← Guides

The résumé keyword myth

A real, documented gap

Users of at least one popular keyword-matching tool report scoring 80% or higher on its match-your-resume-to-this-job-posting feature — and still getting zero callbacks. That's not a fluke or a one-off complaint; it points at a real gap between what keyword-matching tools optimize for and what actually determines whether you hear back.

Why this happens: two different layers, one confused with the other

Our guide on the agentic problem covers how modern CV matching actually works: a lexical layer (keyword overlap, TF-IDF — did the words on the page show up?) sitting underneath a semantic layer (does the meaning of your experience actually match what's being asked for?), often followed by a human reading whatever survives both. A keyword-matching tool that scores you against a job posting is checking the lexical layer only — literally, are these words present. It has no way to evaluate whether you're using those words correctly, in context, backed by real substance, versus dropped into a skills list because the tool told you to add them.

Passing the lexical check gets you past the crudest filter. It says nothing about the semantic layer or the human read that typically follows — and both of those can tell the difference between "genuinely has this experience, described specifically" and "pasted the job posting's own vocabulary into a bullet list."

What keyword stuffing actually looks like, and why it backfires

  • A skills list with terms never mentioned anywhere else in the CV. If "stakeholder management" only appears once, in a list, with no role or achievement that demonstrates it, both a careful human reader and an increasingly capable semantic matching system can tell the difference between a claimed skill and a demonstrated one.
  • Repeating the job posting's exact phrasing without specificity. "Experience with cloud infrastructure" copied verbatim doesn't carry more information than the job posting itself did — "AWS Lambda and DynamoDB, migrated our infra off a monolith EC2 setup" does, and it matches on both the lexical and semantic layers because it's genuinely richer.
  • Padding for the sake of a score, rather than because the experience is real. A tool telling you your match score is 60% isn't necessarily telling you to add more keywords — it might be correctly telling you that the underlying experience genuinely isn't there yet.

What actually works instead

The fix isn't "use fewer keywords" or "use more keywords" — it's specificity over vocabulary matching. Say the real thing, with real detail: the actual tool, the actual scale, the actual outcome. This naturally produces the right keywords (because the right keywords are just the accurate words for what you actually did) while also giving a human reader, and the semantic-matching layer underneath any modern screening system, something real to evaluate — not a vocabulary checklist satisfied on paper.

A high keyword-match score is a floor, not a ceiling. It tells you that you probably won't be filtered out on the crudest pass. It tells you nothing about whether the substance underneath will actually convince whatever reads it next — human or model.

Related guides