My CV had just been redone — clean, two columns, badges, a QR code. I thought it looked great. And that's exactly when the doubt creeps in: great for whom? For me, who built it, or for the recruiter who'll scan it in seven seconds between two meetings?
I'd followed the usual advice, the kind you read on every HR blog. "75% of CVs are rejected by bots before a human sees them." "CVs with numbers get 40% more callbacks." Sentences I'd swallowed without ever clicking the source. So before patting myself on the back, I wanted to know one simple thing: what does science actually say about CVs, and what's just folklore copied from blog to blog?
I pointed AI at it. Not to write my CV — I'd already done that (36 iterations with Claude, the story is here). To dig out the real studies, demand primary sources, and flush out the invented numbers. Three research agents in parallel, one non-negotiable rule: no statistic without a verifiable source, and anything that smells like a myth, you flag it. What came back made me throw out half of what I thought I knew.
The myths I almost optimized for
The worst thing about bad advice is when it's precise. A round number, a source that sounds serious, and boom, you build your CV around it. Three examples you've definitely read, and they're all hot air.
"75% of CVs are rejected by the ATS before a human." False. The line traces back to a 2012 sales pitch by Preptel, a CV-optimization company that shut down in 2013. No study, no methodology, no published data. Ever. The reality, measured by surveying recruiters about their own tools: ATS systems rank and file applications, they almost never auto-reject. The only hard filters are knockout questions the recruiter sets by hand (work authorization, location, required certification). The automatic 75% wall doesn't exist.
Optimizing your CV against a rejecting robot means preparing for a fight that isn't happening. The real reader is still a busy human.
"Numbers get you 40% more callbacks." A ghost citation. The supposed source (TalentWorks) is a dead domain, and the original "40%" didn't even mean callbacks: it referred to meeting 40% of a job's requirements. The figure got laundered from blog to blog until it became a truth nobody checked. To be clear, I'm not saying numbers are useless — quite the opposite (more on that below). I'm saying the precise multiplier is invented.
"Recruiters read in an F-pattern." The famous F-pattern comes from a 2006 Nielsen Norman Group study on reading web pages, not CVs. The CV extrapolation was made by blogs, never by a study on recruiters. NN/g itself lists five reading patterns and notes the F is "rarely a perfect F." What stays true is that the top and the left draw attention. But "your recruiter reads in an F," nobody has shown that on a CV.
The common thread across these three myths: a clean number, a fuzzy authority, and zero link to a study. The reflex that saves you is to ask for the primary source. Nine times out of ten, it doesn't exist.
What's actually proven
The good news is there's real science underneath. Less sexy, more nuanced, but usable. Here's what holds up, with the honest level of evidence next to it.
Writing quality has a measured causal effect. This is the strongest evidence of the lot: a controlled experiment by van Inwegen, Munyikwa and Horton (2023) on roughly 480,000 jobseekers shows that receiving writing assistance raised the probability of being hired by 8%. Not the polish — the clarity. Writing better helps the employer assess your ability, not just guess it. A real randomized trial, large sample. If you keep one thing: sweat the sentence.
Two pages, for an experienced profile, is defensible. A ResumeGo simulation (2018) saw recruiters prefer two-page CVs over one-page ones, especially for managerial profiles. Take it with a grain of salt (study funded by a CV-writing service, and it's a simulation, not real hiring), but it lines up with the industry consensus: one page under ten years of experience, two beyond, three almost always penalized outside academia, medicine and law.
The photo isn't neutral. A study in Management Science (Ruffle & Shtudiner, 2015, 5,312 CVs sent) shows the photo introduces a measurable attractiveness and gender bias that can backfire on you. In the US and UK, omitting it is the norm and cuts discrimination risk. Verdict: no photo is a defensible choice, not negligence.
Vocabulary matters more than the "robot." Harvard's Hidden Workers study (2021), on 8,000+ workers, shows the real problem isn't auto-rejection, it's vocabulary mismatch: if you don't use the words from the posting, the recruiter searching their database doesn't find you. So yes, mirroring the job's language is worth it. Not to clear a fictional wall — to be findable and readable by the human who scans.
And the seven-second scan? The number comes from a TheLadders eye-tracking study, small (30 recruiters), commercial, never replicated. Treat it as a rhetorical image, not a measurement. But the idea behind it holds: the first second decides, and the top of the CV must instantly say "good profile." That part is solid.
The one real technical trap: columns
Where the ATS myth deflates, a very real problem takes its place: parsing. When an ATS reads your CV, it turns it into structured text, line by line, left to right. Give it two columns, and it reads across: your job title on the left blends with a skill from the right column, and the output becomes mush.
Practical tests on five ATS platforms (Workday, Greenhouse, Lever, iCIMS, Taleo) confirm it, with numbers: on a two-column layout, Taleo flat-out lost an entire job and returned an empty skills list. The CSS-styled skill "chips"? Either ignored or shattered into orphan tokens. The QR code? An image, so invisible to the parser. In short, everything that makes my CV pretty to a human eye makes it risky for a machine.
My two-column CV is gorgeous for a human and treacherous for a parser. Both are true at the same time.
The fix isn't to butcher the design, it's to keep two versions. The pretty two-column PDF for what you actually do with a freelance CV: send it directly, link it from your site, hand it over in person. And a single-column version, no chips, standard section headings, "Jan 2021" dates, for the corporate portals with an ATS. Free 30-second test: copy-paste your CV into a plain-text editor. If the columns scramble, the ATS sees the same mess.
The radar: scoring a CV on what actually matters
Once the folklore was set aside, eight criteria with real evidence behind them remained. I turned them into a grid and scored myself with it. No-mercy verdict on my own CV:
| Criterion (evidence-backed) | Score | Why |
|---|---|---|
| Quantified impact | 9/10 | Almost every line carries a number (99.9% uptime, integration complexity −80%, response time ÷4). |
| Seniority & scope | 8/10 | Dev → Lead → Architect progression visible, explicit architecture decisions. Missing 1-2 "chose X over Y" trade-offs. |
| Writing & clarity | 8.5/10 | Action verbs, precise prose. The best-proven axis (the +8% trial). |
| Human readability / design | 9/10 | Good-looking, clear hierarchy, scannable in seconds. |
| ATS compatibility (parsing) | 4.5/10 | The flip side of the design: two columns + chips + QR break parsing. No flat version. |
| Hierarchy / top of CV | 6.5/10 | Experience starts on page 2. The strongest role sits below the fold. |
| Length / concision | 6/10 | Three pages. The seniority consensus caps at two. |
| Proof / portfolio (GitHub, projects) | 8.5/10 | Annotated projects with stack and substance, GitHub and portfolio linked. A real reason to click. |
Overall: 7.5/10. The content is top-tier: quantified, senior, proven. The only two real weaknesses are mechanical, not substantive: ATS parsing caused by the design, and the length that top-loads badly. Exactly the kind of diagnosis you can't make on your own, because you confuse "I like my CV" with "my CV does the job."
What it actually changed
Four levers, in order of impact. The first two are already live on my CV page, downloadable.
Two versions of the CV, not one. The pretty two-column PDF, and a flat single-column version, no chips, parser-friendly for ATS. The pretty one for direct sending and the site, the ATS one for corporate portals. Simple rule: online portal means the ATS version; direct human contact means the pretty one; when in doubt, the ATS one, because it stays perfectly readable by a human while the reverse isn't true.
Current role lifted to page 1. The radar flagged that my experience starts on page 2: the strongest role sits below the fold. I first tried to lift the whole experience section onto page 1, and it orphaned the section heading at the bottom of the page, uglier than before. The real fix was simpler: a "current role" line in the header ("Backend Architect at Goin Invest, fintech under AMF license"). The strongest signal enters the seven-second scan zone without breaking anything. And I owned the three pages: dense content with no filler beats two cramped pages that sacrifice the projects.
A different CV for every application. The best-proven lever almost nobody uses: tailoring the CV to the posting. Since mine is generated from HTML, I produce a variant per posting, profile and skills reordered to match the role's vocabulary. Reorder and rephrase what's true, never invent.
Two substantive tweaks. Add a real architecture trade-off ("modular monolith over microservices because...") and, when the number exists, a business impact in euros or users. That's what moves you from "senior" to "architect" in a hiring manager's eyes.
Conclusion
The real value of AI in this story isn't that it wrote my CV. It's that it did the dirty work nobody does: trace every piece of advice back to its primary source and find that half of them are ghost citations. The CV-advice industry largely runs on folklore repeated with confidence, because nobody clicks the link.
And the lesson that outlives the CV: I'd spent time optimizing my CV against a rejecting robot — a robot that doesn't exist. Meanwhile the real reader, the busy human who decides in a few seconds, was there the whole time. That's often the scam of "best practices": they make you please the wrong judge.