# AI job search: how it works and where it fails

> AI job search is software that reads your resume and a pool of postings, then ranks the postings by how closely they match. It finds roles you would not have searched by title, and it fails on thin postings, stale data and rankings it cannot explain. Six questions tell you whether a tool deserves your time.

Source: https://zorlya.com/blog/ai-job-search-how-it-works-what-it-gets-wrong-and-how-to-judge-a-tool

By Zorlya. Published 2026-10-27. Research and data. 5 min read.

AI job search means software reads your resume and a large pool of postings, then ranks the postings by how closely they match you. It works well for finding roles you would not have searched for by title. It works badly when a posting says little, when the data is stale, or when the tool will not say why it ranked a job where it did.

An AI job finder is a tool that does this ranking for you. AI job matching is the ranking step inside it. This post explains how the step works, where it goes wrong, and how to test any tool, Zorlya included, before you trust its order.

## How does AI job search work?

An AI job search tool does three things in order: it collects postings, it turns text into numbers, and it ranks.

First, it needs postings. A tool can only rank what it holds, so the source list matters more than the model. Zorlya reads public job boards and the systems employers post on, 130+ sources in all, and each posting shows its source and links to the original ([where the jobs come from](/blog/where-the-jobs-come-from)).

Second, it converts your resume and each posting into embeddings, lists of numbers that place text by meaning. Texts about similar work land close together even when they share few words. Zorlya uses lists of 1,024 numbers for this ([how embeddings match your resume to jobs](/blog/how-embeddings-match-your-resume-to-jobs)).

Third, it ranks. Closeness alone is a blunt signal, so better tools add checks that closeness cannot make: which skills the posting names, what it pays, where the job is. A language model may then read the shortlist and write a reason for each rank.

## What does AI job search get wrong?

It gets the same few things wrong across tools, and knowing them lets you correct for them.

- **Thin postings rank oddly.** A posting of two sentences has little meaning to place, so its score says little. A long posting full of generic company text can drift toward everyone.
- **Missing data is guessed.** Most postings publish no salary. A tool that treats unknown pay as an average pulls good jobs down and weak jobs up. Zorlya leaves salary out of the total when neither you nor the posting gives a figure, and scores it 45, a little under the middle, when you set a floor and the posting hides its pay ([how Zorlya scores a job](/blog/how-zorlya-scores-a-job)).
- **Stale postings look live.** A ranking cannot tell that a role was filled last week unless the source says so. Dates on postings are also unreliable ([why a job's posted date is unreliable](/blog/why-a-jobs-posted-date-is-unreliable)).
- **Your resume decides the ceiling.** A resume that lists skills without the work behind them gives the tool little to match. The ranking reflects what you wrote, not what you can do.
- **Coverage has holes.** No tool reads every employer. Some employers post only on their own page, and some boards forbid automated reading ([what a job index holds and what it cannot tell you](/blog/what-a-job-index-holds-and-what-it-cannot-tell-you)).
- **A high rank is not a prediction.** It says the posting resembles your background. It cannot see who else applied or what the hiring manager wants that the posting left out.

## How do I judge an AI job finder?

Run any tool through these six checks before you let its order decide where your time goes.

1. **Does each job name its source?** You should be able to see where a posting came from and open the original. If a tool will not say, you cannot check anything else.
2. **Does it link to the employer's own posting?** A copy can be shortened or out of date. The original is the one to apply through.
3. **Can you see why a job scored as it did?** A bare number cannot be argued with. Look for the parts of the score and a reason written for that job, one that names something from your resume.
4. **What does it do with missing data?** Ask what happens to a job with no salary or no stated location. The honest answers are "left out" or "shown as unknown", not "scored as average".
5. **Does it say what it does not cover?** A tool that claims every job everywhere is wrong about the claim.
6. **Does it hold up on a job you already know?** Take a role you have applied to or turned down. Check whether its score and reason agree with what you learned.

If a tool fails the first three, treat its ranking as a list of ideas and nothing more.

## How does Zorlya do AI job matching?

Zorlya scores each job from 0 to 100 on four parts: resume fit, skills match, salary fit and location fit. Each part is shown. When a posting states a seniority level, how far it is from yours counts too, at 10% of the total. When you refresh your matches, each match Zorlya keeps gets two or three sentences written for that job and your resume. The default weights are 40%, 30%, 20% and 10%, and a track, which is a kind of role such as backend engineering, can change them.

Plain code does the arithmetic, such as skills coverage and salary against your floor. A language model is used where judgement or writing is needed, such as reading a shortlisted posting against your profile ([when Zorlya uses a language model](/blog/when-zorlya-uses-a-language-model)). The [how it works](/how-it-works) page shows the steps in order.

## A short test you can run this week

Pick five roles you could do. Look for each one in the tool you are testing, and note whether it appears, how it ranks, and whether the reason matches your own view. Two or three misses on roles you know exist tells you more about coverage than any claim on a website. Repeat the test with one role that is a poor fit. A good tool ranks it low and says why.

## Questions people ask

### Is AI job search accurate?

It is accurate about resemblance and weaker about everything a posting leaves out. Expect good suggestions for roles close to your background and unreliable ones when a posting is short or its facts are missing. Check the reasons before you trust a rank.

### Can an AI job finder find jobs that job boards miss?

It can surface roles you would not have typed a search for, because it matches meaning and not exact words. It cannot find jobs that are not in its sources. The reach of the source list sets the limit.

### Will AI job matching replace reading the posting?

No. A score tells you where to look first. You still need to read the posting for duties, deadlines and anything the score cannot see ([how to read a job posting](/blog/how-to-read-a-job-posting)).

### Does the tool keep my resume?

Ask this of every tool and read its privacy page. For Zorlya, [what happens to your resume in matching](/blog/what-happens-to-your-resume-in-matching) lays out what is stored and why.

## What we know and what we do not

Zorlya reads 130+ sources, shows the source and the original link on every job, and shows the four parts of every score, with a reason for each match you refresh. It does not read every employer, it does not predict whether you will be hired, and it cannot tell you a role is filled until a source says so. We have not measured how often Zorlya's order agrees with a recruiter's, and we do not publish an accuracy figure.