Home/How the score works
How we calculate the ATS score
The number is a weighted average of 21 sub-scores, computed in your browser from the text of your file. There is no model in the scoring path, no randomness, and no server call. This page documents the arithmetic down to the rounding, so you can check the number instead of trusting it.
Published 27 October 2026. If a weight changes, this page changes with it — the numbers below are read from the same rule definitions the app ships.
Why publish this at all. "ATS score" is sold everywhere with no method attached. A score you cannot audit is a score you cannot act on: you get a number, you get told to improve it, and you have no way to tell whether the advice is connected to anything real. The weights below are opinions — we chose them — but they are stated opinions, and you can disagree with them specifically rather than vaguely.
The arithmetic, in four steps
- Each rule returns a sub-score from 0 to 100. A rule looks at one specific property of the parsed document — say, whether every bullet starts with a verb — and grades it. The sub-score is rounded to a whole number.
- That sub-score is scaled by the rule's weight. Points = sub-score × weight ÷ 100, rounded to one decimal. A rule weighing 12 that scores 50 contributes 6.0 points; a rule weighing 2 that scores 50 contributes 1.0. This is the only place the weights do any work.
- The points are divided by the maximum available, not by a fixed 100. Score = (sum of points ÷ sum of applicable weights) × 100, rounded to a whole number.
- Rules that could not run are removed from the denominator. The four job-description rules are skipped when no job description is supplied, and they contribute 0 points out of 0 available — they are excluded, not counted as failures.
A job description counts as supplied when the pasted text is longer than 30 characters. Below that it is treated as absent, because a company name alone is not a job description and scoring against it would produce a confident, meaningless match.
The 21 rules and their weights
Five groups, 21 rules, 101 weight in total. The order below is the order they run in.
| # | Rule | Group | Weight |
|---|---|---|---|
| 1 | Is there any text in the file? | Parsing | 4 |
| 2 | Two-column layout, tables, text boxes | Parsing | 3 |
| 3 | Are the section headings the conventional ones? | Parsing | 2 |
| 4 | Contact details buried in a header or footer | Parsing | 1 |
| 5 | Page count against years of experience | Structure | 5 |
| 6 | Are contact details complete and parseable? | Structure | 5 |
| 7 | Are date formats consistent? | Structure | 4 |
| 8 | Unexplained employment gaps | Structure | 3 |
| 9 | Photo, age, marital status, nationality | Structure | 3 |
| 10 | Does every bullet start with a verb? | Bullet content | 6 |
| 11 | Quantified results | Bullet content | 8 |
| 12 | Vague filler phrases | Bullet content | 5 |
| 13 | Are the skills concrete tools? | Bullet content | 5 |
| 14 | Hard-skill keyword coverage | Job match | 12 |
| 15 | Job-title alignment | Job match | 8 |
| 16 | Industry terminology hit rate | Job match | 6 |
| 17 | Soft-skill keyword match | Job match | 4 |
| 18 | Concrete outcomes and scale | Detail and risk | 5 |
| 19 | Certifications and licences | Detail and risk | 3 |
| 20 | Spelling | Detail and risk | 4 |
| 21 | Keyword stuffing, including hidden text | Detail and risk | 5 |
Group totals: parsing 10, structure 20, bullet content 24, job match 30, detail and risk 17. Seventeen of the 21 rules run without a job description; only rules 14–17 require one.
The largest single weight is rule 14, hard-skill keyword coverage, at 12. That is deliberate: in the systems we are modelling, the recruiter's first action is a keyword search over the parsed index, so coverage of the posting's actual required skills moves a resume further than any bullet rewrite. The smallest is rule 4, at 1 — contact details trapped in a page footer are a real failure, but a rare one, and it is already partly covered by rule 6.
Why the weights add up to 101, not 100
They were chosen as relative importance, in whole numbers, before anyone added them up. The total happens to be 101. This changes nothing, and here is why: the score divides points by the maximum available, so a single perfect pass still returns exactly 100. The 101 only means every individual point is worth marginally less than one percent.
The alternative — rescaling the weights to sum to 100 — would have meant editing every weight to fix a cosmetic property, and would have made the published numbers differ from the numbers in the code. Matching the code is worth more than a round total.
Without a job description the applicable maximum is 71 (101 minus the 30 carried by rules 14–17), so a resume with no job description is scored out of 71 rather than being capped near 70. A file that is mechanically perfect scores 100 either way — what changes is which failures are available to it.
A worked example you can check by hand
These are illustrative sub-scores, not a real candidate. No job description is supplied, so rules 14–17 are absent and the denominator is 71.
| Rule | Weight | Sub-score | Points |
|---|---|---|---|
| 1 · text present | 4 | 100 | 4.0 |
| 2 · layout | 3 | 100 | 3.0 |
| 3 · headings | 2 | 100 | 2.0 |
| 4 · header/footer contact | 1 | 100 | 1.0 |
| 5 · page count | 5 | 80 | 4.0 |
| 6 · contact complete | 5 | 100 | 5.0 |
| 7 · date formats | 4 | 100 | 4.0 |
| 8 · gaps | 3 | 100 | 3.0 |
| 9 · protected fields | 3 | 100 | 3.0 |
| 10 · verb-first bullets | 6 | 70 | 4.2 |
| 11 · quantified results | 8 | 40 | 3.2 |
| 12 · filler phrases | 5 | 60 | 3.0 |
| 13 · concrete skills | 5 | 100 | 5.0 |
| 18 · outcomes and scale | 5 | 60 | 3.0 |
| 19 · certifications | 3 | 100 | 3.0 |
| 20 · spelling | 4 | 100 | 4.0 |
| 21 · keyword stuffing | 5 | 100 | 5.0 |
| Total | 71 | — | 59.4 |
59.4 ÷ 71 × 100 = 83.66, which rounds to 84. That lands in the second band, "good". Adding each row by hand is the whole audit — there is no hidden term.
What the four bands mean
| Score | Label | Reading |
|---|---|---|
| 85–100 | Excellent | The file parses and the content is findable. Remaining risk is fit, not mechanics. |
| 70–84 | Good | Parses correctly; several content rules are leaking points that a rewrite can recover. |
| 50–69 | Fair | Either a structural problem or a keyword gap large enough to keep you out of searches. |
| 0–49 | Poor | Usually a parsing failure upstream — the content below it rarely matters yet. |
The bands are labels, not probabilities. Nothing in the data supports the sentence "84 means an 84% chance of an interview", and we do not make it.
Why the same file always scores the same
The scoring path contains no language model, no network request and no random number source. The file is parsed in your browser; the 21 rules are ordinary functions over that parse; the arithmetic is arithmetic. Run the same file twice and you get the same number, with the same rules failing for the same stated reasons.
This is the property that makes the score useful as a before-and-after instrument: change one bullet, change one rule, watch one number move. A stochastic score cannot do that — every edit is confounded by the model's variance, and you end up guessing whether the rewrite or the dice moved the number.
The language model is used in exactly one place, and it is not the score: the paid rewrite, which runs after payment and produces text. Its output is not fed back into the scoring, so nothing in the model can move the number you were shown.
One caveat that is not about scoring: the same resume saved by two different exporters can extract differently, because extraction happens before any rule runs. That is a file difference, and it is visible in the app — the parse preview shows what the rules actually saw.
Four ways this number can mislead you
- It is not an employer's ATS score. No vendor's ranking is being modelled here. The number describes whether this file parses cleanly and whether it contains the terms a recruiter would search for. It does not reproduce Workday, Greenhouse or Taleo's ordering.
- The parsing rules are heuristics over our own extractor. Rule 2 does not run your file through every commercial parser. It flags structures that are known to break text extraction in general, which is a strong prior, not a guarantee.
- A high score means findable, not good. Every bullet can start with a verb and every result can carry a number while the experience itself is the wrong experience. Nothing here measures whether a human likes the resume, or whether you can do the job.
- Some failures are not fixable by editing. A scanned PDF with no text layer fails rule 1 for a reason no rewrite addresses: it needs re-exporting. Knockout questions on the application form — work authorisation, years of experience — reject people regardless of what this number says.
See which of the 21 you fail
The free score names every rule you lost points on, with the specific line that triggered it. No account, no email, and the file never leaves your browser.
The rules themselves are documented one by one in what an ATS actually checks, and the pricing comparison against Jobscan and Resume Worded is in the alternative page. Weights and thresholds on this page are read from the same definitions the app ships; if they diverge, that is a bug, not a policy.