Keyword matching vs semantic screening: what actually changes for recruiters

Almost every applicant-tracking system in production today still leans on keyword matching. You paste a job description, the ATS pulls a list of strings, and resumes are scored by how many of those strings appear verbatim. It is fast, transparent, and — for any role that is not a near-exact rerun of one the team has hired for before — wildly incomplete.

Where keyword matching breaks

Two specific failure modes show up in almost every search we audit:

• False negatives from vocabulary drift. A backend engineer writes "built distributed event streams" on their resume. Your JD says "Kafka." A keyword parser scores them at zero. A semantic system reads both phrases and recognises them as the same competency.
• False positives from string presence. A candidate lists "managed Kafka cluster downtime incidents" in their last role. Keyword: 1 match. Semantic: they are not a Kafka platform engineer, they were an SRE on-call. Different shortlist.

The two failure modes do not cancel out. They compound. The recruiter at the end of the funnel sees a pile that is missing the strongest candidates and contains people who match strings but not roles.

What semantic screening changes upstream

If you only swap the matching engine, you mostly get the same shortlist faster. The bigger lever is changing how you write screening criteria, because semantic matching lets you write them as recruiter questions instead of keyword lists.

Before (keyword)

Required skills: Kafka, Spark, Airflow, Python, AWS, Postgres

After (semantic)

Has the candidate built and operated a streaming or batch data pipeline at production scale, with on-call ownership? Evidence: which system, which technologies, which year.

The second formulation is a real screening criterion. It produces evidence, it can be answered yes/no/needs-review, and it survives the candidate using a different stack with the same shape of experience.

Three things to do this week if you screen with keywords

• Take the last role you closed and rewrite three of its must-haves as questions instead of keyword lists. See how much your final shortlist shifts when you re-score.
• Audit any role that has been open more than 60 days. Almost always, the bottleneck is a too-literal must-have that excludes the candidates you would actually want to interview.
• Stop relying on the ATS resume-rank column as the input to your call list. Use it as a sanity check, not a ranker.