Keywords Are Not Hiring Signal: Why Context Has to Come Before the Conversation

Sam RahmanChief Product Officer, Kreativs

Key takeaways

77% of executives told PwC that hiring and retaining talent is their most critical growth driver, yet the technology underneath that priority still ranks people by keyword matching.
Roughly 70% of workers admit to misrepresenting something on a resume, which means the document most systems trust as ground truth is the least reliable input in the process.
Free and paid tools now tell candidates their exact match percentage and the precise keywords to add, which converts screening into a contest of who optimized best rather than who delivered most.
Context and semantic understanding, not keyword overlap, are what produce defensible hiring signal before a recruiter ever picks up the phone.
The goal is not a better filter. The goal is a shortlist that starts from understanding instead of a guess.

The priority is right. The plumbing is broken.

In its Pulse Survey, PwC found that 77% of executives said hiring and retaining talent was their most critical growth driver.¹ That is not an HR statistic. That is a board-level statement about where growth comes from.

Now look at what sits underneath that priority in most organizations. A requisition gets written, posted, and pushed into an applicant tracking system that was built around one core assumption: that the words in a resume, compared against the words in a job description, are a reasonable proxy for whether a person can do the job.

They are not. They never really were. And in 2026, that assumption has collapsed entirely.

Why does keyword matching no longer work?

Keyword matching worked, sort of, in a world where the resume was a slow, handmade artifact and the candidate had no visibility into how it would be parsed. Both of those conditions are gone.

Three things broke at once.

1. The resume stopped being a reliable input

Forbes reported on a ResumeLab study of 1,914 workers that found 70% admitted to lying on their resumes, with 37% saying they do it frequently.² The most common fabrications were exactly the ones a keyword parser cannot detect: inflated job titles, exaggerated scope of management, and stretched employment dates.

Think about what that means mechanically. A keyword system does not evaluate a claim. It only confirms the claim is there. When the claim itself is unreliable, confirming it is present is not screening. It is transcription.

2. The optimization layer became universal

There is now an entire category of tooling built specifically to beat resume parsing. Jobscan, to take the most established example, scores a resume against a job description and recommends a 75% match rate,³ then shows the applicant exactly which hard skills, soft skills, and keywords are missing so they can be added.

That is a legitimate product serving a real need. The point is not that these tools are unethical. The point is what their existence proves: if a match score can be reverse engineered to a number and a checklist, the score was never measuring capability.

And that was before general-purpose AI. Today a candidate can paste a job description and a resume into ChatGPT, Claude, or Gemini and get a rewritten document in under a minute, tuned to whatever system sits on the other side. LinkedIn reported roughly 11,000 job applications submitted per minute, a 45% year over year increase, with AI cited as a major driver of the surge.⁴

The optimization layer is no longer a competitive edge held by a savvy few. It is table stakes for anyone paying attention, and invisible to everyone who is not.

3. The "AI" retrofit did not change the underlying logic

Here is the part the market has not fully priced in. A great deal of the AI and agentic AI functionality bolted onto legacy hiring platforms is presentation-layer work. The interface got smarter. The ranking logic frequently did not. Underneath the new language, many of these systems are still counting how many words a resume shares with a job description, then wrapping that count in confident-sounding scoring.

Buying a keyword engine with a chat interface does not give you a context engine. It gives you the same guess, delivered with more conviction.

Keyword matching turned hiring into a sales contest

This is the consequence that should concern talent leaders most, and it rarely shows up in a report.

When the filter is knowable, the filter becomes the game. Candidates who understand the market dynamic optimize for it and advance. Candidates who do not, and that includes a large share of genuinely excellent operators, engineers, and clinicians who have simply never heard of a match rate, get auto-rejected or buried in a pile they had no fair shot at.

You are not selecting for the best candidate. You are selecting for the best-marketed candidate.

Those two populations overlap far less than most hiring teams assume. A brilliant systems engineer with eleven years of delivery behind her may write a plain, honest, poorly formatted resume and lose to someone who ran the job description through a match-rate tool that morning. The system worked exactly as designed. That is the problem.

The damage compounds from there:

Your funnel skews toward resume skill. Being good at writing a resume has very little to do with being good at the job.
You never see who you lost. Rejected candidates do not tell you they were qualified, so the process looks like it is working.
Trust erodes. Candidates who sense the process is arbitrary disengage, and that experience becomes your employer brand.

The human layer has the same problem

Technology is only half the failure. The other half is structural.

In a large share of hiring processes, the person conducting the first evaluation is a liaison rather than a practitioner. Recruiters are asked to assess technical depth, architectural judgment, clinical competence, or regulatory expertise in domains they were never trained in, using a job description written by someone else, often with limited access to the hiring manager's actual bar.

That is not a criticism of recruiters. It is a criticism of a process that asks them to do something structurally impossible and then treats the result as signal.

The predictable outcome is subjectivity. Two recruiters review the same resume and reach different conclusions. The same recruiter reaches different conclusions on Monday morning and Thursday afternoon. Decisions get justified after the fact with language borrowed from the requisition. Nobody is acting in bad faith, and the process still produces inconsistent, unauditable outcomes.

Add a keyword filter on top of subjective human review and you have a process where neither layer can explain, in specific terms, why one person advanced and another did not.

What does context-based evaluation actually mean?

Context means reading a candidate's history the way a competent hiring manager would read it, and doing that consistently at volume.

The unit of evaluation is not the keyword. It is the delivery. What did this person actually ship, build, fix, close, treat, negotiate, or lead? At what scale? With what constraints? Over what duration? What does the trajectory across roles tell you about depth versus surface exposure?

A context-driven read asks questions a keyword filter simply cannot:

Does the stated seniority match the described scope of work, or is there a gap between title and evidence?
Is a listed skill supported anywhere by delivery, or does it appear only in a skills block at the bottom of the page?
Does adjacent, differently-labeled experience actually satisfy the requirement even though the vocabulary does not match the requisition?
Where is this candidate incomplete against the role, and how material is that gap?

That last question matters as much as the others. Context-based evaluation is not just a better way to find matches. It is a way to describe completeness: what a candidate demonstrably has, what they demonstrably lack, and how much that gap actually costs for this specific role.

Semantic understanding is what makes this possible. A system that reads for meaning rather than matching words can recognize that a candidate who "rebuilt the payments checkout used by 40 million customers a day and cut its failure rate in half" has exactly the depth the role needs, even if the job description asked for "performance optimization experience" and the resume never used that phrase. It can also recognize that a resume listing fourteen tools and skills with nothing to show for any of them is thin, no matter how many boxes it ticks.

Keyword systems reward vocabulary. Context systems reward evidence. Those produce very different shortlists.

How Kreativs approaches this

We built Kreativs Copilot from the ground up on context and semantic learning, not keyword matching, because we did not think a keyword engine could be patched into a context engine.

Copilot reads a resume the way an experienced practitioner would read it. It grades against what the candidate has delivered, the skills that delivery actually evidences, and the completeness of that profile relative to the role. It does not compute keyword overlap and it does not rank by how well a document was optimized. Every assessment comes with the reasoning behind it, so a recruiter is looking at a substantiated read rather than an unexplained number.

And the decision stays human. Nobody is auto-disqualified by the system. The output of evaluation is an informed shortlist, and a person decides who moves forward.

This is the paradigm shift we are arguing for, and it is worth stating plainly: your recruiters should walk into the first conversation already understanding the candidate, not hoping to discover them. Signal should be established before outreach, not reconstructed during a screening call. That protects strong candidates who never learned to game the system, and it gives recruiters a defensible basis for a decision in domains they do not personally practice in.

What changes for your team

If you move evaluation from keywords to context, several things change quickly.

Your shortlist composition shifts. Expect to see candidates surface who would previously have been filtered out, and expect some heavily optimized profiles to fall.
First conversations get better. When a recruiter starts from understanding, the screen becomes a validation and motivation conversation rather than a discovery interview.
Decisions become explainable. "Advanced because of X delivery evidence, flagged on Y gap" is auditable in a way that "85% match" never was.
Gaming the system stops paying. When evaluation reads what someone delivered rather than what words they used, adding twelve keywords to a resume does not change the outcome.

None of this removes the human from hiring. It removes the guess.

The old process is not underperforming. It is broken.

There is a meaningful difference between a process that needs tuning and a process built on an assumption that no longer holds. Keyword matching is the second kind. When 70% of resumes contain misrepresentation, when candidates can query their exact match score for free, and when 11,000 applications hit a single platform every minute, a system that ranks people by how many words they share with a job description is not a filter. It is noise dressed up as a score.

The way forward is not a stricter filter or a smarter-sounding score. It is understanding: what a candidate has delivered, what their experience actually evidences, and how complete they are against the role. Get that right, and the shortlist your recruiters work from starts from an understanding rather than a guess.

That is the standard we think hiring should be held to, and it is what we built Kreativs Copilot to deliver.

Want to see what context-based evaluation looks like on your own pipeline? Talk to us today.

FAQ

Overwhelmingly, yes. Most ATS platforms were built to match the text of a resume against the text of a job description, and much of the AI functionality added to them in recent years still relies on that same keyword matching to rank candidates, even when the screen in front of you suggests otherwise.

Yes, and the tooling is widely available. Products like Jobscan return a percentage match score against a job description and list the specific keywords to add. General-purpose AI assistants including ChatGPT, Claude, and Gemini can rewrite a resume against a posting in seconds. Any filter whose logic can be inferred can be optimized against.

Keyword matching checks whether specific words appear in a document. Semantic evaluation reads what those words actually mean: what the candidate delivered, what scope and constraints they operated under, and whether their stated skills are supported by evidence. The first measures vocabulary. The second measures capability.

By evaluating documented delivery, demonstrated skills, and profile completeness against the actual requirements of the role, rather than by counting shared words. A context-based evaluation produces a substantiated read of what a candidate has done and where they are incomplete, which is enough to decide whether a conversation is warranted.

No. It changes what recruiters are asked to do. Instead of making subjective judgments about domains they may not practice in, they receive an evidence-backed evaluation and make the human decision about who moves forward.

Sources

PwC Pulse Survey, Executive views on business in 2022. https://www.pwc.com/us/en/library/pulse-survey/executive-views-2022.html
Forbes, "70% Of Workers Lie On Resumes, New Study Shows" (ResumeLab survey, n=1,914). https://www.forbes.com/sites/bryanrobinson/2023/11/05/70-of-workers-lie-on-resumes-new-study-shows/
Jobscan match rate methodology. https://www.jobscan.co/
eWeek, "Job Seekers, Some Using AI, Flood LinkedIn With 11,000 Applications a Minute". https://www.eweek.com/news/ai-job-applications-linkedin/