Beyond Keyword Matching: How AI Reads CVs Like a Recruiter

You search for a Senior Cloud Security Engineer. The tool returns forty CVs. Twelve list the right words but have never done the job. And somewhere — outside the list entirely — is the best candidate you've seen in months, filtered out because he phrased his experience differently to your search.

That's what keyword matching costs you. Not just the bad interviews. The missed placements.

What keyword tools actually do

A keyword matcher checks whether words appear on the page. That's the whole job. It doesn't know whether someone used a skill or merely listed it. It doesn't know that "hardened Azure workloads" and "Azure security" are the same experience. It treats a CV as a list of tokens, and tokens lie constantly.

The failure comes in two directions:

Both are expensive. Most recruiters have counted the cost of the first. Almost none have counted the cost of the second.

How a good recruiter actually reads

Watch an experienced recruiter with a pile of CVs. They're not scanning for words. They're reading in three layers:

Layer 1: Claims

A claim is what the candidate says about themselves. "Python." "AWS." "Strong leader." Every skill listed in a skills section is a claim. Claims are cheap — anyone can write one — and a CV is full of them.

Claims aren't useless. They tell you what the candidate wants to be seen as. They're just not evidence.

Layer 2: Evidence

Evidence is the claim with a home: an employer, a date range, and a description of work actually done.

Consider the same claim on two CVs:

Same word. Completely different signal. A keyword matcher cannot tell these apart. A reader can — in about ten seconds.

The test is four questions:

Layer 3: Context

The same evidence means different things in different contexts. Five years of Python at a ten-person startup is a different candidate to five years of Python at a 5,000-person bank — and the role you're filling determines which one you want.

Context questions to ask alongside the evidence:

• Recency. Did they use the skill last month, or five years ago? Skills decay.
• Trajectory. Is the skill growing, plateaued, or abandoned across their career?
• Environment. Startup pace or enterprise process? Regulated sector or fast-moving one? Which matches your client?

A keyword matcher sees "Python" in 2019 and "Python" in 2026 as the same thing. They are not the same thing.

A worked example

Take a real search: Senior Cloud Security Engineer for an energy-sector client. The ad wants Azure security, SIEM experience, infrastructure hardening, and regulated-environment experience. Three candidates:

Candidate 1 lists every keyword in the ad. Azure, Sentinel, ISO 27001, zero trust. Scrolling down: none of these words appear inside any job entry. They appear only in the summary. Claim without evidence — the keyword matcher's favourite. A recruiter reads layer two and moves on.

Candidate 2 writes "managed our cloud estate" with no cloud words at all. But inside the roles: "Consolidated 40+ Azure subscriptions into a governed landing-zone architecture, 2021–2024, energy trading firm." The evidence is all there, phrased differently. The keyword matcher never shows you this CV. It's the best candidate in the pile.

Candidate 3 has genuine cloud security experience — at a five-person e-commerce startup, six months ago. The evidence is real, but the context is wrong for a regulated energy client. Strong, but not for this role. A matcher can't tell; a reader can.

Same forty CVs. Three different decisions. Only one of them is visible to a keyword search.

The 60-second pass

Reading in three layers sounds slow. It isn't — you're not reading more, you're reading for specific things:

1. Seconds 0–10: Skip the summary. Go straight to the most recent role. That's where the evidence lives.
2. Seconds 10–30: Run the four questions on the two or three skills the role actually depends on. Mark each as claim or evidence.
3. Seconds 30–45: Check recency and environment. Same skill, five years ago, at the wrong kind of company — downgrade it.
4. Seconds 45–60: Decide: strong, possible, or out. You now have a reason for the pile you put them in.

You can do this at the speed you'd skim anyway. The difference is you're making the one decision that matters — is this real experience? — instead of the one that doesn't.

Monday morning

• Take your last five shortlisted candidates. Mark every skill as claim or evidence using the four questions. Regrade anyone you got wrong
• Re-run one search with the title swapped for the core skill — read ten CVs the keyword tool would have filtered out
• Write your one-sentence defence for each of this week's shortlist before you send it

---

Keyword matching gives you speed without judgment. Reading in layers gives you both — you spend the same minutes, but they land on the candidates who can actually get hired. The recruiters who read this way don't work harder. They just stop paying for the wrong pile.