Open any stack of 200 inbound resumes and you will find the same pattern. About forty of them describe real, recent, role-relevant work. The other one hundred and sixty describe hours, titles, tools, and tasks that the candidate did, in some order, at some point, for some company. The first forty are signal. The rest is noise. Most hiring teams cannot tell them apart fast enough, and that is the entire reason resume screening still feels broken in 2026.
What signal and noise actually mean on a resume
Signal is information that predicts whether a candidate will do the job you are hiring for. Concrete outcomes (“cut p95 latency from 800ms to 120ms”), relevant scope (“owned the billing service for 18 months”), and decision-shaped achievements (“shipped onboarding redesign that lifted activation 14%”) are signal. Noise is everything else: generic responsibilities, buzzwords, certifications nobody verifies, and tools listed without context. Harvard Business Review estimates that about 72% of resumes are filtered out before a human ever reads them, and the systems doing that filtering mostly look at noise — exact-match keywords, years-of-experience integers, and degree fields. A candidate who shipped the exact thing you need but used different words gets cut. A candidate who wrote a buzzword salad that mirrors your job description gets through. The cost of treating noise as signal is not abstract: SHRM’s 2025 Benchmarking Report puts the average cost-per-hire at $5,475, and a wrong hire compounds for months.
Why keyword matching cannot separate signal from noise
Legacy applicant tracking systems were designed in an era where the bottleneck was paperwork, not judgment. They parse a PDF, extract tokens, compare those tokens to a job description, and rank. This works for compliance forms. It does not work for distinguishing a senior engineer who shipped real systems from a junior who pasted the right words. Three structural problems make keyword matching the wrong tool:
- Vocabulary drift. A candidate who built “event-driven backend services on AWS” does not match a job description asking for “REST APIs and Lambda.” Same skill, different vocabulary, zero score.
- Context blindness. The phrase “led a team of five” means something different at a 10-person seed startup than at a 30,000-person bank. Keyword systems treat them as identical.
- No outcome weighting. Two resumes can list the same skills. One person built a thing that ten million users touched. The other helped maintain a tool used by twelve people internally. Keyword scoring cannot see the difference.
This is why even Greenhouse, Lever, Workable, and Ashby — all of which have added “AI” features over the last two years — still produce shortlists that hiring managers re-rank by hand. The AI is bolted on top of the same keyword foundation. It speeds up the wrong activity.
What we learned from search and recommendations at Amazon
Before founding CurriculoATS in 2024, our founder Dev spent years at Amazon working on search and recommendation systems — the same problem class as resume screening, just dressed differently. The lesson from that work translates almost directly. Ranking systems that win do not ask “which document contains the most query keywords?” They ask “which document is most likely to satisfy the user’s intent?” Two principles drove every win we shipped:
First, evaluate against outcomes the user actually cares about. For Amazon, that was conversion and long-term satisfaction. For hiring, it is whether a candidate can do the job and stick. That means scoring on quantified achievements, experience relevance, career trajectory, and skills alignment — not on whether the word “Python” appears 4 versus 7 times.
Second, expose the reasoning. A black-box ranker that says “trust me” loses to a transparent one every time, because users can audit, correct, and trust the transparent system. We built CurriculoATS so every 0–100 fit score is paired with a written reasoning paragraph explaining what the model saw and weighed. A founder reading that paragraph in our Impact Scoring view can decide in seconds whether the AI got it right, and override when it did not. That is signal-based hiring in practice.
How to test whether your current ATS reads signal or noise
The fastest way to find out what your ATS is actually doing is to run a controlled test that takes about 20 minutes. Pick three resumes you and your team have already evaluated and agreed on: one obvious top candidate (clear quantified outcomes, recent relevant work), one obvious weak candidate (buzzword-heavy with no measurable achievements), and one borderline case (good background but unclear outcomes). Submit all three through your standard application flow. Look at the rankings the ATS produces. If the obvious top candidate is in the top quartile and the buzzword resume is in the bottom half, the system is reading something useful. If the buzzword resume scores higher than the top candidate — which happens roughly 40% of the time on legacy keyword systems we have tested — your ATS is screening on noise. The second part of the test is the reasoning paragraph. Ask the system: why did this candidate score where they did? If the answer is “73% match” or “strong fit,” you do not have explainability. If the answer is a paragraph that names the achievements, the experience signals, the trajectory, and the gaps, you do. Most teams who run this test for the first time discover their ATS has been reranked silently in their heads for months — they just stopped trusting the output without realizing it. The discipline of running the test once a quarter, even after a switch, keeps the model honest. Models drift. Job descriptions drift. The 20-minute test is the cheapest insurance against either kind of drift becoming an expensive misfire.
How a startup founder applies signal-vs-noise thinking
You do not need a machine learning team to use this framework. You need a job description that names outcomes instead of tasks, and a screening loop that rewards them. Five concrete moves:
- Rewrite your JD around outcomes. Replace “manage the marketing funnel” with “own paid acquisition CAC for our SMB segment, target $X by Q3.” Candidates self-select.
- Ask for one before/after number. In your application form, add a field: “Pick one project from the last 18 months. What was the metric before and after you touched it?” Most applicants will leave it blank. The ones who fill it in are pre-screened for signal.
- Read the reasoning, not just the score. If your ATS produces a score with no explanation, you are still doing keyword hiring. If you can read why a candidate ranked where they did, you can correct mistakes early.
- Stop weighing certifications and degrees as primary filters. McKinsey’s research shows skills-based hiring grew from 40% of companies in 2020 to 60% in 2024, and over 15 US states are formally dropping degree requirements for public-sector roles.
- Time-bound the signal. A candidate’s last 36 months of work matter more than the previous 10 years. Weight accordingly.
FAQs founders ask about resume signal
How do I know if my ATS is screening on signal or noise?
Run a test. Take a candidate your team agrees is strong and submit their resume to your current ATS through the normal application flow. Then take a generic, well-formatted resume that uses every keyword from your JD but describes irrelevant work. If the second resume scores higher than the first, your system is screening on noise. Most legacy ATS platforms fail this test, which is the practical reason we built CurriculoATS as an outcome-based screener.
Can AI really distinguish quantified achievements from boilerplate?
Yes, when it is built to. The trick is training the model on the right objective: did this person produce outcomes, or did they list responsibilities? Modern large language models are good at parsing prose for cause-and-effect structure. The harder problem is exposing the reasoning so a hiring manager can audit it. Our scoring writes a paragraph per candidate explaining what it weighed, which gives founders a real check on the model rather than a sealed number.
What about candidates with non-traditional backgrounds?
Signal-based screening helps non-traditional candidates more than it hurts them, because it weighs what they did over where they did it. A self-taught engineer who shipped a production system at a small startup will outscore a credentialed candidate whose resume is full of keywords but thin on outcomes. This is one of the few ways to reduce “pedigree bias” without resorting to manual blinding.
Will candidates game an AI that reads for outcomes?
Some will try. The defense is the same as in search ranking: cross-check claims against context. A claim of “led 50-person team” at a company with 12 employees on LinkedIn is a flag. A claim of “grew revenue 400%” with no role context gets weighted down. Outcome-based scoring is harder to game than keyword scoring because the model is reading for coherence, not token frequency.
What does a high-signal job description look like?
It names outcomes the role owns in the first six months, not duties the role performs daily. “Own paid acquisition CAC for SMB” is signal. “Manage marketing campaigns” is noise. The test is whether a candidate could write a four-sentence cover letter explaining how their last role maps to your outcomes. If they cannot, your JD is too vague and your applicants will be too. Signal-first JDs reduce inbound volume by 30-50% and raise shortlist quality more than any sourcing change a startup can run.
What to do this week
If your inbound is buried under noise, two moves change the trajectory in seven days. First, rewrite one open role’s JD around outcomes and add a single “before/after” question to the application. Second, run your existing pipeline through an outcome-based screener so you can compare what your current ATS surfaces against what a signal-first system surfaces. CurriculoATS is free to start and takes 15 minutes to set up — see the features overview or jump straight to pricing. Then read SHRM’s 2025 cost-per-hire benchmarking report to set expectations for the savings.