AI Candidate Matching: How It Works and How to Check It

How matching tools compare candidates with a role, what a match score is built from, and how to test one before you trust it

A recruiter in a white sweater studies a ranked candidate list on her laptop at a desk in a bright, glass-walled office, checking why each person matched before shortlisting

AI candidate matching is the technology behind the match percentage or ranked shortlist that some applicant tracking systems (ATS) show next to each candidate. When a list arrives already sorted, the useful question is what put one person above another.

AI candidate matching software compares a candidate's profile with a role's requirements and returns a score or ranked list. The score reflects how closely a record fits the criteria the tool was given, which is a different question from how well the person would do the job.

Key takeaways

  • A match score measures fit against the criteria and data the tool was given, not how someone will perform in the role.
  • Keyword matching, semantic matching and skills inference each pass over different candidates, so one person can score differently across tools.
  • Running a tool on a role you have already filled shows more about its ranking than a vendor demo can.
  • Research on biased language models and on reviewers who follow AI picks supports sampling low scores yourself.

How AI candidate matching compares a candidate with a role

Every match has two sides: the candidate's record and the role's criteria. A tool first turns resumes, profiles, and notes into structured fields such as titles, dates, skills, and locations. That step is called resume parsing, and an error there carries through to every score that follows.

The role side comes from a job description, an intake form, or a plain-language search. AI candidate-matching software measures overlap in one of three ways, or combines them, and the mix shapes who rises to the top.

Keyword matching

Keyword matching looks for the exact terms in your criteria inside the candidate's record. Because you can see which terms a candidate matched, it suits hard requirements such as a named license or certification. Its weakness is that people who describe the same work differently can drop down the list.

Semantic matching

Semantic matching compares meaning rather than exact wording. Many tools do this with language models that convert text into numerical representations, placing similar meanings close together. A title like "talent partner" then lands near "recruiter" even though the words differ.

Consider a search for a customer success manager. A strong candidate titled "account manager" whose work covers onboarding and retention could be missed by keyword matching and surfaced by semantic matching. The reverse can also happen: someone may rank highly for a skill that sounds close but differs in practice. These matches are harder to trace, so they're worth reading before you advance anyone.

Skills inference

Skills inference credits candidates with skills they didn't list, based on titles, employers and described work. A data engineer who mentions building pipelines may be credited with SQL even if the word never appears. This helps with thin records and fits skills-based hiring, but an inferred skill is still an estimate, and not every tool shows which skills it inferred.

The table below sets out three ways AI candidate matching works and where each can fall short.

MethodWhat it comparesWhere it helpsWhere it misses
Keyword matchingExact words in the profile and the criteriaPrecise requirements such as a named certificationCandidates who describe the same work in other words
Semantic matchingMeaning of titles and skillsRelated titles and varied industry vocabularySkills that sound close but differ in practice
Skills inferenceSkills implied by titles and past workThin profiles with little detailSkills the person never used or no longer has

Weights, must-haves and filters

Tools also apply weights that decide how much each criterion counts. If every line of a job description carries equal weight, a candidate with several nice-to-haves can outrank one who meets the most important requirement. Writing criteria that reflect the role rather than a wish list is where an ideal candidate profile helps.

Filters remove candidates outright instead of lowering their score, so you never see who was excluded. Location, work authorization, and relocation are frequent filters, and each depends on current data. Treating location as a weight early in a search, where a tool allows it, keeps people with outdated records visible.

What a match score shows and what it can miss

A match score measures how well a record fits the criteria and data the tool was given, often relative to other candidates in the same search. It doesn't predict job performance, and scores from different tools sit on different scales.

One candidate can score differently across tools for reasons you can trace. One tool may weight titles heavily while another leans on skills, and each may read a different version of the same record. A low score can reflect missing data or unfamiliar vocabulary as much as a weak fit. For that reason, the order within one search often says more than a fixed cutoff.

The most useful score comes with an explanation of which criteria the candidate met, which they missed, and which skills were inferred. That lets you judge the reasoning and explain a shortlist to a hiring manager or client without pointing at a percentage.

See why each candidate matched

Book a demo

Where candidate matching is used

Candidate matching shows up at three points in a recruiter's work, and the data behind the score changes at each. SHRM's 2026 research, a survey of 1,908 HR professionals, found that recruiting leads AI use in HR. Screening and matching resumes ranked among the most common uses, in a survey of in-house HR rather than agencies.

Applicants to one open role

Here, AI candidate matching ranks applicants, and the main risk is a qualified applicant who ranks low and never gets opened. Varied resume formats strain parsing. This use overlaps heavily with AI resume screening.

Past candidates in your own database

Rediscovery searches your existing records for people who fit a new role. Older records may carry an outdated title, city, or skill set, and a matcher can only score what the record says. Steady talent pool management gives rediscovery better data. In-house teams can add context from the human resources information system (HRIS), within whatever their data policy allows.

External profiles when sourcing

When you source, the tool searches profiles of people who haven't applied, and you have less control over what each record contains. AI sourcing tools vary in how much data they cover and how fresh it is. A strong match here is a reason to reach out, not a sign the person wants to move.

Agency recruiters may match one candidate against several clients' roles, so saving criteria per client keeps priorities separate. When a client asks how candidates were selected, it's easier to share a ranking explanation than a score.

Where matching goes wrong

When AI candidate matching produces a poor shortlist, the cause often lies in the inputs rather than the method. Thin or outdated records, parsing errors and criteria copied from a padded job description all distort scores.

Proxies are a quieter problem. A criterion that looks neutral, such as a commute radius or a list of preferred schools, can track characteristics you aren't allowed to consider. Checking that each heavily weighted criterion is needed for the job helps catch this.

In a 2024 University of Washington study, three open-source language models ranking resumes favored white-associated names 85% of the time and Black-associated ones 9%. Those were research models, not commercial hiring tools, so the finding is a reason to test the tools you use, not proof about any product.

A second study looked at people reviewing AI output. In a 2025 lab study with simulated AI, 528 online U.S. participants showed little bias without recommendations but mirrored the AI once recommendations appeared. With severely biased recommendations, choices followed the AI around 90% of the time, according to the same study.

The participants weren't recruiters at work, but the finding suggests reading some profiles before their scores to keep part of your review independent. The same concern about human oversight runs through AI in hiring generally.

Ways to test candidate matching software before you rely on it

The most informative test is to run AI candidate matching on a role you filled recently, because you know who the strong candidates were. Use the criteria you had at the start, not what you learned later. If the tool ranks your eventual hire low or misses them, you have learned something a demo wouldn't show.

  • Run a known search: Compare the tool's top results for a filled role with your shortlist, noting who it surfaced and who it buried.
  • Ask for explanations: Ask the vendor to show why each candidate matched and which skills were inferred rather than stated.
  • Adjust the weights: Mark must-haves and nice-to-haves, then check whether the ranking moves the way you'd expect.
  • Compare with Boolean: Run the same role as a Boolean search and look at who each approach finds that the other misses.
  • Sample low scores: Read a handful of profiles from the bottom of each list to catch strong candidates the tool undervalued.
  • Recheck after updates: Save your known search and rerun it when the vendor changes its model or scoring, since rankings can shift.
  • Plan for candidates: Decide what you'll tell candidates about automated ranking and how someone can ask for a human review.

Rules that can apply to matching tools

Rules on automated hiring tools depend on where the role is based, so check with counsel for each location. New York City's Local Law 144, enforced since July 5, 2023, requires a bias audit within one year of use, public audit information, and candidate notices.

Illinois Public Act 103-0804, effective January 1, 2026, bars AI that has a discriminatory effect and uses zip codes as proxies for protected classes. The same act requires employers to notify employees when they use AI. California's Civil Rights Council rules, effective October 1, 2025, require employers to keep employment records, including automated-decision data, for at least four years.

At the federal level, the U.S. Equal Employment Opportunity Commission (EEOC) recordkeeping summary says covered private employers keep application forms and hiring records for one year. Ask your vendor what ranking data it stores and for how long.

Check a search you know before you trust the score

A good first test is a role filled this year whose hire had an unconventional background, because it shows how the tool handles unfamiliar titles. If the ranking tracked your judgment, you can use it as a reading order. If you can't see why it ranked candidates as it did, treat the score as one input until the vendor shows you.

Check every match against your own criteria

Book a demo

Frequently asked questions

Should hiring managers see match scores?

That depends on whether they also see the explanation behind each score. A bare number can shape a manager's view of a candidate before they read the profile. Sharing why a candidate matched gives them reasoning they can agree or disagree with.

Does AI candidate matching work for niche or senior roles?

It can, but small pools and unusual titles give the tool less to work with. Senior profiles often describe scope and outcomes rather than listing skills, which keyword matching can miss. Reviewing more of the list by hand makes sense when the pool is small.

Can a match score be the only reason to reject a candidate?

Relying on a score alone removes the human review that catches parsing errors and missing data. Some state and local rules also attach audit, notice or recordkeeping duties to automated tools. Check with counsel before setting any automatic rejection threshold.

About the author

Mihailo Bozic
Mihailo Bozic

Founder & CEO @ Rotto

Founder & CEO of Rotto, building tools that help tech recruiters source better candidates, faster.

Related articles

Rotto

Built for recruiters who work from a brief, not a boolean.

rotto.com