If you have ever wondered why a clearly qualified candidate vanished into your ATS without a phone screen, the answer is almost always at the parsing layer, not the scoring layer. The candidate had a two-column resume. Their name was in a header. Their dates were in text boxes. The system received the document, returned a clean parse for 80% of it, and silently dropped the parts it could not read. The candidate looked weak on screen because the screen was looking at half a resume. This happens more than founders realize, and it is the single biggest reason ATS scoring gets blamed for problems that are actually parsing problems.
What an ATS actually does, in one paragraph
An applicant tracking system is software that collects, parses, organizes, ranks, and stores job applications. Most modern ATS platforms run a three-stage pipeline: parse the document into structured data, evaluate the parsed data against the job requirements, and rank candidates so a human can review the strongest first. The implementation has evolved meaningfully since 2020. Earlier ATS platforms relied almost entirely on hard keyword matching, where a job description listing “Python” would only match resumes containing the literal token. Current systems including CurriculoATS combine semantic matching, multi-signal scoring, and large language model reasoning to evaluate context, not just tokens. The market for these platforms has expanded accordingly. Fortune Business Insights estimates the global ATS market at multi-billion-dollar scale and growing at high single-digit CAGR through 2030, driven primarily by AI-enabled screening replacing legacy keyword tools.
Why the architecture matters more than the brand
Two ATS platforms with the same feature list can produce wildly different outcomes depending on how their parsing and scoring layers are designed. The brand on the box does not predict the screening quality; the architecture under the hood does. A founder evaluating ATS options should care less about logos and more about three specific questions: how does the parser handle non-standard layouts, what signals does the scoring layer evaluate, and does the system expose its reasoning in a way you can audit.
Stage 1: parsing, where most failures happen
The parsing engine extracts structured data from the resume document: name, contact information, work experience entries with dates and titles, education, and skills. Industry-side analysis suggests roughly 30% of all ATS rejections originate at the parsing stage, before scoring even runs. The most common parsing failures are predictable. Multi-column layouts get linearized incorrectly, so a candidate’s third bullet point under their current job ends up appearing to belong to the previous role. Text boxes are skipped entirely by older parsers because they live outside the document’s main content stream. Headers and footers, where many designers place contact information, are sometimes ignored. Image-based name banners are invisible. Creative section headings (“Adventures” instead of “Experience”) confuse the classifier. Each of these results in a candidate looking weaker than they actually are. CurriculoATS uses a modern parsing pipeline that handles standard layouts robustly but, like every parser, struggles with extreme creative formatting. The advice that has remained constant for a decade still holds: a clean single-column resume in standard sections parses cleanly across every major ATS.
What the parser sees vs. what the human sees
A useful exercise: copy and paste your resume into a plain text editor. What you see is approximately what the parser sees. If your contact information is missing, your dates are out of order, or your bullets are interleaved with previous roles, that is what the scoring layer will evaluate. Most parser failures are visible immediately at this step.
Stage 2: scoring, where ATS platforms diverge
Once the resume is parsed, the system evaluates the structured data against the job requirements. This is where ATS platforms differ most. Three approaches dominate the market in 2026. First, hard keyword matching, the legacy approach: the job description’s keywords are searched literally in the resume text and a frequency score is produced. This is fast and explainable but brittle. A candidate who described their AWS experience as “running production services on Amazon’s cloud” gets a low score against a job requiring “AWS” because the literal token is missing. Second, semantic matching with NLP: the system understands that “AWS” and “Amazon Web Services” and “running services on Amazon’s cloud” are equivalent. This catches more candidates correctly but still scores at the keyword level, just with a thicker dictionary. Third, multi-signal scoring with reasoning: the system evaluates the resume on multiple dimensions (quantified achievements, experience relevance, career trajectory, skills alignment) and produces a composite score plus a written reasoning paragraph. CurriculoATS uses this third approach, which we call Impact Scoring, because keyword-level evaluation, even with semantic enrichment, misses the strongest signal in a resume: outcomes attached to specific roles.
How does CurriculoATS score candidates differently?
The composite score is computed from four signals weighted toward outcomes: quantified achievements (numbers attached to results), experience relevance (whether the candidate has solved problems near the open role), career trajectory (whether responsibility is growing role-over-role), and skills alignment (whether claimed skills match the resume’s evidence). The output is a 0-100 score and a written paragraph explaining why the score is what it is, so a founder can audit the scoring decision in seconds. This is the difference between “keyword density 0.78” and “strong fit because the candidate’s last three roles each shipped measurable revenue impact in the same domain as your open role.”
Stage 3: ranking and the human handoff
The third stage is presentation: ordering candidates so a human reviewer reads the strongest first. This is more important than founders realize. A pipeline that delivers 200 candidates sorted by application date forces the founder to read all 200 to find the top 8. A pipeline that delivers the same 200 candidates ranked by composite score lets the founder read the top 8 and make a confident triage decision in 30 minutes. The ranking is also where the system’s biases become operationally visible. NYC Local Law 144 requires bias audits for automated employment decision tools used in NYC hiring, and the EU AI Act’s Annex III classifies recruitment AI as high-risk, requiring documentation, human oversight, and ongoing monitoring. Modern ATS platforms designed for AI-assisted scoring should publish their reasoning so that bias audits can be conducted meaningfully and so that hiring teams can sanity-check the ranking output before making interview decisions.
What’s the candidate experience side of this look like?
The candidate sees an apply form, a confirmation email, and either an invitation to phone screen or a polite rejection. Behind that simple surface, a parser, a scorer, and a ranker have run; the candidate’s resume has been compared against several hundred others; and a structured decision has been made about whether they advance. Modern ATS platforms also provide candidate-facing transparency: telling applicants whether automated tools were used in their evaluation is now legally required in NYC, increasingly required across the EU under the AI Act, and broadly best practice elsewhere.
What we learned at Amazon about ranking systems
Before CurriculoATS, our founder Dev worked on Amazon’s search and recommendation systems. The lesson that translated most directly: a ranking system is only as good as its feedback loop. Amazon’s product rankings improve because every click, purchase, and return feeds signal back into the model. Most ATS platforms are designed without this loop; they score a candidate, the candidate is hired or rejected, and nothing about the outcome flows back to recalibrate the score. The result is a system that gets stale. CurriculoATS is built with the loop closed: when a hired candidate’s 90-day performance is logged, the system learns which signal patterns predicted success in your specific company. The general scoring model is a starting point; the team-specific calibration is what makes the score reliable over time.
Why explainable scoring beats black-box scoring
If a system gives you a 0-100 score with no explanation, you cannot tell whether the score is right. If it gives you the same score with a written reasoning paragraph, you can audit it in seconds. The audit catches the cases where the model misread a non-traditional career path or weighted a keyword the wrong way. Explainability is also what makes regulatory compliance under NYC Local Law 144 and the EU AI Act tractable; bias audits on opaque models are mostly performative.
Frequently asked questions
Are 75% of resumes really rejected by ATS?
The 75% number is widely repeated but its original source is thin. The accurate version is that a meaningful percentage of resumes get filtered out at the parsing or keyword stage before a human reads them. The exact percentage depends heavily on the ATS, the resume format, and the role. What is true is that resumes with parsing-hostile formatting (multi-column, image-based names, text-boxed contact info) are deprioritized in nearly every system.
How does CurriculoATS handle non-standard resumes?
The parser handles standard single-column resumes robustly. For multi-column layouts, the system reconstructs the reading order using layout analysis. For text-boxed content, the parser includes the text in the main content stream. The scoring layer then evaluates the parsed content on outcomes, not just keywords, so a candidate whose resume describes work in their own words is not penalized for missing a specific token. Read more on the AI resume screening page.
What’s the difference between keyword matching and AI screening?
Keyword matching counts tokens. AI screening evaluates context, including outcomes, trajectory, and relevance. Keyword matching is fast and explainable but brittle. AI screening is slightly slower and requires reasoning, but produces a meaningful ranking that correlates with hiring success. The CurriculoATS Impact Scoring layer combines a fast keyword pass with a multi-signal scoring pass, so both speed and quality are achieved.
Do I need to optimize my resume for the ATS?
For the parsing layer: yes, by using clean single-column formatting, standard section headings, and avoiding text boxes or images for critical content. For the scoring layer: depends on the ATS. Older keyword-based systems require explicit token matching; modern systems including CurriculoATS understand context, so writing clearly about outcomes matters more than stuffing in keywords.
How do bias audits actually work for ATS platforms?
Under NYC Local Law 144, an independent auditor evaluates the AEDT for disparate impact across protected categories using historical data. The audit must be completed and published within 12 months of use. The EU AI Act under Annex III requires similar documentation, plus human oversight, monitoring, and documentation of training data. Modern ATS vendors should support these audits by exposing scoring reasoning and outcome data; opaque models make audits much harder to conduct meaningfully.
Take the next step
Understanding how an ATS works is the first step in choosing one that fits your team. The architectural decisions, parsing approach, scoring methodology, and ranking transparency, matter more than the feature checklist. If you want to see how a modern ATS evaluates candidates with explainable reasoning, the free CurriculoATS Starter plan covers one active job. The Impact Scoring page walks through the four-signal evaluation in detail. The right ATS is the one whose architecture matches how you actually want to make hiring decisions; the brand on the box rarely tells you that.