What we measure, how we measure it, and what we cannot know.

A model reads the whole posting the way a recruiter would, and answers a fixed set of questions about it: what does this “job” ask you to do with money, cheques, parcels, crypto or identity documents; is anyone actually hiring for a seat; what does it require. It is run at temperature zero, so the same text gets the same read, and it is asked for the words in the posting behind every claim — a quote we cannot find in the text is dropped before you see it.

Then the posting is read a second time, by a separate reader that never sees the first read’s conclusion and is asked only one question: what does this require the applicant to do? It is told to assume the company name is fake, and to ignore the About section, the benefits, the salary band and the equal-opportunity paragraph entirely — those are free to add, and a careful scammer adds them. It reports money you are asked to pay, money or goods passed through you, identity documents requested before anyone has spoken to you, applications routed somewhere that is not the employer’s own site, and a stated absence of any live interview. It can only raise the alarm; it has no way to clear a posting, so a mistake there costs a false warning rather than a missed fraud. Alongside it, deterministic checks run on every scan and every tier — among them whether the hiring link belongs to the company named in the posting at all. These override the model in one direction only: if one fires, the verdict is scam whatever the model concluded, and none of them can ever approve a posting. Everything that fires is shown to you with the phrase that triggered it. Green signals are never shown beside a red one, and are withheld entirely on any posting we have not fully cleared.

Why both: rules alone were beaten by rewording — a list of personal data padded with six more words walked past a proximity rule. A model alone can be talked round. The read handles wording; the rules hold the line on mechanics.

The same read runs the culture decoder, which translates stock phrases into what they commonly signal — “fast-paced environment” as a marker for chronic understaffing, “we’re like a family” as one for blurred boundaries. A translation of a phrase, not a judgment of the employer.

This is the part we consider the actual product. Rather than scoring your resume against a bag of keywords, we extract the posting’s individual requirements and check each one against your resume separately. Every requirement is classified by category (experience, skill, domain, education, certification, logistics), by importance (must-have or preferred), and by verdict — covered, partial, or missing.

About the ATS score specifically

We do publish a match percentage, because it is genuinely useful for ranking your own resume versions against one job and because it is what people search for. It is the requirement table counted — not a keyword list — so “Postgres” and “PostgreSQL” are the same thing and a satisfied “or” is not a gap. It is always computed against a specific posting; a resume score with no job attached measures nothing.

Where our published numbers come from

Statistics we cite in marketing and on the blog are attributed to named third parties. We do not publish first-party statistics from our own scan data yet — our scan volume is not large enough for that to be honest, and we will say so on the day it changes.

• Ghost-job prevalence (~43% of postings): reporting from NPR, Forbes, and ZipRecruiter research.
• Job-scam losses: US Federal Trade Commission consumer reports and the FBI's IC3 annual report.
• ATS usage among large employers and resume filtering rates: Harvard Business School's "Hidden Workers" (2021) and Jobscan's ATS usage research.
• Employer detection of AI-generated applications: published survey work on recruiter screening practices.

Questions about the method