Our score is a composite of five pillars, each weighted by how much it drives whether a resume makes it to a human recruiter in 2026. The weightings reflect how the current generation of enterprise Applicant Tracking Systems rank candidates against a specific job description, not a generic quality score.

The 5-pillar rubric

1. Keyword density and contextual match (30%)

Every modern ATS ranks a candidate on how closely the words in the resume map to the words in the job posting. We look at both exact matches (an identical tool name, certification, or job title) and contextual matches (a synonym the parser is likely to normalize, like "React.js" and "ReactJS"). We penalize keyword stuffing: a term appearing far more often than it would naturally in a resume is a signal that an older ATS might flag, and that a human recruiter will dismiss on a skim.

The weighting is highest here because the first filter inside a Workday or Taleo pipeline is almost always a keyword or Boolean match. If a resume misses the first filter, the other four pillars never get read.

2. Hard and soft skill alignment (25%)

A skill section is only valuable if the skills it names are the ones the role calls for. Our model separates hard skills (specific tools, languages, platforms, certifications) from soft skills (communication, leadership, stakeholder management) and checks each list independently against the posting. Hard skills count for more than soft skills, because they are the ones most parsers can confirm. A match on "stakeholder management" is useful; a match on "Snowflake" is actionable.

3. ATS format and layout parsability (20%)

Even a perfect keyword match can fail if the ATS can't read the file. The format pillar scores how cleanly the resume tokenizes on a typical enterprise parser:

- Single-column, text-selectable PDF or DOCX, not an image scan.
- Standard section headings (Experience, Education, Skills), because parsers key off those labels.
- Readable dates in a consistent format (MMM YYYY or YYYY).
- No critical content inside headers, footers, text boxes, or SVG graphics, which many parsers skip.
- Simple bullet characters, because exotic glyphs sometimes get dropped during tokenization.

4. Experience and impact relevance (15%)

Years of experience, seniority markers, and quantified impact. We weight experience that is semantically close to the role above raw tenure. A software engineer applying for a backend role gets more credit for three years of backend work than for seven mixed years. We also look for impact phrasing: numbers, percentages, named projects, outcomes. "Reduced API latency 42% by moving to Rust" is more parseable to both a human and an LLM-powered screener than "worked on performance."

5. Education and certification fit (10%)

Education is the smallest pillar because many 2026 roles no longer gate on a specific degree. Where education matters, it's usually because the posting lists a required credential (bar passage, PE license, PMP, AWS certification) or a specific field of study. We check the match and flag missing credentials that the posting explicitly requires.

How enterprise ATS parsers read your resume

No two ATS platforms tokenize a resume the same way, but the common path looks like this. A file (PDF or DOCX) is handed to a parser that extracts raw text, segments it into sections based on heading patterns, and emits a structured record: name, contact info, experience, education, skills, free text. The record is indexed, and the job posting is converted into a query against that index.

Workday

Workday's parser is strict on structure. Section headings matter more than visual formatting. Candidates often get asked to re-enter data by hand after upload because the parser could not confidently fill a field. A single-column layout with clear "Experience" and "Education" headings reduces that re-entry.

Taleo

Taleo (now Oracle Recruiting Cloud for many enterprise customers) leans heavily on Boolean keyword fields configured by the recruiter. Exact keyword matches against those fields drive ranking. Synonyms often don't help here; the exact term from the posting does.

Greenhouse

Greenhouse puts more weight on scorecards and tagged attributes added by recruiters. The parser itself is friendlier to modern formats. Keyword density still matters, but a messy layout fails less often than it does in Workday.

Lever

Lever's parser handles multi-column PDFs better than most. It still prefers clean section labels, and its tag-based search means the exact skill name used in the posting is the one worth including.

iCIMS

iCIMS supports several parser backends, so behavior varies. The safest assumption is strict: single column, standard headings, text-based PDF.

Jobvite and SmartRecruiters

Both lean on structured fields and tagged skills added during application. The resume still gets parsed and ranked, so keyword and format rules apply, but recruiters usually have more flexibility to re-rank manually.

How we use AI

The scoring itself is grounded in the rubric above. We use a large language model to apply the rubric at scale: it reads the resume text and the job description, scores each pillar, extracts matched and missing keywords, and generates specific quick wins. The model does not invent facts about you; it reads only what you upload. Our primary model in October 2026 is a Llama-family model served by Groq, with a Gemini fallback if the primary is unavailable. Model names change over time; the current chain is documented in the open-source codebase.

Privacy

Uploaded resumes are deleted within sixty seconds of a scan completing. The analysis result is kept only when you have signed in and chosen to save it to your history. Anonymous rate limiting uses a one-way hash of your IP address, not the IP itself. We do not sell resumes, we do not share uploaded files with third parties beyond the AI provider that performs the analysis, and we do not run display or retargeting ads.

Updates to this methodology

Parser behavior, model names, and recruiter conventions change. We revisit this page at least once per quarter and note the date at the top. If a change affects how your past scans would be interpreted, we call it out explicitly in the changelog section of the page.

Reviewed by Nauman Pathan on 14 October 2026.