TL;DR — A job matching algorithm ranks listings against your profile so you see the most relevant first. Three kinds are in use: keyword matching (fast, literal, misses synonyms), embedding similarity (understands meaning, ignores hard requirements) and LLM scoring (reads both sides, costs more). Most sites use the first; the ranking you get is only as good as the profile you gave it.
A job matching algorithm answers one question: given everything known about you and thousands of open roles, which ones go at the top? Every job site runs one, whether or not it calls it AI, and its choices decide what you ever see — the roles it buries are, for you, roles that do not exist.
The three approaches
Keyword matching. Your resume and the posting are compared as bags of words. If both contain "Kubernetes", that is a point. Cheap, instant, and how most job boards actually work.
What it gets wrong: it is literal. "K8s" does not match "Kubernetes". A "Growth Marketer" does not match a "Demand Generation Manager" doing the same job. And it cannot tell a requirement from a mention — a posting that says "no Python needed" scores as a Python match.
Embedding similarity. Both texts are converted into vectors — long lists of numbers positioned so that similar meanings sit close together — and ranked by distance. This solves the synonym problem: "K8s" and "Kubernetes" land near each other without anyone writing a rule.
What it gets wrong: it measures overall similarity, not eligibility. A senior role and a junior role in the same field look nearly identical to it, as do a US-only role and a global one. Similarity is not the same as "you could get this job".
LLM scoring. A language model reads your profile and the posting together and produces a score with reasons. It handles the things the first two miss: seniority, hard requirements, "5+ years" versus your three, a language you do not speak.
What it gets wrong: cost and consistency. Reading every posting against every profile is expensive, so systems that do this must first narrow the field — usually with one of the two methods above. And a model asked the same question twice can answer differently unless the system is built to prevent it.
Most serious systems combine them: a cheap method narrows thousands of roles to dozens, then an expensive method ranks those. Ours does exactly that — embedding similarity to shortlist, then a language model to score each shortlisted pair 0–100 with a written reason.
Why matching so often feels wrong
Three causes, in order of how often they are the real one:
1. Your profile is thin. Every method above works from what you gave it. A two-line profile produces two-line relevance. 2. The system is optimising for something else. Job boards earn from employers. "Relevant to you" and "roles the employer paid to promote" are not the same ranking, and few sites tell you which one you are seeing. 3. Hard requirements are being ignored. This is the specific failure of similarity-only matching, and it is why you get shown roles you are plainly not eligible for — wrong seniority, wrong country, wrong work authorisation.
How to tell which one you are using
- Does it explain itself? A system that tells you why a role scored well is doing more than keyword counting — and lets you correct it when the reason is wrong.
- Does it respect hard filters? Set a salary floor and a location. If roles that violate both keep appearing, ranking is not really personalised.
- Does feedback change anything? Dismiss five roles of a type. If the sixth is identical, nothing is learning.
- How fresh is the pool? The best ranking over a stale index still shows you filled jobs.
Frequently Asked Questions
Is a job matching algorithm the same as an ATS?
No. An ATS is the employer's software for managing applicants. A matching algorithm is the job seeker's side: which listings you get shown, and in what order.
Why do job sites show me irrelevant jobs?
Usually because they rank by text similarity without checking eligibility, or because paid placement outranks relevance. A role you cannot legally take or are plainly too senior for is the signature of similarity-only matching.
Can I improve my matches?
Yes, and it is the highest-leverage thing available to you: give the system more to work with. A complete profile with real titles, technologies and a stated salary floor changes the ranking far more than any tweak on the site's side.
What is a good match score?
It depends on the scale, and a score is only useful if it comes with a reason. On our 0–100 scale, 60+ means the core requirements line up; the reason attached to it tells you whether the gap is something you can address in a cover letter.
RemoteHunt shortlists by meaning and then scores each role 0–100 against your resume with a written reason — so you can see why, and disagree. Start for free.