Recagn
Sign in

Guide · Shortlisting

How to shortlist candidates from resumes: criteria first, then evidence

Decide what the role needs before you open the first CV. Then read each CV for evidence, and treat every line as a claim until the interview.

What CV shortlisting criteria should you set before reading any CV?

Set two lists before reading any CV: a few pass-or-fail must-haves, and a weighted list of what separates a good candidate from a great one. The must-haves rule candidates out, and the weights set the order of everyone left. You write both lists, and the client signs them off before the first CV.

Together, the two lists are the role's scorecard. Every CV in the pile gets read against the same one, and the interview later scores against it too: see how to build an interview scorecard.

Must-haves

Keep must-haves for what truly disqualifies: the work location, a licence the law requires, a stated minimum of experience. Each one empties part of the pile, so too many leave you with nobody to read.

A useful test: if the client names six must-haves, ask which candidate they would still meet if one were missing. That one is a weighted criterion, not a must-have.

Weighted criteria

Write each weighted criterion as something a candidate did, not a word they used. "Handled payment reversals in production" can be checked against a CV and tested in an interview. "Payments" can only be matched as a word.

Give each criterion a weight that matches how much the client cares. Two criteria the client calls equally important carry the same weight.

Then anchor the scale. For each criterion, write down what a 1 looks like and what a 5 looks like, in the role's own terms. Without anchors, two recruiters on one desk read the same CV and mark it differently.

For "handled payment reversals in production", a 1 might be no payments work at all. A 3 might be work on a payments team without owning the reversal flow. A 5 might be owning that flow and being able to say how failed reversals were caught and fixed. Anchors like these turn "seems strong" into a mark another recruiter would give too.

Lock them before the first CV

Write the criteria before you open the pile, and don't change them halfway. If you change a criterion after reading 100 CVs, those 100 were read against a different role.

The questions that get these answers out of a client belong on the intake call. They are set out stage by stage in the candidate screening process for recruitment agencies.

Should you shortlist CVs by keyword match or by criteria?

Shortlist by criteria: read each CV against the criteria you set, because a keyword only shows that a word appears, not what the candidate did. Keyword filters suit must-haves a CV states plainly, such as a city; AI matching reads for meaning, but from the same unproven CV. A ranking only orders the pile, and you choose who to invite.

The three ways of matching a CV to a role differ in what they read and what they miss.

Keyword matchCriteria match, by handAI match against criteria
What it readsWords on the CVThe CV, against each criterionThe CV, against each criterion
What it catchesA missing hard requirement you can name: a city, a degree, a certificationWhat the candidate did, if the CV says itThe same, across the whole pile
What it missesSynonyms, "used" against "ran", anything the CV words differentlyConsistency: two readers mark the same CV differently without anchorsAnything the CV doesn't say, and any claim that isn't true
Speed on a large pileSecondsSlowestFast
Who decidesYouYouYou: the ranking is a recommendation

Source: how each method works · a description, not a ranking

The input is the same CV in all three columns. None of them can tell a true line from a padded one, because the CV is a document the candidate wrote about themselves.

That is why matching is only half of shortlisting. The other half is deciding which claims the interview has to test.

How do you read a CV for evidence, not keywords?

Read each CV line for what the candidate did, owned and measured, and mark each criterion as shown, claimed or missing. "Payments experience" names a field, while "refunds and reversal flows for UPI" names work, and only work can be tested in an interview.

Shown, claimed or missing

Give every criterion one of three marks as you read:

  • Shown. The CV holds something you can check on paper: dates, job titles, a degree, a certification.
  • Claimed. The CV states ownership, depth or a result that only the interview can test.
  • Missing. The CV is silent. Silence is a question to ask, not a no.

Most lines that matter land in "claimed". That is normal: it tells you what to ask, not who to drop.

Verbs and skills lists

Watch the verbs. "Used", "worked on" and "exposure to" describe contact with a thing. "Built", "ran" and "owned" describe responsibility for it, and those are the lines worth testing.

A skills list is a keyword list. It shows what the candidate wants you to search for, and the experience lines show what they did with it.

Numbers need their count. "Cut cloud cost by roughly a quarter" is a claim until the candidate says what the bill was before and after, and how much of the change was theirs.

Turn each claimed line into an interview question before you finish the CV. For the cost line: what was the monthly bill before, what was it after, which change saved the most, and who else worked on it. A candidate who did the work answers those easily. One who rounded up a team's result usually says so when asked.

One CV, two readings

Here is a line from an invented CV we wrote for a senior Java full stack role. In her skills list, Sneha K. writes "Kubernetes (as a user)". Her experience line says she deploys "to a Kubernetes cluster that the platform team runs".

A keyword filter for "Kubernetes" passes her. A criterion that reads "administers Kubernetes clusters" marks it missing, because her own CV says someone else runs the cluster.

Neither reading is a rejection. The marks are your reading of the CV, and you decide what to ask her and whether she goes forward.

How do you spot fake experience and proxy interviews?

Check whether the dates support the claimed years, whether the interview answers back up the CV, and whether the person interviewing is the person you'll place. No single check is proof, because people round up dates and count internships, so a mismatch is a question for the candidate. You decide whether a mismatch matters, after you ask.

Red flags

Three flags come up often enough to check on every senior role. For each, here is what it looks like and how to check it.

Inflated experience. A summary line claims more years than the roles add up to, or a claim the dates allow that the work doesn't support. Check the dates first, then ask about the work. Which check settles which kind of fake experience, and how to verify work experience in India, is in the guide to fake experience on a resume.

Proxy interviews. Someone else answers the interview, or the call, in the candidate's place. Check identity more than once, from the first round to joining, and ask follow-up questions about the candidate's own past work. How to spot a stand-in or a deepfake candidate, and what a face match can tell you, is in the guide to proxy interviews.

AI-written CVs. Fluent, generic wording proves nothing on its own, either way. Treat a polished CV like any other: a set of claims to test in the interview.

We haven't tested whether AI-written CVs can be detected, so this guide makes no claim about it. Tools that read interview answers for signs of read-aloud or AI-written text judge the answers, not the CV.

How common it is

We found figures for how common fake experience is among candidates in India, but only as second-hand reports. We have not checked them at their source, so this guide quotes none.

What a mismatch means is still your call. Ask the candidate first, note what they say, and then decide whether it changes your view.

How do you reuse your candidate database for a new role?

Search your own candidate database with the new brief's criteria before you open a new pile, and rank past candidates by the CV and by how they did on earlier roles that tested the same skills. You decide who to approach again.

Rediscovery

Rediscovery means going back to candidates you have already screened. Search by the brief's criteria, not the job title, and confirm notice period, CTC and interest before you put anyone forward again. How to search old CVs, rank what you find and tell when your database has nobody is in the guide to candidate rediscovery.

How do you screen 500 resumes fast without missing good candidates?

Screen 500 resumes in two passes: drop only CVs that clearly fail a must-have, then rank the rest against the weighted criteria. Decide how many to invite by working back from the shortlist the client wants and how many interviews you can run, not from a fixed percentage. The ranking orders the pile, and you pick who is invited.

Pass one: clear failures only

The first pass is mechanical. Only CVs that clearly fail a must-have go: a location the client won't accept, a licence the role legally needs.

A CV that is silent on a must-have goes through. The interview can ask, and a silent CV is not a failed one.

Pass two: rank, then read

The second pass ranks everyone left against the weighted criteria. Then spend your reading time where it counts:

  • The top. Read closely, marking each criterion shown, claimed or missing.
  • The middle. Skim for criteria the ranking may have missed, such as relevant work under an unusual job title.
  • The bottom. Leave it. Don't delete it: it may fit the next role.

Your candidate database comes before all of this. Past candidates with relevant history go into the ranking alongside the new pile, not after it.

How many to invite

Work back from the end. Start with how many candidates the client wants on the shortlist. Then decide how many must finish an interview for you to pick that many with confidence. Then invite more than that, because some won't answer.

We found no published benchmark, with a method behind it, for how many CVs an agency should invite per shortlist place, so this guide gives none. The "10% of CVs received" rule of thumb on some guides comes without a source, and we don't repeat it.

No method in this section comes with a time saving we could check. The speed comes from skipping close reads of CVs that the criteria already place low.

When should you not shortlist from the CV alone?

Don't shortlist from the CV alone when it can't show what the role needs: for freshers, career changers and many sales roles. There, a short first-round interview with the same questions for everyone tells you more than a closer read. You decide which roles skip the CV ranking.

  • Freshers and campus hires. The CV is thin by design. A ranking mostly sorts by college and grades, which says little about the job.
  • Career changers. The right skills sit under the wrong job titles, and a ranking against the role's criteria pushes them down.
  • Sales and client-facing roles. The CV's numbers are claims, and the skill that matters most, how they talk to a customer, isn't on paper.
  • Roles that turn on availability. Shift work, travel or an immediate joining date decide these roles, and a CV rarely states them. Ask first, then read.
  • A small pile. If you can read every CV in an afternoon, read them all. A two-pass method adds steps without saving time.

In each case, invite more widely and let the same questions do the sorting. The criteria you wrote still apply; they move from the CV to the interview.

How Recagn does this

Recagn is our product. This section takes one brief through the method above, using a real run scored by Recagn on 8 Nov 2026. The brief is adapted from a real public job posting for a senior Java full stack engineer, with company details removed. The candidates are invented, and their answers were performed from a script.

The criteria

1–10 of 10

Recagn drafted a scorecard, which Recagn calls a rubric, from the brief. It has four hard gates: 5+ years of relevant experience, including 4+ years on Kubernetes and GitLab CI/CD; hands-on Java full stack work; platform and cloud operations; and remote availability.

It has six weighted criteria: Java full stack engineering, 25%; AWS and infrastructure as code, 20%; containers and platform engineering, 20%; Linux and CI/CD automation, 15%; databases and SQL, 10%; and collaboration, 10%. The cut-offs are 3.8 for Qualified and 3.0 for Maybe.

Turning the brief into a rubric comes off your plate. What the role needs stays yours: you edit and lock the rubric before anyone is invited.

Reading the CVs

Recagn reads the profile it parses from each CV against the locked rubric, before any interview. That profile holds the CV's skills, years of experience, education, projects and locations, and Recagn's CV read is not sent the name or contact details.

This step is CV screening, which Recagn calls a Match Assessment. It treats every line as an unproven claim, so a hard gate can fail on the CV but can only pass in the interview. It recommends Strong, Possible or Weak, and the recommendation decides nothing.

This run's export holds no match results, so none are shown here. How Recagn's resume screening works, and what it costs, is on its own page.

Working back from the shortlist, as in the 500-CV section above, looks like this in an invented example on Recagn's homepage: 1,240 candidates in the database, 212 who fit the brief's filters, 30 invited and 21 who answered. The numbers are made up for the demo, not a benchmark.

Searching the candidate database and reading every CV come off your plate. Who gets invited stays yours.

Evidence, not keywords

After her interview, Recagn scored Sneha K. 2 out of 5 on containers and platform engineering. Its reasoning says she uses Kubernetes "only as a developer" and "does not administer clusters".

That is the reading from the evidence section above, now made from her interview answers rather than her CV. Each criterion is scored against the rubric's anchors, as how Recagn scores an interview sets out, and every score cites the answer it rests on.

A score is a recommendation. You read the answers and decide.

The claim check

Vikram S.'s CV claim check

Read what the CV says, then what he said when asked. Recagn counts what it found; the counts never change his score.

Vikram S. Senior Java full stack and platform engineer

  1. From the CV we wrote

    Designed the company's Terraform estate across AWS; 7+ years of Terraform in production.

    His answer · Describe a time you designed or optimised AWS infrastructure using Terraform or Ansible. What did you optimise for (cost, availability, security) and what was the measurable impact?

    I own our whole AWS Terraform estate and have done this for over seven years. I optimise for cost and availability. We use reserved instances and autoscaling, and everything is in modules. When I joined in 2019 we were still creating most things from the console, and Terraform came in later, but I led that move. The impact was significant savings across the board.

  2. From the CV we wrote

    Authored the internal developer portal on Backstage, including 15 custom plugins.

    His answer · Tell us about your experience administering Kubernetes or OpenShift clusters and, if applicable, building internal developer portal features with Backstage. What problems were you solving?

    I run a platform of twelve clusters on Kubernetes and OpenShift serving hundreds of microservices. I built our developer portal on Backstage with fifteen custom plugins. The main problem we solved was giving teams one place to see everything about their services. The plugins integrate with our tools across the stack.

Profile against interview: 1 contradiction, 4 inflations, 7 corroborations No effect on scores.

Your decision: not made yet

Scored by Recagn on 8 Nov 2026 · Senior Java full stack and platform engineer · invented candidate; the CV and answers were written for this example

Vikram S.'s CV, which we wrote for this example, says "7+ years of Terraform in production" under a role that started in Apr 2019. By the run date that role is seven years and seven months old, so a dates check alone lets the claim through. That arithmetic is ours, on the CV.

In his interview, he wrote that when he joined in 2019 most things were still created from the console, and Terraform came in later.

After the interview, Recagn compares what the CV claims with what the candidate said. Across his answers, Recagn's check found 1 contradiction, 4 inflations and 7 corroborations.

The check does not contact employers, check documents or run a background check. It compares the CV with the interview, and nothing else.

A finding never changes the interview score or the recommendation. You decide whether it matters, and you ask the candidate first.

Vikram's own recommendation is Disqualified, for a different reason. His criterion scores work out to 2.85, below the 3.0 cut-off for Maybe, because the evidence on Java full stack delivery and databases was thin. The report shows each criterion's score and the bucket, and the 2.85 is worked out from those scores by the scoring rule. Only you can reject him.

Checking CV claims comes off your plate. Whether a finding matters stays yours.

Proxies and borrowed answers

Recagn's authenticity check reads each interview answer for signs it wasn't the candidate's own work: read aloud, read off a screen, or written by AI. It gives each interview one reading from 0 to 10, where 0 means cheating is very unlikely and 10 very likely. Vikram S.'s reading was 3.

The reading sits beside the score. It never changes the score and never rejects anyone.

What it does not do matters as much. It does not check identity, so it cannot tell you the person answering is the person on the CV; Recagn's separate face match, from Starter, compares the candidate's ID photo with a selfie, but does not check that the ID is genuine. The authenticity check does not read the CV for AI writing either.

Flagging likely cheating comes off your plate. Whether a flag matters stays yours.

Your candidate database

Recagn's ranking can move a past candidate up or down by their own history. It counts their results on earlier roles that share a criterion, including claim-check findings and criterion scores, your decisions, the client's outcomes, and whether they replied.

The move is capped at 15 points either way on the 0 to 100 match score, and each move carries its reason. Each kind of history fades by half-life:

  • whether they replied: 14 days;
  • the client's outcomes and your decisions: 90 days;
  • claim-check findings: 90 days;
  • earlier criterion and gate results: 135 days, or 180 days for a score of 1 or 5.

Skill and outcome history count only when the new rubric shares a criterion theme with the old one. Reply history counts on any role.

The ranking never changes a recommendation or an interview score, and it never invites or rejects anyone. You decide who to approach again.

In this run, Recagn drafted the rubric, scored four interviews and checked the CV claims. The rubric, the invitations and every decision stayed with the recruiter.

See a worked example

Questions people ask

Can Recagn tell if someone else took the interview?

Partly. Recagn's authenticity check reads each answer for signs it was read aloud, read off a screen or written by AI, but on its own it does not check identity. With an ID check on, from Starter, Recagn also compares the face on the candidate's ID with a selfie; it does not check that the ID is genuine. Check identity yourself more than once, from the first round to joining. Candidates are told before the first question what the interview records, that AI helps score their answers, and that a person reviews their answers and makes every decision.

What does it cost to screen a resume with AI on Recagn, and on Hyring?

At most ₹3.33 a CV on Recagn Starter before GST (₹3.93 with GST), and about ₹46 a CV on Hyring Pro billed monthly. Recagn's figure is a ceiling, because the Starter price also pays for interview credits; Starter includes 900 CV reads a month, and on paid plans a senior CV read counts as two. Hyring's figure assumes every credit goes on screening: its AI Resume Screener uses 0.25 credits a screening, and Pro gives 50 credits a month for $99 billed monthly (Hyring's pricing page, checked 10 Nov 2026, GST not stated, at ₹93.56 to the dollar, Wise mid-market rate, 17 Nov 2026). Free plans go the other way: Hyring Free's 50 credits a month cover up to 200 CV screenings if none go on interviews, and Recagn Free includes 100 CV reads a month after its first 30 days.

Can my ATS shortlist CVs for me?

Some can rank them. Manatal, for example, lists automatic candidate scoring by AI on its plans, going by Manatal's pricing page, checked 10 Nov 2026. A ranking built from the CV alone still leaves every claim untested until the interview, and you still choose who goes on the shortlist.

How many candidates should be on a shortlist for a client?

As many as the client agreed to see on the intake call, so ask, because clients differ. We searched in November 2026 and found no published benchmark for shortlist size with a method behind it, so this guide doesn't suggest a number.

Should you reject a candidate whose CV years don't add up?

Not before you ask. People round up dates, count internships or describe a team's work as their own. Ask the candidate, note the answer, and then decide whether the gap changes your view.

Does Recagn look at a candidate's name, photo or college when it reads a CV?

Not the name or photo, but it does see the college. Recagn's CV read is sent the skills, years, education, projects and locations parsed from the CV, without the name or contact details. Education includes the institution, and its instructions forbid using or inferring protected attributes or the prestige of a university. You decide who is invited.

What is candidate rediscovery?

Candidate rediscovery is searching your own candidate database with a new brief's criteria, not the job title, before you source new CVs. Rank the people you find by the CV and by how they did on earlier roles that tested the same skills, and confirm their notice period, CTC and interest before you put anyone forward again.

Can you tell if a CV was written with ChatGPT?

We haven't tested whether that can be done reliably, so this guide makes no claim either way. It matters less than it seems: a polished CV makes the same claims as any other, and the interview tests them. Recagn's authenticity check reads interview answers, not CVs.

See one real brief taken from CVs to a shortlist.