Candidate matching software: how AI matching works and how to evaluate it

By Drippay, Inc.Published 8 min read

Candidate matching software takes a job and returns a ranked list of people who may fit it, drawn from your own ATS, an external market of profiles, or both. Vendors rarely disclose exactly how the ranking is produced, so the useful questions are which pool it ranks, what inputs it uses, whether a recruiter can see why someone ranked where they did, and how it performs on your reqs.

This guide covers what matching does, the techniques vendors describe, what match quality depends on, a fair way to evaluate tools, and a checklist for the buying decision.

What is candidate matching software?

Matching sits between search and screening. Search returns records that satisfy a query, possibly ranked. Screening evaluates one candidate against a role, often with questions or an assessment. Matching takes a role and a candidate pool and orders the pool by estimated fit, so a recruiter reviews a short list rather than everything the query returned.

The category includes matching built into an ATS or CRM over the firm’s own records, sourcing platforms that also rank an external market, and rediscovery tools that match a new req against past applicants. Many products do more than one of these, so ask which pool a ranking came from before comparing rankings.

Related: Candidate rediscovery: re-engaging the candidates already in your ATSCandidate sourcing software: how to evaluate sourcing tools

Techniques vendors describe

Product pages describe inputs and outputs more than internals, and the same product may combine several approaches. The table lists what vendors say and what each approach implies for a buyer.

Approaches described by candidate matching vendors and their trade-offs
ApproachWhat the vendor describesImplication for the buyer
Search built from the job descriptionBullhorn describes auto-building a search from the job description with a keyword library that suggests terms, returning top matches with relevancy scoresThe req text is the query; a vague req produces a vague ranking
Joint analysis of profile and requirementsLoxo describes AI that analyzes candidate profiles and job requirements to surface matches; Spott describes semantic search matchingLikely handles synonyms and adjacent experience; ask how a single result is explained
Outcome-informed scoringAlex describes relevance scores combining skills, experience, location, and performance data from past interviewsUses history as a signal; ask what history is used and how it is refreshed
Continuous matchingBullhorn describes finding and ranking candidates continuously, and Alex describes recommending candidates for other open rolesResults arrive as notifications; decide who reviews them and how outcomes are logged

Vendor descriptions as of September 2026 from the pages linked under Sources. Embedding-based similarity is one possible technique behind semantic matching; vendors generally do not publish their architecture.

What match quality depends on

Both the product and the inputs matter, and only the inputs are under your control. Bullhorn’s Amplify guidance says matching performance improves significantly when candidate and job data is standardized and up to date, and recommends title and skill normalization, standardized employment history, starting with a small, high-quality pool of recently engaged candidates, and deploying matching first on repeatable roles and on jobs the team cannot prioritize.

Three things to check in your own workflow: how specific your reqs are as written, how consistent your titles, skills, and employment fields are, and whether interview and placement outcomes flow back into the tool at all. You can measure the data effect directly by running the same tool on a role family before and after a normalization pass.

Related: Recruitment database search and maintenance

How to evaluate matching tools fairly

Demos run on the vendor’s data. Run the test on yours, in two parts.

  • Retrospective: pick several reqs you filled in the last year. Load each req as it was written at the time and rank records as they stood before the hire, without placement notes or post-hire updates the tool could not have seen. Treat the people you interviewed and placed as reference cases, not a perfect answer key; a highly ranked candidate you never spoke to may have been a real miss on your part.
  • Independent relevance review: have a recruiter who did not work the req rate the top twenty for each tool blind, marking clear misses on location, credential, or seniority. Compare the ratings across tools, and check whether each tool can explain its misses.
  • Prospective: run the shortlisted tools on live reqs for a few weeks. Count how many ranked candidates are submitted and interviewed, and how much recruiter time the ranking saves or costs.

A worked example: a clinical team tests two tools on filled travel-nurse reqs. One ranks the placed nurse highly on most reqs but surfaces candidates without the required state license; the other never misses a license because licensure is a hard filter, but its rankings otherwise look weaker in the blind review. The test has told you exactly what to ask each vendor: a license filter from the first, and a ranking explanation from the second.

Human review and job-related evidence

A ranking becomes a decision when it determines who a recruiter looks at. Keep a person between the ranking and any rejection, score on evidence tied to the job (credentials, experience, location, availability) rather than on proxies, document the criteria used for each req, and ask vendors what documentation they provide about how their matching is tested for fairness. Requirements for automated hiring tools vary by jurisdiction; confirm what applies in your markets with counsel.

Buying checklist

Take these into every demo.

  • Which pool does it rank: our ATS, an external market, or both? Can we restrict to one?
  • Can a recruiter see why a candidate ranked where they did, in terms they could repeat to a client?
  • Which criteria can be hard filters, and which are scored?
  • Where do results appear, and who reviews them?
  • How do interview and placement outcomes flow back, and can the vendor demonstrate a ranking that changed because of them?
  • What data leaves the ATS, where is it stored, and how is it deleted when we leave?
  • What is the pricing unit, and what does it cost when the database doubles?

Discuss this workflow with dreach

dreach is a managed service for staffing firms, run by hand in the pilot. It watches the employers your firm knows and the ones it can reach through a warm path, such as a current or past client, a person your firm placed who now works at the hiring company, or a contact your team authorizes. It flags new openings with the source and date, confirms the hiring manager to contact, and shows the warmest way in. For a live role, it ranks people your firm already knows from its own ATS, past placements, and network your firm authorizes, and shows why. Your team makes every contact; dreach contacts no one. Scope, data access, integration requirements, and availability are confirmed before starting.

Want to test this against a real req? Bring one and we will walk through your existing records and what fit and relationship evidence actually holds up.

Discuss your candidate workflow

Candidate matching software is a ranking layer whose quality depends on the product and on the reqs and records you give it. Test on your own filled reqs without post-hire leakage, add a blind recruiter review and a prospective run, insist on explainable results and hard filters for non-negotiables, and keep a person between the ranking and any decision.

ask away

Questions

What is candidate matching software?

Software that takes a job and returns a ranked list of candidates who may fit it, drawn from your own ATS, an external market of profiles, or both. Vendors describe building searches from the job description, analyzing profiles and requirements together, and in some cases using past interview data in the score.

How does AI candidate matching work?

Vendors rarely publish their architecture. Descriptions on product pages include auto-building a search from the job description with suggested terms, analyzing candidate profiles and job requirements together to rank by fit, and combining skills, experience, location, and past interview data into a relevance score. Semantic similarity from language models is one possible technique behind these descriptions.

Is AI candidate matching accurate?

It varies by product and by your data. Vendors advise that matching improves when titles, skills, and employment history are standardized and reqs are specific. The reliable way to know is to test on reqs you already filled, without post-hire information, add a blind recruiter review of the rankings, and then run the tool on live reqs for a few weeks.

What is the difference between candidate matching and candidate rediscovery?

Rediscovery is a use case: filling a new req from people already in your database. Matching is a technique that rediscovery tools, ATS features, and sourcing platforms use to rank candidates against a job. A rediscovery tool matches against your history; a sourcing tool matches against an external market as well.

Does dreach match candidates to jobs?

dreach is a managed service for staffing firms, run by hand in the pilot. It watches the employers your firm knows and the ones it can reach through a warm path, such as a current or past client, a person your firm placed who now works at the hiring company, or a contact your team authorizes. It flags new openings with the source and date, confirms the hiring manager to contact, and shows the warmest way in. For a live role, it ranks people your firm already knows from its own ATS, past placements, and network your firm authorizes, and shows why. Your team makes every contact; dreach contacts no one. Scope, data access, integration requirements, and availability are confirmed before starting.