&lt;/a&gt;, and read the order that comes out.&lt;/li&gt;
&lt;li&gt;Keep long URLs out of narrow columns.&lt;/li&gt;
&lt;li&gt;Export a text-based PDF. If you can't select the text, nothing else works.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  FAQ
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Does an ATS reject resumes with columns?&lt;/strong&gt;&lt;br&gt;
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.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Is DOCX safer than PDF?&lt;/strong&gt;&lt;br&gt;
Not in our test. DOCX avoided the ordering problems but was the only format where header and footer content disappeared.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Do icons hurt?&lt;/strong&gt;&lt;br&gt;
Icons next to text did not remove the text. Text drawn as an image did.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Does a high checker score mean my resume will parse correctly?&lt;/strong&gt;&lt;br&gt;
No. The checker starts at 100 and subtracts points when a rule fires, and it has no reading-order rule. Three layouts with changed order still score 100.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Is this representative of real templates?&lt;/strong&gt;&lt;br&gt;
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.&lt;/p&gt;

&lt;h2&gt;
  
  
  Methodology and data
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Date run:&lt;/strong&gt; 2026-11-17. One fictional resume, no real person.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Variants:&lt;/strong&gt; 14 PDFs from Google Chrome 157.0.8103.79 (headless print-to-PDF, Letter) and 6 DOCX files from &lt;code&gt;docx&lt;/code&gt; 8.6.0. Image variants are screenshots of the same HTML.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Extraction:&lt;/strong&gt; &lt;code&gt;pdf-parse&lt;/code&gt; 1.1.1 and &lt;code&gt;mammoth&lt;/code&gt; 1.11.0, as in the checker. PDFs were also run through poppler &lt;code&gt;pdftotext&lt;/code&gt; 26.07.0 (default mode) as a cross-check; both sets of numbers are in the CSV.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Scoring:&lt;/strong&gt; &lt;code&gt;ats-scorer.ts&lt;/code&gt; from the CVPosh repository at &lt;code&gt;origin/main&lt;/code&gt; commit &lt;code&gt;184d4b4&lt;/code&gt;. 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. Update 2026-11-19: the files were re-scored with the current checker (Word list items via &lt;code&gt;parseDocxWithListBullets&lt;/code&gt;, bullet rule of PR #118); numbers are in &lt;code&gt;results-after-bullet-fix.csv&lt;/code&gt; and &lt;code&gt;results-docx-list-bullets.csv&lt;/code&gt; in the same folder, and six files (PDF ➤ bullets, CSS-drawn bullets, DOCX baseline, footer, table layout, two-column table) were run through the live checker, which returned 100, 90, 100, 80, 100 and 100.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Items:&lt;/strong&gt; 37; an item is found when its text appears after whitespace and case normalisation, bullets as complete sentences.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Limits:&lt;/strong&gt; 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.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Reproduce:&lt;/strong&gt; scripts and CSV at &lt;a href="https://cvposh.com/research/resume-layout-test/results.csv" rel="noopener noreferrer"&gt;cvposh.com/research/resume-layout-test&lt;/a&gt;.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;em&gt;CVPosh is our product. The checker is free and needs no sign-up. &lt;a href="https://cvposh.com/tools/how-the-ats-checker-scores" rel="noopener noreferrer"&gt;How its score is calculated&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>career</category>
      <category>resume</category>
      <category>webdev</category>
      <category>testing</category>
    </item>
    <item>
      <title>We checked 100 well-known sites for AI-crawler readiness: llms.txt, robots.txt, schema, sitemaps</title>
      <dc:creator>Kien Phan</dc:creator>
      <pubDate>Thu, 19 Nov 2026 09:30:00 +0000</pubDate>
      <link>https://dev.to/kienpt91/we-checked-100-well-known-sites-for-ai-crawler-readiness-llmstxt-robotstxt-schema-sitemaps-f8g</link>
      <guid>https://dev.to/kienpt91/we-checked-100-well-known-sites-for-ai-crawler-readiness-llmstxt-robotstxt-schema-sitemaps-f8g</guid>
      <description>&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;38 of 100 well-known sites serve an &lt;code&gt;llms.txt&lt;/code&gt; file (42% of the 91 we could check). The split is sharp: 25 of 35 developer-tool sites, 2 of 18 news and media sites.&lt;/li&gt;
&lt;li&gt;24 of the 96 sites whose &lt;code&gt;robots.txt&lt;/code&gt; we could read (25%) block at least one of GPTBot, ClaudeBot and PerplexityBot from &lt;code&gt;/&lt;/code&gt;. Among news and media sites it is 14 of 17 for ClaudeBot. Only 10 of 96 block OAI-SearchBot, the crawler OpenAI documents for ChatGPT search.&lt;/li&gt;
&lt;li&gt;Our own audit tool had two blind spots here: it could not load 27 of the 100 homepages, and it reported an &lt;code&gt;llms.txt&lt;/code&gt; for 5 sites where we found only HTML pages. We report numbers from direct fetches, not from audit scores.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;This is a snapshot of 100 sites we chose by hand on one day. It is not a ranking of the web, and it does not say whether any of these settings changes whether AI products cite a site.&lt;/p&gt;

&lt;h2&gt;
  
  
  What we measured
&lt;/h2&gt;

&lt;p&gt;We wanted a few facts that anyone can re-check with &lt;code&gt;curl&lt;/code&gt;:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Does &lt;code&gt;https://&amp;lt;domain&amp;gt;/llms.txt&lt;/code&gt; return a text file?&lt;/li&gt;
&lt;li&gt;What does &lt;code&gt;robots.txt&lt;/code&gt; say about GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-SearchBot, PerplexityBot, Google-Extended and CCBot for the path &lt;code&gt;/&lt;/code&gt;?&lt;/li&gt;
&lt;li&gt;Does the site declare a sitemap, or serve &lt;code&gt;/sitemap.xml&lt;/code&gt;?&lt;/li&gt;
&lt;li&gt;Does the homepage HTML contain JSON-LD and an &lt;code&gt;og:title&lt;/code&gt;?&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The domains are a fixed list of 100: 35 SaaS and developer tools, 10 AI companies, 18 news, media and publishing sites, 20 consumer and marketplace sites, and 17 education, reference and other sites. The list is in the data folder. We also sent every domain through SEOBolt's free audit endpoint, then compared its output with the direct fetches. SEOBolt is our product; the comparison is there to show where it is wrong.&lt;/p&gt;

&lt;h2&gt;
  
  
  Results
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Check&lt;/th&gt;
&lt;th&gt;Result&lt;/th&gt;
&lt;th&gt;Out of&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Serves a text &lt;code&gt;llms.txt&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;38 (38%)&lt;/td&gt;
&lt;td&gt;100; 9 could not be checked&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Declares a &lt;code&gt;Sitemap:&lt;/code&gt; in robots.txt&lt;/td&gt;
&lt;td&gt;79 (82%)&lt;/td&gt;
&lt;td&gt;96 readable robots.txt&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Declares a sitemap or serves &lt;code&gt;/sitemap.xml&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;85 (85%)&lt;/td&gt;
&lt;td&gt;100&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Homepage has JSON-LD&lt;/td&gt;
&lt;td&gt;51 (61%)&lt;/td&gt;
&lt;td&gt;83 homepages we could fetch&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Homepage has &lt;code&gt;og:title&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;73 (88%)&lt;/td&gt;
&lt;td&gt;83 homepages we could fetch&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Blocking means a &lt;code&gt;Disallow&lt;/code&gt; rule wins for the path &lt;code&gt;/&lt;/code&gt; in the group that applies to that crawler (its own group, otherwise &lt;code&gt;*&lt;/code&gt;):&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Crawler&lt;/th&gt;
&lt;th&gt;Blocks &lt;code&gt;/&lt;/code&gt;
&lt;/th&gt;
&lt;th&gt;Of 96&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;GPTBot&lt;/td&gt;
&lt;td&gt;20 (21%)&lt;/td&gt;
&lt;td&gt;96&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;ClaudeBot&lt;/td&gt;
&lt;td&gt;24 (25%)&lt;/td&gt;
&lt;td&gt;96&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;PerplexityBot&lt;/td&gt;
&lt;td&gt;18 (19%)&lt;/td&gt;
&lt;td&gt;96&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;OAI-SearchBot&lt;/td&gt;
&lt;td&gt;10 (10%)&lt;/td&gt;
&lt;td&gt;96&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;ChatGPT-User&lt;/td&gt;
&lt;td&gt;11 (11%)&lt;/td&gt;
&lt;td&gt;96&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Claude-SearchBot&lt;/td&gt;
&lt;td&gt;15 (16%)&lt;/td&gt;
&lt;td&gt;96&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Google-Extended&lt;/td&gt;
&lt;td&gt;13 (14%)&lt;/td&gt;
&lt;td&gt;96&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;CCBot&lt;/td&gt;
&lt;td&gt;24 (25%)&lt;/td&gt;
&lt;td&gt;96&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;By category (blocking counts are out of the sites with a readable robots.txt):&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Category&lt;/th&gt;
&lt;th&gt;llms.txt&lt;/th&gt;
&lt;th&gt;GPTBot blocked&lt;/th&gt;
&lt;th&gt;ClaudeBot blocked&lt;/th&gt;
&lt;th&gt;PerplexityBot blocked&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;SaaS and developer tools&lt;/td&gt;
&lt;td&gt;25 of 35&lt;/td&gt;
&lt;td&gt;2 of 35&lt;/td&gt;
&lt;td&gt;2 of 35&lt;/td&gt;
&lt;td&gt;1 of 35&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;AI companies&lt;/td&gt;
&lt;td&gt;4 of 10&lt;/td&gt;
&lt;td&gt;0 of 9&lt;/td&gt;
&lt;td&gt;0 of 9&lt;/td&gt;
&lt;td&gt;0 of 9&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;News, media, publishing&lt;/td&gt;
&lt;td&gt;2 of 18&lt;/td&gt;
&lt;td&gt;10 of 17&lt;/td&gt;
&lt;td&gt;14 of 17&lt;/td&gt;
&lt;td&gt;12 of 17&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Consumer and marketplaces&lt;/td&gt;
&lt;td&gt;3 of 20&lt;/td&gt;
&lt;td&gt;6 of 19&lt;/td&gt;
&lt;td&gt;6 of 19&lt;/td&gt;
&lt;td&gt;3 of 19&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Education, reference, other&lt;/td&gt;
&lt;td&gt;4 of 17&lt;/td&gt;
&lt;td&gt;2 of 16&lt;/td&gt;
&lt;td&gt;2 of 16&lt;/td&gt;
&lt;td&gt;2 of 16&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;h2&gt;
  
  
  llms.txt: common among developer tools, rare elsewhere
&lt;/h2&gt;

&lt;p&gt;&lt;code&gt;llms.txt&lt;/code&gt; is a proposal published at &lt;a href="https://llmstxt.org" rel="noopener noreferrer"&gt;llmstxt.org&lt;/a&gt; by Jeremy Howard in 2024. It is a Markdown file at the site root that summarises a site for language models. We count presence only. We did not check the format or content, and we make no claim that having one changes how AI products treat a site.&lt;/p&gt;

&lt;p&gt;Presence tracks category more than anything else. 25 of 35 developer-tool domains serve one, and so do 4 of 10 AI companies, against 2 of 18 news sites and 3 of 20 consumer sites. Nine domains answered our requests with 401, 403, 405 or no response, so we cannot say either way. File sizes among the 38 ranged from about 2.8 KB to about 2.5 MB (twilio.com, 2,523,347 bytes). Correction, 2026-11-18: this post first said "about 1 MB" for the largest file; the results CSV shows 2.5 MB.&lt;/p&gt;

&lt;p&gt;Six domains returned HTTP 200 for &lt;code&gt;/llms.txt&lt;/code&gt; with an HTML page (a redirect to another page, a branded not-found page or a bot challenge). We did not count those. This matters for audit tools, as the section below shows.&lt;/p&gt;

&lt;h2&gt;
  
  
  robots.txt: training crawlers and search crawlers are different decisions
&lt;/h2&gt;

&lt;p&gt;The vendors document these crawlers separately. OpenAI says &lt;a href="https://platform.openai.com/docs/bots" rel="noopener noreferrer"&gt;OAI-SearchBot and GPTBot settings are independent&lt;/a&gt;: a site can allow OAI-SearchBot to appear in search results while disallowing GPTBot, which OpenAI describes as crawling content that may be used to train its foundation models. Anthropic's &lt;a href="https://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler" rel="noopener noreferrer"&gt;help article&lt;/a&gt; describes ClaudeBot as collecting web content that could contribute to training, and &lt;a href="https://docs.perplexity.ai/guides/bots" rel="noopener noreferrer"&gt;Perplexity's documentation&lt;/a&gt; describes PerplexityBot as surfacing sites in Perplexity search results and says it is not used to crawl for foundation models.&lt;/p&gt;

&lt;p&gt;That is why the gap matters. 20 sites block GPTBot but only 10 block OAI-SearchBot. Of the 96 readable robots.txt files, 24 block at least one of GPTBot, ClaudeBot or PerplexityBot and 14 block all three. Most name the crawler in its own group: 18 of the 20 GPTBot blocks, 22 of 24 for ClaudeBot and 16 of 18 for PerplexityBot. The other 2 in each case come only from a &lt;code&gt;User-agent: *&lt;/code&gt; rule with &lt;code&gt;Disallow: /&lt;/code&gt;. We did not examine whether any block is deliberate.&lt;/p&gt;

&lt;p&gt;One caution: robots.txt is a request. &lt;a href="https://www.rfc-editor.org/rfc/rfc9309" rel="noopener noreferrer"&gt;RFC 9309&lt;/a&gt; states that its rules "are not a form of access authorization." A &lt;code&gt;Disallow&lt;/code&gt; shows a site's stated preference, not that crawling is prevented.&lt;/p&gt;

&lt;h2&gt;
  
  
  Sitemaps, JSON-LD and Open Graph
&lt;/h2&gt;

&lt;p&gt;85 sites declared a sitemap in robots.txt or served one at &lt;code&gt;/sitemap.xml&lt;/code&gt;. Only 52 served a valid sitemap at that exact path, so a root-path check alone would under-count sites that use another location. JSON-LD appeared in the raw homepage HTML of 51 of 83 fetchable homepages; client-side-only injection would not show up in this method. Open Graph titles were the most common of the four signals at 73 of 83.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where our own audit disagreed with the direct fetches
&lt;/h2&gt;

&lt;p&gt;We called &lt;code&gt;POST /api/free-audit&lt;/code&gt; once per domain. Every call returned HTTP 200, and scores ranged from 44 to 100 (median 93). Comparing it with the direct fetches turned up two problems:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;It could not read 27 of the 100 homepages.&lt;/strong&gt; For those it reports "couldn't fetch the homepage" and skips structured-data and Open Graph checks. These are mostly large sites behind bot protection. 21 audits also reported &lt;code&gt;robots.txt&lt;/code&gt; or &lt;code&gt;llms.txt&lt;/code&gt; as unverifiable.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;It reported an &lt;code&gt;llms.txt&lt;/code&gt; for 5 sites where we found none.&lt;/strong&gt; For sendgrid.com, trello.com, booking.com, spotify.com and netflix.com the audit said an &lt;code&gt;llms.txt&lt;/code&gt; exists but "needs more context". Our direct fetch returned an HTML page for each (HTTP 200 for four of them, a 404 page for booking.com). A check that trusts the HTTP status gets this wrong. It is a flaw in our audit, so until it is fixed treat an &lt;code&gt;llms.txt&lt;/code&gt; audit result as unconfirmed until you open the URL.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;On ChatGPT-search crawler access the audit flagged 9 sites as blocked; the direct check found 11 for OAI-SearchBot or ChatGPT-User, and the two lists differ for 4 domains. That is why every figure in this post comes from the direct fetches.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to do with this
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;Open &lt;code&gt;https://yourdomain/robots.txt&lt;/code&gt; and read it. Decide separately about training crawlers (GPTBot, ClaudeBot) and search or user-fetch crawlers (OAI-SearchBot, PerplexityBot, Claude-SearchBot).&lt;/li&gt;
&lt;li&gt;Declare your sitemap in robots.txt.&lt;/li&gt;
&lt;li&gt;If you add an &lt;code&gt;llms.txt&lt;/code&gt;, check with &lt;code&gt;curl -i&lt;/code&gt; that it returns a text file and not your homepage.&lt;/li&gt;
&lt;li&gt;Check JSON-LD in the raw HTML (&lt;code&gt;curl -s https://yourdomain | grep ld+json&lt;/code&gt;), not in the browser inspector.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  FAQ
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Does blocking GPTBot keep me out of ChatGPT answers?&lt;/strong&gt;&lt;br&gt;
OpenAI documents GPTBot and OAI-SearchBot as independent settings and ties GPTBot to training. We did not test what happens to citations either way.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Does an llms.txt file help me get cited?&lt;/strong&gt;&lt;br&gt;
We don't know and did not test it. This study counts how many sites have one.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Is the sample representative of the web?&lt;/strong&gt;&lt;br&gt;
No. It is 100 well-known domains chosen by hand, skewed toward tech companies. Percentages describe this list only.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Why are some checks "unverifiable"?&lt;/strong&gt;&lt;br&gt;
Some sites return 401, 403 or 405 to automated requests. We retried once with a desktop browser user agent and, where it still failed, marked the check unverifiable instead of guessing.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Can I re-run this?&lt;/strong&gt;&lt;br&gt;
Yes. The domain list, the audit and direct-fetch scripts and the CSV are published with this post.&lt;/p&gt;

&lt;h2&gt;
  
  
  Methodology and data
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Date:&lt;/strong&gt; 2026-11-17 (UTC). Audit requests ran from 09:38 to 11:20; direct fetches and their retry pass ran between 09:39 and 09:46.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Domains:&lt;/strong&gt; 100, fixed and hand-picked before running anything; none excluded afterwards.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Audit:&lt;/strong&gt; &lt;code&gt;POST https://seobolt.app/api/free-audit&lt;/code&gt; with only &lt;code&gt;{"domain": ...}&lt;/code&gt;, 5 requests then a 305-second pause to stay inside the documented limit of 5 per 5 minutes per IP. No request returned 429. This is the scoring version live after PR #167 ("honest audit scoring"); the audit code was unchanged between the live revisions &lt;code&gt;522a570&lt;/code&gt; and &lt;code&gt;1bdd25a&lt;/code&gt; seen during the run.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Direct fetches:&lt;/strong&gt; GET of &lt;code&gt;/&lt;/code&gt;, &lt;code&gt;/robots.txt&lt;/code&gt;, &lt;code&gt;/llms.txt&lt;/code&gt;, &lt;code&gt;/sitemap.xml&lt;/code&gt;, first as &lt;code&gt;seobolt-research/1.0&lt;/code&gt;, then retried once with a desktop Chrome user agent for any answer other than 200, 404 or 410. 7 domains changed as a result.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Definitions:&lt;/strong&gt; a valid &lt;code&gt;llms.txt&lt;/code&gt; is a 200 non-HTML non-empty response. Robots rules are evaluated with our own RFC 9309 implementation (specific group over &lt;code&gt;*&lt;/code&gt;, longest match, Allow wins ties) for the path &lt;code&gt;/&lt;/code&gt; only; partial blocks of sub-paths are not counted. We compared the results with the raw robots.txt of four sites by hand.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Limits:&lt;/strong&gt; one day, one network location, hand-picked sample, JSON-LD and Open Graph read from raw HTML only, sitemaps checked at the root path only, 4 domains had no readable robots.txt.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Data:&lt;/strong&gt; &lt;a href="https://seobolt.app/research/ai-search-readiness-100-sites/results.csv" rel="noopener noreferrer"&gt;results CSV, domain list and scripts&lt;/a&gt;.&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>seo</category>
      <category>ai</category>
      <category>webdev</category>
      <category>llm</category>
    </item>
    <item>
      <title>Why email domain verification fails, and how to make it work</title>
      <dc:creator>Kien Phan</dc:creator>
      <pubDate>Wed, 18 Nov 2026 03:26:56 +0000</pubDate>
      <link>https://dev.to/kienpt91/why-email-domain-verification-fails-and-how-to-make-it-work-3ioa</link>
      <guid>https://dev.to/kienpt91/why-email-domain-verification-fails-and-how-to-make-it-work-3ioa</guid>
      <description>&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Of 23 customer domains added to AISend (accounts not ours, not banned), 4 verified and 19 are still pending. The 4 verified in a median of about 8 minutes; 17 of the 19 pending ones were added more than 30 days ago. People either finish quickly or never.&lt;/li&gt;
&lt;li&gt;A public-DNS check of the 19 pending domains on 2026-11-17 found that 11 do not exist (NXDOMAIN at the registrable domain). Of the other 8, six had no DKIM records published at all, one was behind a Cloudflare proxy and one is unclear; five already had an SPF record that a second one would break.&lt;/li&gt;
&lt;li&gt;The fixes are mechanical: publish the DKIM CNAMEs exactly as given and DNS-only, keep one SPF record per name, and verify a subdomain such as &lt;code&gt;mail.acme.com&lt;/code&gt; rather than your root domain.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Why I counted
&lt;/h2&gt;

&lt;p&gt;I run &lt;a href="https://aisend.app" rel="noopener noreferrer"&gt;AISend&lt;/a&gt;, a Resend-compatible email API that sends through Amazon SES. Domain verification is where new users get stuck, so I counted how and where. Domain names below are placeholders (&lt;code&gt;acme.com&lt;/code&gt;); no customer domain appears in this article.&lt;/p&gt;

&lt;h2&gt;
  
  
  What "verified" actually checks
&lt;/h2&gt;

&lt;p&gt;With Amazon SES, the thing that proves you own a domain is DKIM. AWS describes DKIM as the method SES uses to verify domain ownership, and a verified domain can send from any subdomain or address under it (&lt;a href="https://docs.aws.amazon.com/ses/latest/dg/creating-identities.html" rel="noopener noreferrer"&gt;SES identities&lt;/a&gt;). When you add a domain, SES issues three DKIM tokens and you publish them as three CNAME records. Only those gate verification in AISend. SPF and the return-path record are optional, and DMARC is recommended afterwards.&lt;/p&gt;

&lt;p&gt;Resend works the same way at the dashboard level: add a domain, publish the records, wait. Resend says a domain often verifies within 15 minutes and sometimes takes up to 72 hours (&lt;a href="https://resend.com/docs/add-a-domain" rel="noopener noreferrer"&gt;Resend: add a domain&lt;/a&gt;). AWS gives the same 72-hour ceiling for DKIM (&lt;a href="https://docs.aws.amazon.com/ses/latest/dg/troubleshoot-dkim.html" rel="noopener noreferrer"&gt;SES DKIM troubleshooting&lt;/a&gt;).&lt;/p&gt;

&lt;h2&gt;
  
  
  The records, with placeholders
&lt;/h2&gt;

&lt;p&gt;For &lt;code&gt;mail.acme.com&lt;/code&gt; the set looks like this. The &lt;code&gt;&amp;lt;token&amp;gt;&lt;/code&gt; values come from your dashboard; never invent them.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Type   Host (full name)                        Value
CNAME  &amp;lt;token1&amp;gt;._domainkey.mail.acme.com       &amp;lt;token1&amp;gt;.dkim.amazonses.com
CNAME  &amp;lt;token2&amp;gt;._domainkey.mail.acme.com       &amp;lt;token2&amp;gt;.dkim.amazonses.com
CNAME  &amp;lt;token3&amp;gt;._domainkey.mail.acme.com       &amp;lt;token3&amp;gt;.dkim.amazonses.com
TXT    mail.acme.com                           v=spf1 include:amazonses.com ~all     (only if mail.acme.com has no SPF yet)
TXT    _dmarc.mail.acme.com                    v=DMARC1; p=none; rua=mailto:dmarc@mail.acme.com
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Most DNS providers append your zone, so in the "Name" box you type &lt;code&gt;&amp;lt;token1&amp;gt;._domainkey.mail&lt;/code&gt;, not the full name.&lt;/p&gt;

&lt;h2&gt;
  
  
  Decision table: what you see, what it means
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;What you see&lt;/th&gt;
&lt;th&gt;Likely cause&lt;/th&gt;
&lt;th&gt;How to check&lt;/th&gt;
&lt;th&gt;Fix&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Pending for hours, records "added"&lt;/td&gt;
&lt;td&gt;Host typed with the zone twice: &lt;code&gt;&amp;lt;token&amp;gt;._domainkey.mail.acme.com.acme.com&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;dig +short CNAME &amp;lt;token1&amp;gt;._domainkey.mail.acme.com&lt;/code&gt; returns nothing&lt;/td&gt;
&lt;td&gt;Type only &lt;code&gt;&amp;lt;token1&amp;gt;._domainkey.mail&lt;/code&gt;
&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Pending, DNS shows an IP at the DKIM name&lt;/td&gt;
&lt;td&gt;Cloudflare proxy (orange cloud) hides the CNAME&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;dig +short A &amp;lt;token1&amp;gt;._domainkey.mail.acme.com&lt;/code&gt; returns Cloudflare addresses&lt;/td&gt;
&lt;td&gt;Set each CNAME to DNS only&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Pending, CNAME exists&lt;/td&gt;
&lt;td&gt;Wrong value or a stale record&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;dig +short CNAME&lt;/code&gt; shows a different target&lt;/td&gt;
&lt;td&gt;Replace with the exact value&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Nothing resolves anywhere&lt;/td&gt;
&lt;td&gt;Domain not registered, typo, or expired&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;dig NS acme.com&lt;/code&gt; returns NXDOMAIN&lt;/td&gt;
&lt;td&gt;Fix the name or register the domain&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;No DKIM records found, zone is live&lt;/td&gt;
&lt;td&gt;Records added at a DNS host the domain does not use&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;dig NS acme.com&lt;/code&gt; differs from where you edited&lt;/td&gt;
&lt;td&gt;Add them at the provider the nameservers point to&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Verified, but old mail starts failing SPF&lt;/td&gt;
&lt;td&gt;A second SPF record at the same name&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;dig +short TXT acme.com&lt;/code&gt; shows two &lt;code&gt;v=spf1&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;Merge into one record&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;"Already added by another account"&lt;/td&gt;
&lt;td&gt;First account to add the name holds it&lt;/td&gt;
&lt;td&gt;Error from the API or dashboard&lt;/td&gt;
&lt;td&gt;Contact support&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;h2&gt;
  
  
  Cloudflare: turn the proxy off
&lt;/h2&gt;

&lt;p&gt;Cloudflare can only proxy A, AAAA and CNAME records, and its docs say CNAMEs that prove domain ownership should not be proxied (&lt;a href="https://developers.cloudflare.com/dns/proxy-status/" rel="noopener noreferrer"&gt;Cloudflare proxy status&lt;/a&gt;). Resend gives the same instruction for its CNAMEs (&lt;a href="https://resend.com/docs/add-a-domain" rel="noopener noreferrer"&gt;Resend docs&lt;/a&gt;). A proxied CNAME answers a lookup with Cloudflare IP addresses, so the verifier cannot see the target. The fix is one click per record: grey cloud, "DNS only".&lt;/p&gt;

&lt;h2&gt;
  
  
  SPF at the root: the collision
&lt;/h2&gt;

&lt;p&gt;AISend lists an SPF record at the domain itself: &lt;code&gt;v=spf1 include:amazonses.com ~all&lt;/code&gt;. If you verify &lt;code&gt;acme.com&lt;/code&gt; and you already send mail through, say, Google Workspace, you already have an SPF record there. Adding this one next to it produces two. RFC 7208 section 4.5 says a receiver that finds more than one SPF record returns &lt;code&gt;permerror&lt;/code&gt;, so SPF fails for all mail from that name, including your existing mail (&lt;a href="https://datatracker.ietf.org/doc/html/rfc7208#section-4.5" rel="noopener noreferrer"&gt;RFC 7208&lt;/a&gt;).&lt;/p&gt;

&lt;p&gt;The merge is a single record:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight diff"&gt;&lt;code&gt;Before:  v=spf1 include:_spf.google.com ~all
After:   v=spf1 include:_spf.google.com include:amazonses.com ~all
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h2&gt;
  
  
  Why the collision is less scary than it sounds
&lt;/h2&gt;

&lt;p&gt;Two things. First, SES does not need your SPF record for verification, and by default SES uses an &lt;code&gt;amazonses.com&lt;/code&gt; subdomain as the MAIL FROM, which passes SPF without you (&lt;a href="https://docs.aws.amazon.com/ses/latest/dg/send-email-authentication-spf.html" rel="noopener noreferrer"&gt;SES SPF&lt;/a&gt;). Second, a subdomain avoids the question entirely, because &lt;code&gt;mail.acme.com&lt;/code&gt; has no SPF record to collide with.&lt;/p&gt;

&lt;p&gt;AISend's Verify screen reports each record as found, wrong value, proxied, wrong host, duplicate or merge needed, and says which one to fix.&lt;/p&gt;

&lt;h2&gt;
  
  
  Start on a subdomain
&lt;/h2&gt;

&lt;p&gt;Resend recommends sending from a subdomain (&lt;code&gt;updates.acme.com&lt;/code&gt;) to isolate sending reputation and signal intent (&lt;a href="https://resend.com/docs/dashboard/domains/introduction" rel="noopener noreferrer"&gt;Resend: verified domains&lt;/a&gt;). The same reasoning applies to any SES-based sender. A subdomain also leaves the root domain's existing mail and SPF untouched. The exception is a root domain whose reputation you want to keep: then keep it, and warm up as your provider advises.&lt;/p&gt;

&lt;h2&gt;
  
  
  Our numbers: the cohort
&lt;/h2&gt;

&lt;p&gt;Snapshot: 2026-11-17, read-only counts from AISend's database. Verified domains took a median of 7.8 minutes from being added (range 1.2 minutes to 1.5 hours; n = 4, so treat it as an anecdote). Of the 19 pending, 2 were added in the last 7 days and 17 more than 30 days ago; none sit in between. Apex domains verified 3 of 16 times and subdomains 1 of 7. That sample is too small to say subdomains are better.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Step&lt;/th&gt;
&lt;th&gt;Domains&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Rows in the domains table (33 accounts)&lt;/td&gt;
&lt;td&gt;55&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Our own product domains&lt;/td&gt;
&lt;td&gt;-17&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Founder or internal test accounts&lt;/td&gt;
&lt;td&gt;-8&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Banned accounts&lt;/td&gt;
&lt;td&gt;-7&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Customer cohort&lt;/td&gt;
&lt;td&gt;23&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Verified&lt;/td&gt;
&lt;td&gt;4 (17%)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Pending&lt;/td&gt;
&lt;td&gt;19 (83%)&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;h2&gt;
  
  
  Our numbers: what public DNS shows
&lt;/h2&gt;

&lt;p&gt;I looked up the 19 pending names in public DNS (1.1.1.1 and 8.8.8.8) on the same day. Of the 8 domains that do exist, 5 already have an SPF record at the name, so an unmerged second record would have broken their mail. I did not expect the biggest bucket: more than half of the stuck domains do not exist in DNS. Typos and unregistered names are the main failure here, ahead of the DNS subtleties above. AISend rejects malformed names, reserved names and free-subdomain hosts when you add a domain, but it does not currently check that a well-formed name is actually registered, which is why these rows exist.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Finding&lt;/th&gt;
&lt;th&gt;Domains&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Registrable domain returns NXDOMAIN (does not exist in DNS)&lt;/td&gt;
&lt;td&gt;11&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Zone exists, no DKIM CNAMEs published&lt;/td&gt;
&lt;td&gt;6&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Zone exists, DKIM names answer with Cloudflare addresses (proxied)&lt;/td&gt;
&lt;td&gt;1&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Zone exists, an address record answers instead of the CNAME&lt;/td&gt;
&lt;td&gt;1&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Doubled-host records (&lt;code&gt;...acme.com.acme.com&lt;/code&gt;)&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;h2&gt;
  
  
  Doing it from an agent
&lt;/h2&gt;

&lt;p&gt;AISend's MCP server exposes &lt;code&gt;add_domain&lt;/code&gt; and &lt;code&gt;verify_domain&lt;/code&gt;. &lt;code&gt;add_domain&lt;/code&gt; returns the records to publish and suggests a subdomain; you publish them yourself, since the tool cannot change DNS. &lt;code&gt;verify_domain&lt;/code&gt; returns each record's live state with a plain-language hint, such as a Cloudflare proxy or a second SPF record. The sequence is add, publish, verify, then send.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently asked questions
&lt;/h2&gt;

&lt;h3&gt;
  
  
  How long does domain verification take?
&lt;/h3&gt;

&lt;p&gt;Usually minutes once the CNAME records are visible in public DNS. Our four verified customer domains took a median of about 8 minutes. Resend and AWS both say it can take up to 72 hours, so wait that long before assuming a fault.&lt;/p&gt;

&lt;h3&gt;
  
  
  Why does my domain stay pending after I added the records?
&lt;/h3&gt;

&lt;p&gt;Check three things in order: the domain exists (&lt;code&gt;dig NS acme.com&lt;/code&gt;), the records are at the DNS provider your nameservers point to, and the DKIM CNAMEs are not proxied and not typed with the zone twice. In our pending set, 11 of 19 domains did not exist in DNS.&lt;/p&gt;

&lt;h3&gt;
  
  
  Can I have two SPF records on one domain?
&lt;/h3&gt;

&lt;p&gt;No. Receivers treat more than one &lt;code&gt;v=spf1&lt;/code&gt; record on a name as a permanent error, and SPF fails for all mail from it. Merge the includes into a single record.&lt;/p&gt;

&lt;h3&gt;
  
  
  Should I verify my root domain or a subdomain?
&lt;/h3&gt;

&lt;p&gt;A subdomain such as &lt;code&gt;mail.acme.com&lt;/code&gt; is the safer default: it keeps sending reputation separate and avoids touching the root domain's SPF. Keep the root domain only if its reputation is something you need to preserve.&lt;/p&gt;

&lt;h3&gt;
  
  
  Does Cloudflare break domain verification?
&lt;/h3&gt;

&lt;p&gt;Only if the DKIM CNAMEs are proxied. Set them to DNS only. Cloudflare says records that prove domain ownership should not be proxied.&lt;/p&gt;

&lt;h2&gt;
  
  
  Methodology and data
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Read-only snapshot of AISend's production database, 2026-11-17 around 09:40 UTC. Counts only; no customer domain is shown.&lt;/li&gt;
&lt;li&gt;Cohort: domains added by accounts that are not ours. Excluded: 17 domains on our own products, 8 from founder or internal test accounts, 7 from banned accounts.&lt;/li&gt;
&lt;li&gt;"Verified" means the domain's status is &lt;code&gt;verified&lt;/code&gt;. Time to verify is &lt;code&gt;verified_at&lt;/code&gt; minus &lt;code&gt;created_at&lt;/code&gt;, which is the first successful check, not the moment DNS propagated.&lt;/li&gt;
&lt;li&gt;DNS findings come from public lookups on the same day and are a point-in-time view; a domain fixed since would show differently.&lt;/li&gt;
&lt;li&gt;Queries and the DNS script are published at &lt;a href="https://aisend.app/data/domain-verification-queries.md" rel="noopener noreferrer"&gt;aisend.app/data/domain-verification-queries.md&lt;/a&gt;.&lt;/li&gt;
&lt;li&gt;Stated plainly: n = 23, one product, one day. Not an industry statistic.&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>email</category>
      <category>dns</category>
      <category>devops</category>
      <category>tutorial</category>
    </item>
    <item>
      <title>89,499 production errors from 16 apps: what our own error tracker showed us</title>
      <dc:creator>Kien Phan</dc:creator>
      <pubDate>Wed, 18 Nov 2026 03:20:33 +0000</pubDate>
      <link>https://dev.to/kienpt91/89499-production-errors-from-16-apps-what-our-own-error-tracker-showed-us-409f</link>
      <guid>https://dev.to/kienpt91/89499-production-errors-from-16-apps-what-our-own-error-tracker-showed-us-409f</guid>
      <description>&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Between 2026-03-13 and 2026-11-17, 16 of our own apps sent 89,499 error occurrences to Bugsly. They grouped into 1,195 issues. Twelve issues (1%) made up 78.7% of the volume; 465 issues (38.9%) happened exactly once.&lt;/li&gt;
&lt;li&gt;The biggest categories were upstream HTTP 4xx/5xx errors (32.5% of volume), plain code bugs such as type and null errors (21.4%) and ClickHouse limits (15.6%). Postgres connection and pool errors were 0.8%; Redis was 0.2%.&lt;/li&gt;
&lt;li&gt;I read the 12 largest issues by hand. Two (19.4% of all volume) needed only a code change. Six (48.3%) needed config, infrastructure or a migration that a human applies. Four (11.0%) carried no usable signal.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;We build &lt;a href="https://bugsly.dev" rel="noopener noreferrer"&gt;Bugsly&lt;/a&gt;, a hosted error tracker that accepts Sentry SDKs, and run it on our own products. This is what that data says; the queries to reproduce every number are published at &lt;a href="https://bugsly.dev/data/production-errors-16-apps-queries.md" rel="noopener noreferrer"&gt;bugsly.dev/data/production-errors-16-apps-queries.md&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;One caveat first. 56,302 of the 89,499 occurrences (62.9%) came from a single project: Bugsly's own Python backend (the project is still called Watchdog). Read this as what happens in a small portfolio of products, not as an industry benchmark.&lt;/p&gt;

&lt;h2&gt;
  
  
  What we counted
&lt;/h2&gt;

&lt;p&gt;The scope is one organization (ours) with 17 projects; 16 reported at least one error: 11 Next.js, 2 Node, 2 Ruby and 1 Python. The Python project is the 62.9% outlier above; Next.js apps contributed 27.1% (24,223) of the occurrences across 955 issues.&lt;/p&gt;

&lt;p&gt;Totals use each issue's occurrence counter, which has run since 2026-03-13. The raw events in ClickHouse expire after 90 days, so they are used only for the latency numbers further down. All other organizations are excluded, including one external tenant with a much larger volume.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where the volume went
&lt;/h2&gt;

&lt;p&gt;I bucketed every issue by pattern-matching its title. The buckets are heuristic; the rules are in the published queries.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Category&lt;/th&gt;
&lt;th&gt;Issues&lt;/th&gt;
&lt;th&gt;Occurrences&lt;/th&gt;
&lt;th&gt;Share&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;HTTP 4xx/5xx from an upstream API&lt;/td&gt;
&lt;td&gt;93&lt;/td&gt;
&lt;td&gt;29,049&lt;/td&gt;
&lt;td&gt;32.5%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Plain code bugs (type, null, import)&lt;/td&gt;
&lt;td&gt;90&lt;/td&gt;
&lt;td&gt;19,189&lt;/td&gt;
&lt;td&gt;21.4%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;ClickHouse limits and connectivity&lt;/td&gt;
&lt;td&gt;32&lt;/td&gt;
&lt;td&gt;13,953&lt;/td&gt;
&lt;td&gt;15.6%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Network failures (fetch, DNS, socket)&lt;/td&gt;
&lt;td&gt;37&lt;/td&gt;
&lt;td&gt;9,091&lt;/td&gt;
&lt;td&gt;10.2%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Schema drift (missing column or table)&lt;/td&gt;
&lt;td&gt;20&lt;/td&gt;
&lt;td&gt;4,597&lt;/td&gt;
&lt;td&gt;5.1%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Framework, proxy and ORM misc&lt;/td&gt;
&lt;td&gt;90&lt;/td&gt;
&lt;td&gt;3,518&lt;/td&gt;
&lt;td&gt;3.9%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Auth, tokens, unverified identities&lt;/td&gt;
&lt;td&gt;55&lt;/td&gt;
&lt;td&gt;3,169&lt;/td&gt;
&lt;td&gt;3.5%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Quota, credits, rate limits&lt;/td&gt;
&lt;td&gt;63&lt;/td&gt;
&lt;td&gt;2,349&lt;/td&gt;
&lt;td&gt;2.6%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;React hydration mismatch&lt;/td&gt;
&lt;td&gt;132&lt;/td&gt;
&lt;td&gt;1,060&lt;/td&gt;
&lt;td&gt;1.2%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Postgres connection or pool&lt;/td&gt;
&lt;td&gt;48&lt;/td&gt;
&lt;td&gt;673&lt;/td&gt;
&lt;td&gt;0.8%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Redis&lt;/td&gt;
&lt;td&gt;8&lt;/td&gt;
&lt;td&gt;211&lt;/td&gt;
&lt;td&gt;0.2%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Other (long tail)&lt;/td&gt;
&lt;td&gt;527&lt;/td&gt;
&lt;td&gt;2,640&lt;/td&gt;
&lt;td&gt;2.9%&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;h3&gt;
  
  
  Volume is concentrated, and a lot of issues are one-offs
&lt;/h3&gt;

&lt;p&gt;The top issue is 24.5% of all occurrences; the top 3 are 50.7%, the top 10 are 75.9%. At the other end, 465 issues were seen once: 0.5% of the volume but 38.9% of the list you would have to triage.&lt;/p&gt;

&lt;h3&gt;
  
  
  The biggest source was one of our own services refusing another
&lt;/h3&gt;

&lt;p&gt;Twenty-seven issues name our own email API in the title. They total 23,725 occurrences, 26.5% of everything. By volume, nearly all of it is the backend's billing and notification tasks getting a 403 from that API. The issue data does not record why, so I will not guess. What it does show is an uncomfortable pattern: the alerting path itself was the loudest failure.&lt;/p&gt;

&lt;h3&gt;
  
  
  The classic suspects were rare
&lt;/h3&gt;

&lt;p&gt;Postgres connection and pool issues: 48 issues, 673 occurrences. Redis: 8 issues, 211. Issues whose title names OpenAI: 8 issues, 81 occurrences. Any AI or media provider (OpenAI, Gemini, Veo, TTS and similar): 55 issues, 5,474 occurrences (6.1%).&lt;/p&gt;

&lt;p&gt;Grouping was the bigger problem. React hydration errors produced 132 separate issues for one class of bug, because the minified message carries variable arguments. Normalizing digits and cutting titles to 70 characters collapses all 1,195 issues into 556 title families. Fix grouping before you automate anything.&lt;/p&gt;

&lt;h2&gt;
  
  
  Do issues come back?
&lt;/h2&gt;

&lt;p&gt;Of 1,195 issues, 1,069 are resolved, 59 (4.9%) are regressions (marked resolved, then seen again), 44 are ignored and 23 are unresolved. The regressions account for 13.6% of occurrences. "Resolved" here means someone clicked resolve; we do not record whether a code change shipped, so this is not a fix rate.&lt;/p&gt;

&lt;h2&gt;
  
  
  How fast does an error arrive?
&lt;/h2&gt;

&lt;p&gt;I can measure one part of "time to alert" and not the other.&lt;/p&gt;

&lt;p&gt;Measured: the gap between the SDK's event timestamp and the moment Bugsly received the event, for 15,190 stored events. Median 108 ms, 90th percentile 17.2 s, 99th percentile 154.4 s. The timestamp comes from the client, so offline queues in mobile apps and clock skew inflate the tail.&lt;/p&gt;

&lt;p&gt;Not measured: the time from event to alert email. Alert delivery is not timestamped in our data, so I will not publish a number. (32 active rules cover these projects: 16 new-issue, 16 regression, all email.)&lt;/p&gt;

&lt;h2&gt;
  
  
  What could an AI agent fix?
&lt;/h2&gt;

&lt;p&gt;We have not measured an agent's fix rate on this data. What follows is my own triage of the 12 largest issues (70,418 occurrences, 78.7% of the total), judged from the title and stack frame. "Code only" means the stack trace is enough to write the patch.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;#&lt;/th&gt;
&lt;th&gt;Issue&lt;/th&gt;
&lt;th&gt;Occurrences&lt;/th&gt;
&lt;th&gt;What closes it&lt;/th&gt;
&lt;th&gt;Bucket&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;1&lt;/td&gt;
&lt;td&gt;Our email API returned 403 to a billing task&lt;/td&gt;
&lt;td&gt;21,926&lt;/td&gt;
&lt;td&gt;Fix a credential or sender setting&lt;/td&gt;
&lt;td&gt;Config&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;2&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;TypeError: object of type 'bool' has no len()&lt;/code&gt; in an event insert&lt;/td&gt;
&lt;td&gt;12,348&lt;/td&gt;
&lt;td&gt;Add a type guard in one function&lt;/td&gt;
&lt;td&gt;Code only&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;3&lt;/td&gt;
&lt;td&gt;ClickHouse memory limit exceeded (code 241)&lt;/td&gt;
&lt;td&gt;11,123&lt;/td&gt;
&lt;td&gt;Tune the query or raise limits&lt;/td&gt;
&lt;td&gt;Infra&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;4&lt;/td&gt;
&lt;td&gt;Browser &lt;code&gt;Failed to fetch&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;5,371&lt;/td&gt;
&lt;td&gt;Nothing to fix in code&lt;/td&gt;
&lt;td&gt;No signal&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;5&lt;/td&gt;
&lt;td&gt;Video API returned 400 for an unsupported parameter&lt;/td&gt;
&lt;td&gt;4,984&lt;/td&gt;
&lt;td&gt;Remove the parameter&lt;/td&gt;
&lt;td&gt;Code only&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;6&lt;/td&gt;
&lt;td&gt;Celery &lt;code&gt;KeyError&lt;/code&gt;: task not registered&lt;/td&gt;
&lt;td&gt;3,541&lt;/td&gt;
&lt;td&gt;Align worker and scheduler versions&lt;/td&gt;
&lt;td&gt;Config&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;7&lt;/td&gt;
&lt;td&gt;Postgres &lt;code&gt;column does not exist&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;3,147&lt;/td&gt;
&lt;td&gt;Apply the missing migration&lt;/td&gt;
&lt;td&gt;Migration&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;8&lt;/td&gt;
&lt;td&gt;AWS SES &lt;code&gt;InvalidSignatureException&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;2,273&lt;/td&gt;
&lt;td&gt;Check credentials and region&lt;/td&gt;
&lt;td&gt;Config&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;9&lt;/td&gt;
&lt;td&gt;Browser &lt;code&gt;Failed to fetch&lt;/code&gt; (second source)&lt;/td&gt;
&lt;td&gt;1,742&lt;/td&gt;
&lt;td&gt;Nothing to fix in code&lt;/td&gt;
&lt;td&gt;No signal&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;10&lt;/td&gt;
&lt;td&gt;Browser &lt;code&gt;Failed to fetch&lt;/code&gt; (injected script)&lt;/td&gt;
&lt;td&gt;1,483&lt;/td&gt;
&lt;td&gt;Nothing to fix in code&lt;/td&gt;
&lt;td&gt;No signal&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;11&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;Unknown Error&lt;/code&gt; with no frames&lt;/td&gt;
&lt;td&gt;1,245&lt;/td&gt;
&lt;td&gt;Add instrumentation first&lt;/td&gt;
&lt;td&gt;No signal&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;12&lt;/td&gt;
&lt;td&gt;Same 403 as #1, in the notification dispatcher&lt;/td&gt;
&lt;td&gt;1,235&lt;/td&gt;
&lt;td&gt;Same fix as #1&lt;/td&gt;
&lt;td&gt;Config&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Code only: #2 and #5, 17,332 occurrences (19.4% of all volume). Config, infra or migration: #1, 3, 6, 7, 8 and 12, 43,245 (48.3%). No signal: #4, 9, 10 and 11, 9,841 (11.0%).&lt;/p&gt;

&lt;p&gt;My read: an agent that can read a stack trace and your repository helps on the one-fifth of volume that is a code change. For the rest, the useful output is a diagnosis ("this migration was never applied") that a human still has to act on.&lt;/p&gt;

&lt;h2&gt;
  
  
  How we wired it
&lt;/h2&gt;

&lt;p&gt;Every app uses a stock Sentry SDK. The only change per app was the DSN:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="nc"&gt;Sentry&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;init&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="n"&gt;dsn&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"https://&amp;lt;public_key&amp;gt;@bugsly.dev/&amp;lt;project_id&amp;gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Stored events show the mix: &lt;code&gt;sentry.javascript.nextjs&lt;/code&gt; (2,690), &lt;code&gt;sentry.python.fastapi&lt;/code&gt; (1,058), &lt;code&gt;sentry.ruby.sidekiq&lt;/code&gt; (709), &lt;code&gt;sentry.ruby&lt;/code&gt; (706), &lt;code&gt;sentry.dart&lt;/code&gt; (613). Guide: &lt;a href="https://bugsly.dev/migrate-from-sentry" rel="noopener noreferrer"&gt;bugsly.dev/migrate-from-sentry&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;For agents, Bugsly has an MCP server. In Claude Code the hosted version is one command, followed by an OAuth consent page:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;claude mcp add --transport http bugsly https://bugsly.dev/api/mcp
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The local package takes a token instead: &lt;code&gt;claude mcp add bugsly -e BUGSLY_TOKEN=&amp;lt;token&amp;gt; -- npx -y @bugsly/mcp-server&lt;/code&gt;. Triage uses &lt;code&gt;list_issues&lt;/code&gt;, &lt;code&gt;get_issue_events&lt;/code&gt;, &lt;code&gt;search_errors&lt;/code&gt; and &lt;code&gt;resolve_issue&lt;/code&gt;; the hosted server caps event size and omits user email, IP and username. A prompt that works: "Use Bugsly to read the most frequent unresolved issue, find the failing line in this repository, and propose a fix. Do not change the issue's status." Bugsly does not write the patch; your editor's agent proposes it and you review it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently asked questions
&lt;/h2&gt;

&lt;h3&gt;
  
  
  What are the most common production errors?
&lt;/h3&gt;

&lt;p&gt;In our data: upstream HTTP 4xx/5xx errors (32.5% of occurrences), plain code bugs such as type and null errors (21.4%) and ClickHouse limits (15.6%). This is a case study of 16 apps, not a benchmark.&lt;/p&gt;

&lt;h3&gt;
  
  
  How much of the error volume is noise?
&lt;/h3&gt;

&lt;p&gt;465 of 1,195 issues (38.9%) were seen exactly once, but they are only 0.5% of occurrences. Four of the 12 largest issues (11.0% of all occurrences) had no usable signal: browser &lt;code&gt;Failed to fetch&lt;/code&gt; errors and an &lt;code&gt;Unknown Error&lt;/code&gt; with no stack frames.&lt;/p&gt;

&lt;h3&gt;
  
  
  How quickly does an error tracker receive an error?
&lt;/h3&gt;

&lt;p&gt;For 15,190 stored events the median gap between event timestamp and receipt was 108 ms, the 90th percentile 17.2 s and the 99th 154.4 s. The tail is likely client-side (offline queues, clock skew). This is ingest latency, not time to alert.&lt;/p&gt;

&lt;h3&gt;
  
  
  Can an AI agent fix production errors on its own?
&lt;/h3&gt;

&lt;p&gt;Partly. In my manual triage of the 12 largest issues, 19.4% of all occurrences needed only a code change a stack trace is enough to write, and 48.3% needed configuration, infrastructure or a migration that a person must apply. I did not run an agent on this data, so this is not a measured fix rate.&lt;/p&gt;

&lt;h3&gt;
  
  
  Do I need to rewrite my instrumentation to try Bugsly?
&lt;/h3&gt;

&lt;p&gt;No. Bugsly accepts Sentry SDK envelopes, so you change the DSN and keep the SDK. Test the SDK features you rely on before moving production traffic. Sentry is a trademark of its owner; Bugsly is not affiliated with or endorsed by Sentry.&lt;/p&gt;

&lt;h2&gt;
  
  
  Methodology and data
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Read-only snapshot, 2026-11-17 around 09:40 UTC, from production Postgres (issue counters) and ClickHouse (ingest latency).&lt;/li&gt;
&lt;li&gt;Included: organization 1 (our own), 17 projects, 16 with at least one error, 1,195 issues, 89,499 occurrences, first seen 2026-03-13.&lt;/li&gt;
&lt;li&gt;Excluded: every other organization, including one external tenant.&lt;/li&gt;
&lt;li&gt;Categories are regular expressions on issue titles; the rules and every query are in the published queries. Hand triage of the top 12 is my judgment, not an agent run.&lt;/li&gt;
&lt;li&gt;Ingest latency: events where receive time is after event time and within one day (15,190 of 15,291 stored events); older than 90 days is deleted.&lt;/li&gt;
&lt;li&gt;Disclosure: we build and sell Bugsly. It is hosted and closed source.&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>webdev</category>
      <category>monitoring</category>
      <category>devops</category>
      <category>ai</category>
    </item>
    <item>
      <title>How ATS Software Actually Parses and Rejects Your Resume</title>
      <dc:creator>Kien Phan</dc:creator>
      <pubDate>Wed, 03 Jun 2026 09:07:35 +0000</pubDate>
      <link>https://dev.to/kienpt91/how-ats-software-actually-parses-and-rejects-your-resume-3ipi</link>
      <guid>https://dev.to/kienpt91/how-ats-software-actually-parses-and-rejects-your-resume-3ipi</guid>
      <description>&lt;p&gt;You applied. You heard nothing. You assumed you weren't qualified.&lt;/p&gt;

&lt;p&gt;You might have been perfectly qualified — and a parser threw your resume in the bin before any human laid eyes on it.&lt;/p&gt;

&lt;p&gt;Most ATS advice stays on the surface: "use keywords," "avoid tables." Useful, but it skips the more interesting question: &lt;em&gt;why&lt;/em&gt; do these things matter? What is actually happening inside the system when it reads your resume?&lt;/p&gt;

&lt;p&gt;This is that explanation. As someone who spent time building an ATS checker, I've seen the output from enough parsed resumes to know exactly where things go wrong.&lt;/p&gt;

&lt;h2&gt;
  
  
  What an ATS Parser Actually Does
&lt;/h2&gt;

&lt;p&gt;When you upload a resume to an applicant tracking system, the first thing that happens is not a keyword search. It's a text extraction step. The system converts your document — PDF, DOCX, or otherwise — into a flat, unformatted string of text. Then it tries to segment that string into structured fields: name, contact info, job titles, companies, dates, education, skills.&lt;/p&gt;

&lt;p&gt;Think of it like running OCR on a handwritten letter, then having a script guess where each paragraph begins and ends. More chaos in the original means worse output.&lt;/p&gt;

&lt;p&gt;The structured data it extracts gets stored in a database. When a recruiter searches "5 years Python" or "product manager fintech," they're querying that database — not your original document. Your formatting no longer exists at this point. Only the extracted fields do.&lt;/p&gt;

&lt;p&gt;This is the thing most candidates don't internalize: &lt;strong&gt;your resume is converted to data before it is ever evaluated.&lt;/strong&gt; If the conversion is wrong, the evaluation is wrong.&lt;/p&gt;

&lt;h2&gt;
  
  
  How Keyword Matching and Scoring Work
&lt;/h2&gt;

&lt;p&gt;Once parsed, most ATS platforms score your resume against the job description using keyword matching. Workday, Greenhouse, Lever, iCIMS, and Taleo all have different scoring logic, but the general mechanics are similar.&lt;/p&gt;

&lt;p&gt;The system identifies required and preferred keywords in the job description, then checks your parsed resume for those terms. The resulting score determines where you appear in the recruiter's queue.&lt;/p&gt;

&lt;p&gt;A few things worth knowing about how this actually works:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Exact phrasing often beats semantic equivalence.&lt;/strong&gt; "Stakeholder management" and "managing stakeholders" may be treated as different tokens in older systems. Mirror the job description's exact phrasing wherever it accurately describes your experience.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Keyword placement affects score weight in some systems.&lt;/strong&gt; Terms in your summary or early in experience bullets tend to carry more weight than identical terms buried at the bottom.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Newer platforms with NLP check context, not just presence.&lt;/strong&gt; "Managed a Python codebase" scores better than "familiar with Python" for a senior engineering role.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Your score is relative, not absolute.&lt;/strong&gt; You're ranked against other applicants for that posting. A 60% match against a niche role with 40 applicants might get you through. The same score against a viral posting with 800 applicants might put you on page 12.&lt;/p&gt;

&lt;h2&gt;
  
  
  What "Parse Failure" Looks Like — And Why It Happens
&lt;/h2&gt;

&lt;p&gt;Parse failures are the silent killers. Your resume looks correct to you. After extraction, the data in the ATS is garbled, incomplete, or misattributed. You never find out.&lt;/p&gt;

&lt;h3&gt;
  
  
  Tables and Multi-Column Layouts
&lt;/h3&gt;

&lt;p&gt;This is the most widespread failure. When a parser hits a table, it reads across rows instead of down columns. Two parallel columns get merged into incoherent strings.&lt;/p&gt;

&lt;p&gt;A two-column resume with job history on the left and skills on the right extracts as alternating fragments of both. Your job title, company, location, and dates can collapse onto one line, and bullets stop being associated with the right role — so "5 years of engineering experience" may fail to map to any recognized work history record.&lt;/p&gt;

&lt;h3&gt;
  
  
  Headers and Footers
&lt;/h3&gt;

&lt;p&gt;Contact information in the document header is a classic mistake. Many ATS platforms extract the document body only, skipping headers and footers entirely. Your name and email may not exist in the system. The recruiter who wants to call you back literally cannot find your contact details.&lt;/p&gt;

&lt;p&gt;Put your name, email, phone, and LinkedIn URL in the body of the document, at the top, as plain text.&lt;/p&gt;

&lt;h3&gt;
  
  
  Text Boxes and Sidebars
&lt;/h3&gt;

&lt;p&gt;Content in a text box is often ignored entirely, or extracted out of order. If your skills section or summary is in a sidebar text box because your template looks slick, it may be completely missing from the parsed output.&lt;/p&gt;

&lt;h3&gt;
  
  
  Images and Icons
&lt;/h3&gt;

&lt;p&gt;Inline icons — a phone icon next to your number, a LinkedIn logo next to your URL — are invisible to text parsers. If your name or any section header is rendered as an image (common in design-heavy templates), the parser won't find it. Submitting a scanned resume as a JPEG produces the same problem at scale: the entire document is one large image with no extractable text.&lt;/p&gt;

&lt;h3&gt;
  
  
  Non-Standard Section Headings
&lt;/h3&gt;

&lt;p&gt;ATS systems are trained to recognize section boundaries. They know "Work Experience," "Professional Experience," "Employment History." They are less reliable with creative alternatives. Headings like "Where I've Made an Impact," "My Story So Far," or "Toolkit" may fail to map to a recognized field — and content beneath them may be skipped or misattributed.&lt;/p&gt;

&lt;h3&gt;
  
  
  Date Format Inconsistency
&lt;/h3&gt;

&lt;p&gt;Parsers need to extract date ranges to calculate years of experience. "Jan 2019 – Mar 2022" is fine. Problems arise with unusual formats, ambiguous year-only ranges, or dates embedded in sentences. If a parser can't extract your employment dates reliably, it may assign zero years of experience to a role — which collapses your seniority score against requirements.&lt;/p&gt;

&lt;h2&gt;
  
  
  The Plain-Text Test
&lt;/h2&gt;

&lt;p&gt;The simplest way to catch parse failures before you submit: select all text in your resume and paste it into a plain text editor. Look at the output.&lt;/p&gt;

&lt;p&gt;Contact details missing? They were in a header. Columns jumbled? You have a table or multi-column layout. Sections out of order? Non-linear structure. Whatever looks broken in plain text is exactly what the ATS will see.&lt;/p&gt;

&lt;p&gt;A well-formatted resume should paste as clean, readable, top-to-bottom text with clear section breaks.&lt;/p&gt;

&lt;h2&gt;
  
  
  What This Means for Your Format
&lt;/h2&gt;

&lt;p&gt;Given how parsers work, the formatting rules stop being arbitrary:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Single column.&lt;/strong&gt; Always. The visual appeal of a two-column template is not worth the parsing risk.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;No tables.&lt;/strong&gt; Use bullet points and line breaks instead.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;No text boxes.&lt;/strong&gt; If your template uses them, find a different template.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;No images.&lt;/strong&gt; Not icons, not photos, not decorative elements.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Contact info in the body.&lt;/strong&gt; First thing in the document, plain text.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Standard section headings.&lt;/strong&gt; Work Experience, Education, Skills, Certifications. Boring is correct.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Consistent date formatting.&lt;/strong&gt; Pick one format and use it everywhere.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;A clean single-column resume with good typography and strong content looks more professional than most over-designed templates — and it parses cleanly.&lt;/p&gt;

&lt;h2&gt;
  
  
  One Thing Worth Doing Before Every Application
&lt;/h2&gt;

&lt;p&gt;After fixing your formatting, the remaining variable is keyword alignment with each specific job description. No single resume scores well against every posting. The discipline is tailoring: checking which terms appear in the job description and making sure they appear in your resume where accurate.&lt;/p&gt;

&lt;p&gt;I built CVPosh's free ATS checker to automate exactly this — upload your resume, paste the job description, and get your match score plus which specific terms are present or missing. It takes about 90 seconds and removes the guesswork.&lt;/p&gt;

&lt;h2&gt;
  
  
  The Bigger Picture
&lt;/h2&gt;

&lt;p&gt;ATS systems aren't trying to screen out qualified candidates. They're trying to help recruiters manage volume. "Beating" the system misframes the goal.&lt;/p&gt;

&lt;p&gt;The real goal is making sure your resume converts to accurate, complete data. If the parser reads your work history, skills, and dates correctly, the system can evaluate you fairly. The formatting rules above aren't tricks — they're the conditions under which automated parsing works correctly.&lt;/p&gt;

&lt;p&gt;Get the conversion right. Make sure the content is strong. That combination — clean parse, relevant content — is what actually moves your resume to the top of the queue.&lt;/p&gt;

</description>
      <category>career</category>
      <category>webdev</category>
      <category>productivity</category>
      <category>ai</category>
    </item>
  </channel>
</rss>
