Do I need the exact keywords on my resume?

For some words on your resume, yes. For most of them, no. The dividing line is not the one the advice usually gives you, and it is easier to apply: copy a keyword exactly when it is a name, and write everything else in your own words.

A name is a tool, a certification, a standard, a framework or a job title. Kubernetes. CIPD Level 5. IFRS. Staff Engineer. Those need to match, and the reason is not that software demands a literal string. It is that somewhere behind the posting, a person types that word into a search box.

Nobody searches for "fast-paced environment".

The exact-match step is usually a human with a search field

This is the part the category leaves out, and it changes what you do with the posting.

Greenhouse documents a full text search that a recruiter switches on from the candidate list, then queries with boolean operators. Its support page explains that AND "limits your search because you're requiring that any search term following AND must appear in the results", alongside OR, NOT, quotation marks, parentheses and a * wildcard, and says results reference "resumes, notes, or scorecards".

Teamtailor is more explicit about what gets searched. Its filtering documentation lists the fields a recruiter can query by name, and resume_text is one of them, next to cover_letters, pitch, comments and reviews. The same page documents AND, OR and NOT, with Developer AND (Frontend OR Backend) as the worked example.

Read that as a job seeker and the advice rewrites itself. A boolean search returns exactly what was typed. If the recruiter searches the product name and your resume calls it something else, you are not in the result set, and no amount of context helps you. If the recruiter searches a vague adjective, which they will not, the whole question is moot.

So the exact-match rule is real and narrow. It applies to the words a person would plausibly type, which are overwhelmingly nouns that name things.

One vendor's scoring layer has stopped counting words

The newer feature in this space is not a keyword filter at all, and it arrived quietly enough that almost nothing written for job seekers mentions it.

Greenhouse shipped Talent Matching, an add-on that a site admin switches on, documented in a support article last updated on 9 September 2026. It says the feature uses "AI and your defined calibration criteria" to highlight relevant skills, "provides match scores", and sorts applicants into five bands: strong match, good match, partial match, limited match, and needs manual review. Then the sentence that matters for how your application is seen: "Candidates are grouped by match score and ordered by application date within each group."

That is automatic reordering of the queue. It is not automatic rejection, and Greenhouse says so twice in its own FAQ: the feature "does not automatically disposition any candidates nor does it make any hiring decisions", and it "does not auto-reject or auto-advance any candidate".

Ashby documents AI differently again. Its Filter Assistant turns "a plain language prompt" into filters in candidate search, which is a recruiter asking for something in a sentence rather than a system ranking you on its own.

None of these is the thing most pages describe, which is an applicant tracking system that counts how many posting words appear in your file and drops you below a threshold. If you want the longer version of why that number is not a measurement, what an ATS score measures covers it.

Does repeating the literal word still move anything?

Two studies bear on this, and they point the same way.

Samadi, Banerjee and Nilizadeh tested adversarial text insertion against recruitment ranking algorithms, and report that "attacks against TF IDF is more successful compared to USE". TF-IDF counts word frequency. USE compares meaning rather than strings. Inserting keywords moved the old method more than the newer one.

Bevara and colleagues went at it from the other side, benchmarking embedding models against a conventional keyword-based applicant tracking system across five job domains, with forty student evaluators as the reference ranking. The embeddings won on normalised discounted cumulative gain in three of the five, by 15.85% in mechanical engineering.

The two domains where the keyword system won are the useful part. It beat the embeddings on Operations Management and Software Testing, and the authors explain why: those job descriptions "rely on explicitly stated keywords or technical skills, making the ATS's keyword matching slightly more effective".

Same rule, different direction. Where a posting names specific things, naming them back matters. Where it describes work, your phrasing matters much less than whether the work is there.

Two limits on that paper, because they are real. The evaluators were students rather than recruiters, which the authors flag themselves. And the baseline is never named, so this is evidence about keyword matching as a method, not about Greenhouse or Workday.

Which lines in the posting to ignore

Half of a job description is not addressed to you, and copying it in is how a resume ends up written for software.

Skip the equal-opportunity statement, the benefits list, the sentence about being a fast-paced and collaborative environment, the legal boilerplate at the bottom, and the "nice to have" block until the required block is covered. Nobody searches those, no scoring layer rewards them, and every line you spend on them is a line not spent on evidence.

What is left is usually short: the named tools, the named standards, the qualifications, the scope words such as team size or budget or region, and the three or four outcomes the role is accountable for. That list is what to check your resume against, and the method is in how to tailor your resume to a job description.

What happens if you stuff keywords in anyway?

The measured answer is that it mostly fails, and on two conditions out of thirteen it worked completely, which makes it a long shot rather than a tactic.

Kevin Webster ran a resume-screening audit across eight consumer AI assistants, published as a preprint rather than peer-reviewed work. One condition compared an irrelevant but well-written service-industry resume against the same resume "heavily populated with finance-related terminology, but with those terms used in nonsensical or contextually inappropriate ways", for a Finance Director role. One version referenced overhearing conversations about working capital.

Eleven of the thirteen model conditions scored the stuffed version at 45 or below, most of them at or near the no-keyword score of 10. Two were fooled completely: one went from 10 to 89.5, another from 15 to 92. Each resume was submitted three times per model, so these are small numbers, and the systems tested are general-purpose chatbots rather than applicant tracking products.

Still, the shape is clear enough. Stuffing is a long shot on which reader you get, and the readers it fools are the ones least likely to be the final one. The hidden version of the same tactic has already been measured at scale and is covered in does white text on a resume work.

The claim you will also see, that applicant tracking systems flag keyword density above some repetition count, is worth naming separately. It appears on resume-tool blogs and nowhere else. No vendor documentation describes a density rule and no study found one. It is advice with no source, which is the same problem as the tactic it warns against.

Four steps that follow from all of this

1. Pull the names out of the posting first. Tools, certifications, standards, systems, the job title. These are the search terms, so these are the ones to match literally.
2. Check each one against what you did. If it is true and missing from your resume, add it where the work happened. If it is not true, it does not go on the page, and no scoring band is worth the interview where it comes up.
3. Write the rest in your own words. The requirements that describe work rather than name things do not need the posting's phrasing, and a resume that echoes a posting sentence by sentence reads as exactly that to the person at the end of the queue.
4. Look at the parse before you look at any number. A term you did include is worth nothing if it landed in the wrong field, which a multi-column layout makes more likely. The twenty-second check is in is my resume ATS friendly.

Upload your resume to see what a parser reads from your file, field by field, and the findings it raises. No account and no card.

The short version

Exact keywords matter for names and not for descriptions, because the literal matching that still happens is a recruiter typing into a boolean search field, and people type nouns. The scoring layer above it reads meaning, and the one vendor that documents it reorders a queue rather than rejecting anyone. Copy the named things when they are true. Write everything else the way you would say it.

Rzume does not give you a keyword percentage. It shows the parse first, field by field, then raises findings that quote your own lines back to you, which is a list you can act on rather than a number you can chase. In the rewrite it asks instead of inventing: skip the question and the line ships with the gap visible in brackets, not with a figure nobody supplied.

Upload your resume, or read what each plan includes.

Sources

• Greenhouse, Search candidates using Boolean queries, support documentation, page last updated 2 March 2026. The AND quotation, the operator list and the scope over resumes, notes and scorecards.
• Teamtailor, Filter, segment, and export candidates, support documentation dated 1 September 2026. The resume_text searchable field and the boolean operators.
• Greenhouse, Talent Matching, last updated 9 September 2026, and Talent Matching FAQ, last updated 3 September 2026. The five match bands, the grouped ordering, and both no-auto-rejection statements.
• Ashby, AI Features in Ashby, knowledge base. The Filter Assistant description.
• Samadi, A., Banerjee, D. and Nilizadeh, S., Attacks against Ranking Algorithms with Text Embeddings: a Case Study on Recruitment Algorithms, submitted 12 August 2021. The TF-IDF versus embedding comparison.
• Bevara, R. V. K., Mannuru, N. R., Karedla, S. P., Lund, B., Xiao, T., Pasem, H., Dronavalli, S. C. and Rupeshkumar, S., Resume2Vec: Transforming Applicant Tracking Systems with Intelligent Resume Embeddings for Precise Candidate Matching, Electronics 14(4), 794, published 18 February 2025. Forty student evaluators across five of ten domains, the 15.85% nDCG figure, and the two domains where keyword matching won.
• Fuller, J., Raman, M., Sage-Gavin, E., Hines, K. et al., Hidden Workers: Untapped Talent, Harvard Business School and Accenture, September 2021. The filter that removes a resume on an employment gap, which reads a date rather than a keyword.
• Webster, K. T., Fairness Is Not Enough: Auditing Competence and Intersectional Bias in AI-powered Resume Screening, preprint, version 3 dated 27 July 2026. The keyword-stuffed irrelevant resume condition and the two model conditions it fooled.

All accessed 4 October 2026.