You export one PDF, run it through two CV checkers on the same afternoon, and get 62 percent from one and 89 percent from the other. Nothing about the file changed in between. That gap is not a bug in either tool, and understanding where it comes from is worth considerably more than either number.
Four different questions wearing the same badge
The phrase "ATS checker" covers at least four products that do unrelated work. They get lumped together because they all accept a CV and hand back a verdict, and the verdict almost always looks like a score out of 100. Underneath, they are not answering the same question, so there is no reason for their answers to match.
Parse fidelity: can a machine read this file at all
A parser test pushes your file through text extraction and shows you the structured fields that came out the other side: name, contact details, employers, job titles, dates, education, skills. That is an objective question with a checkable answer. Either your most recent job title survived extraction or it did not.
The failure modes here are mechanical and dull, which is exactly why they are fixable. Text baked into an image. A two-column layout that interleaves your responsibilities with your language skills. Dates parked in a header that extraction drops. An icon standing in for a field label, so a phone number arrives with nothing attached to say what it is. None of that is a matter of taste, and a parse test is the only tool in this category that produces a fact.
Keyword matching: does this document echo this ad
Paste a vacancy, upload the CV, receive an overlap percentage and a list of terms the tool believes are missing. This score has a property people forget within about ten seconds of seeing it: it is entirely relative to the ad you pasted. The same CV scores differently against two adverts for the same role at two employers, because the wording differs.
It is also easy to move. Copy a dozen nouns out of the ad into a skills block and the number climbs, without you becoming a stronger candidate by any measure anyone cares about. A score you can raise that cheaply is not a grade. It is a word-overlap statistic, useful in exactly that narrow way and misleading everywhere else.
All-in-one optimizers: a blend with the recipe hidden
The most familiar category takes parse checks, keyword overlap, formatting rules and readability heuristics, then mixes them into one figure. Many add opinions on top: bullet length, action verbs, whether you quantified anything, how long the file is.
Those weightings are a product decision and they are rarely published. One tool treats a missing skills section as a serious fault; another barely registers it and docks you for a second page instead. Both hand you a number out of 100, and neither shows the split. So when two optimizers disagree, they are frequently not disagreeing about your CV at all. They are disagreeing about what a CV is for, and the single percentage hides that argument rather than settling it.
The employer's own preview: the only output that binds
When an application portal parses your upload and pre-fills its form with whatever it extracted, you are looking at the real output of the real system for that specific employer. It carries no score. It is also the only screen in this whole category with any authority, because it is the version of you that recruiter searches inside that company will actually run against.
Why the numbers diverge
Once you separate the categories, the disagreement stops being mysterious:
- Different questions. A parse test and a keyword scorer are not measuring the same property, so comparing their outputs is a category error.
- Different inputs. One needs a job ad and one does not. A score that depends on pasted text is not a property of the document.
- Undisclosed weights. Every all-in-one tool has a weighting model, and yours is not visible.
- Different extraction engines. Two tools can genuinely read the same PDF differently, particularly with tables, text boxes and unusual fonts.
- A house style to sell. A tool that also sells templates has a structural reason to score anything unlike its templates a little lower.
If two checkers disagree, both are usually right about the question they asked. The mistake is reading either answer as a verdict on the document.
The percentage is the least useful thing on the screen
A single number compresses everything you actually need into a figure with no units. Scroll past it. The parts underneath are where the value sits.
Read the parsed output first, field by field, and compare it against what you wrote. Then read the missing-terms list as a list, not as a deficit: which of those terms are true of you and simply absent from the file, and which are not true of you at all. The first group is an edit. The second group is a signal about whether this job is a fit, which is more useful than any score. Finally, read the specific flags. "No dates detected in the experience section" is actionable. "Formatting score: 71" is not.
Agreement is the part worth acting on
Here is the useful consequence of all this. Because these tools measure different things with different engines, agreement between them is meaningful in a way that any single score is not.
If three checkers independently fail to extract your employment dates, your file is broken. Not suboptimal, not unfashionable, broken, and the fix is structural rather than cosmetic. The same goes for job titles that come back empty, an employer name absorbed into a bullet point, or a skills section that no tool can find. That is free, triangulated evidence about the mechanics of your document, and it is the main thing multiple tools are good for.
Note the converse too. When one tool flags something and three others do not, you have found that tool's house preference, not a defect.
What none of them can see
No checker knows whether you are a credible candidate for the role. It cannot tell whether your last position was a stretch or a step down, whether your results are impressive for your market, or whether the gap in 2024 needs explaining.
It also cannot tell you whether a human finds the document persuasive. Recruiter attention is the actual constraint in most hiring processes, and a CV stuffed with terms lifted from an ad reads badly to the person the stuffing is meant to reach. There is a real tension between a high keyword score and a document someone enjoys reading, and no tool surfaces that tension because no tool measures the second half.
Most importantly, none of them know how that employer has configured its system: which fields are searched, which knockout questions run before a human ever sees the file, how the recruiter phrases their own queries. That configuration varies from company to company, and it is invisible from outside.
A sequence that holds up
- Fix parsing first, because it is the only objective part. Run a parse test, read the extracted fields, repair anything that came back wrong or empty. Stop when the extraction matches the document. Starting from a structured source helps here, which is part of what a LinkedIn-to-CV tool like Postulit does, though no generator can tell you whether the content is convincing.
- Treat keyword scores as a per-application checklist. One ad, one pass, keep the terms that are honestly true of you, ignore the number attached.
- Never trade readability for points. If an edit raises a score and makes a sentence worse, the edit is wrong. The score has no vote at the interview.
- Re-test once, then stop. Iterating against a scoring model you cannot see is optimising for the model, not for the job.
So when two checkers disagree, do not pick the one you prefer, and do not average them. Ask what each one measured, act on the mechanical failures both of them found, and spend the time you save writing better bullet points.