Data-Driven Recruitment, Sourcing First: 9 Metrics, Clean Data and the Decisions They Should Change

By Sparsh Goyal, Founder at The Cognitive. Published 2026-10-04 · Last updated 2026-10-05

• Data-driven recruitment means making sourcing, screening and hiring decisions from recorded evidence, with each metric tied to the decision it changes.
• Start at the top of the funnel: in Gem's 2026 benchmarks, job boards and company marketing produced about 90% of applications but only about half of hires.
• The Cognitive produces sourcing yield, contact reveal outcomes, interview dates and scored reports, and leaves funnel analytics to your ATS.

Data-driven recruitment is the practice of making sourcing, screening and hiring decisions from recorded evidence instead of memory. In practice that means tagging where every candidate came from, dating every stage change, scoring interviews against a fixed rubric, and agreeing in advance which number is allowed to change which decision.

Most teams start measuring at the wrong end. In Gem's 2026 Recruiting Benchmarks Report, job boards and company marketing produced roughly 90% of applications but only about half of hires, while direct sourcing produced 11% of hires from 2.6% of applications. The cheapest gains usually sit at the top of the funnel, in who you find and who replies, long before an offer is on the table.

Full disclosure: I run The Cognitive, which sources candidates across ~900M public profiles and interviews them live with AI. It produces sourcing data and per-candidate interview reports. It does not do funnel analytics, and this guide says exactly where that line sits.

Key takeaways

• Data-driven recruiting is a decision habit. A metric earns its place only if you can name the decision it changes.
• Start with 9 metrics, and give the first 4 to sourcing: source of hire, sourcing yield, contact reach rate and reply rate.
• Clean data comes from 3 habits: a controlled source list, a date on every stage change, and a reason on every rejection or withdrawal.
• Structured interviews are the evaluation data worth keeping. In Sackett and colleagues' 2022 re-analysis they had the highest mean validity of the procedures reviewed.
• For benchmarks, use the stage rates in our recruitment funnel conversion rates guide, then beat your own trailing numbers.

What is data-driven recruitment?

Data-driven recruitment, also called data-driven recruiting, data-driven hiring or data-driven talent acquisition, is hiring where each recurring decision has a number behind it. Which channel gets next month's sourcing hours. Which message gets rewritten. Which interview criterion gets dropped because it predicts nothing. The label covers the same work under any of those names.

The test I use is simple. If a metric moved by half tomorrow, would anyone do something different? If nobody would, move it to an appendix and keep the dashboard for numbers that trigger action.

Recruitment data comes in 4 kinds, and each answers a different question:

• Sourcing data: searches run, people found, people shortlisted, contacts found, replies. It tells you whether you are fishing in the right pool.
• Process data: the date each candidate entered and left each stage. It tells you where people wait and where they drop out.
• Evaluation data: interview scores per criterion, written evidence and verdicts. It tells you why people pass or fail.
• Outcome data: offers, acceptances, 6- and 12-month performance and retention. It tells you whether the other three were measuring the right things.

"Big data for recruitment" mostly means the first kind bought at scale: talent mapping data, salary sets from salary benchmarking tools, profile databases. Data analysis for recruitment is the work of joining all 4 kinds by candidate, so you can see that the people from one search, contacted with one message, scored well and stayed.

How do you build a data-driven recruitment strategy?

A data-driven recruitment strategy is a short written plan that fixes your definitions, your sources of truth and your review rhythm before anyone builds a chart. Mine has 5 steps, in this order:

1. Write down the decisions. List the 5 or 6 calls your team makes every month, such as channel mix, message choice and interview loop design.
2. Pick one metric per decision. Use the table in the next section. A decision with 3 metrics gets argued; a decision with 1 gets made.
3. Name the system of record for each number. Source of hire lives in the ATS. Reply rate lives in your outreach tool. Never let 2 tools both claim the same number.
4. Clean the inputs for 4 weeks before you read anything. The section on clean data below is the checklist.
5. Review on 2 clocks. Sourcing and outreach numbers weekly, because you can change a search on Tuesday. Quality of hire and cost per hire quarterly, because they need time and volume.

Which recruitment metrics matter, and how do you calculate them?

These 9 are the ones I would track first. The third column is the point: each metric is listed with the decision it should change, so nobody tracks it for its own sake. The formulas are plain arithmetic. For what "good" looks like at each stage, use the dated benchmarks in our funnel conversion rates guide rather than any number repeated without a source.

2 of these need a written definition before you trust them. Quality of hire has no standard formula, so pick one proxy and hold it still for a year. Cost per hire breaks the moment someone adds recruiter salaries in one quarter and leaves them out in the next. If you want the longer stage-by-stage list, our 14 recruitment KPIs post has it.

How do you collect clean recruitment data?

Clean recruiting data is a by-product of a few boring rules that every recruiter follows every day. No analytics tool fixes inputs that were never recorded.

Fix the source field first

Give the source field a short, controlled list of 8 to 12 values and remove free text. Set it at first touch, never at hire, or every hire gets credited to whatever happened last. Split "sourced" from "applied" and give "rediscovered" its own value. Gem's report found that 46% of sourced hires now come from rediscovered candidates already in a CRM or ATS, up from 26% in 2021, and you can only see that shift if the tag exists. Our candidate rediscovery guide covers how to search that pool.

Date every stage change

Moving a card is not enough. Each candidate needs an entry date and an exit date for every stage, or stage conversion and time to first interview are guesses. Name stages after decisions, such as "hiring manager interview passed", not after activities, such as "chatted". Our funnel guide walks through renaming stages without breaking your history.

Record a reason on every exit

Keep a reason list of about 6 codes for rejections and 6 for withdrawals. "Withdrew: pay" and "Withdrew: took another offer" point to different fixes. Make the field required at the moment a candidate is closed out.

Tag people who arrive from outside tools

Candidates found in a sourcing tool often enter the ATS by hand, and that is where source data gets lost. The Cognitive does not write people it sourced into your ATS automatically. When you add one, set the source to something like "Sourced: The Cognitive" and note the search name, so source of hire still adds up at the end of the quarter.

Run a 15-minute weekly check

Once a week, pull 3 lists: candidates with no source, candidates in a stage with no entry date, and likely duplicates. Fix them by hand, and give the job to one named person so it happens in busy weeks too.

What should a recruitment KPIs dashboard look like?

A recruitment KPIs dashboard should read top to bottom in the order candidates move, with sourcing first and outcomes last. Here is the layout I would build. It is illustrative and holds no real numbers; the labels show what each panel contains.

1. Panel 1, Sourcing (weekly, per role): searches run, people returned, people shortlisted, sourcing yield. One line per open role.
2. Panel 2, Reach and replies (weekly): people contacted, contact reach rate, replies by intent (interested, needs information, not interested, opted out), reply rate split by channel and by message version.
3. Panel 3, Process (weekly): a stage conversion strip for each role, the median time to first interview, and a list of candidates who have sat in one stage longer than your own usual time.
4. Panel 4, Evaluation (per role, as interviews finish): interviews scheduled, completed and missed, plus the spread of scores on each rubric criterion.
5. Panel 5, Outcomes (quarterly): source of hire, offer acceptance, the quality of hire proxy and cost per hire.

2 rules for every tile: print the count behind each rate (n), and print the one-line definition under the number. A rate with no n hides whether it came from 4 people or 400. If you need a recruitment dashboard sample to fill in each week, the recruiting metrics guide has a one-screen weekly report, and our recruitment analytics software review covers the tools that build the charts.

How do you use sourcing data to fix the top of the funnel?

Sourcing data shows which searches, profiles and channels turn into conversations, and it moves faster than any other recruitment data. You can change a search today and read the result tomorrow, which you cannot do with quality of hire.

Read sourcing yield per search before you send anything

If a search returns 50 people and you would shortlist 3, fix the brief before you touch the outreach. Change the filters before a single message goes out, because a weak list makes every downstream number look worse. In The Cognitive's AI sourcing tool, the brief is read into editable filters you can check. In my own test, "Senior backend engineer with 5+ years of Go who has built payments systems, Remote EU" came back with a location filter reading Remote EU as Germany plus 7 other countries, which is exactly the kind of line you want to see before judging the results.

Compare who you shortlist with who replies

Put shortlisted, contacted, replied and interviewed side by side for one role, then group by a profile attribute: current title, company type, years in role or location. Patterns show up quickly. Senior people at large companies might shortlist well and reply rarely, while people 2 years into a scale-up reply often. Shift the next search toward the group that replies and keep the shortlist bar where it is. If your outreach tool can't split replies by message version, our recruiting outreach software comparison shows which ones can.

Weight channels by hires, not by volume

Gem's 2026 figures are the case for this: referrals converted at 11 times the rate of inbound applicants (the case for a real employee referral program), internal mobility at 32 times, and sourced candidates were nearly 8 times more likely to be hired than inbound applicants. Your own ratios will differ, which is why source of hire needs the controlled tag from the previous section.

Let shortlist decisions train the ranking

Every Shortlist and Hide click is data about what you actually want, often more honest than the job description. In The Cognitive, once a role has 8 or more decisions, including 3 or more shortlists, what you shortlist and pass on shapes later ranking. Everyone found stays with the role, and later searches skip people you have already seen, so yield on the next search counts new people only. When the pool is thin, the search relaxes its softest filters and says what it broadened. Note that in your log, because it changes what the yield means.

How do structured interviews turn evaluation into data?

Structured interviews turn evaluation into data because every candidate for a role is scored on the same criteria and the same scale, so the scores can be compared. That comparability is also why they predict well. In the 2022 re-analysis by Paul Sackett and colleagues Zhang, Berry and Lievens in the Journal of Applied Psychology, structured interviews had the highest mean operational validity of the selection procedures reviewed, at .42, as summarised by SIOP's TIP.

Unstructured interview notes can't be compared across candidates, so they can't be analysed. If you want evaluation data, record 3 things for each candidate: a score per criterion on a fixed scale, the evidence behind it in writing, and the verdict kept separate from the score. After 2 or 3 quarters you can check which criteria actually line up with your quality of hire proxy, and drop the ones that don't.

The Cognitive's AI interviewer runs this as a live, two-way AI video interview of 10 or 20 minutes, in 9 languages. Candidates book their own slot without an account, and the invitation tells them the interview is run by AI. The rubric is fixed per role and the questions adapt live to each answer. Each report gives a 1 to 5 score per criterion, a weighted score out of 100, a suggested verdict, overall written feedback, the transcript and the recording. It also checks up to 5 resume claims and marks each one verified, refuted or unclear, with evidence from the interview. Nothing is rejected automatically, and integrity flags are logged, not scored.

This 5-minute recording shows a live AI interview for a go-to-market role. The AI asks the candidate about an outbound email campaign they ran, then follows up on reply rates and on how they qualified leads, adjusting each question to the previous answer.

What are the pitfalls of data-driven recruiting?

Data-driven recruiting goes wrong in 5 predictable ways. Each one has a cheap fix.

Vanity metrics

Applications received, messages sent and profiles viewed all rise with effort and say little about hiring. Gem's split of roughly 90% of applications against about half of hires shows how far volume and outcome can drift apart. Replace each volume number with the rate that follows it: applications become stage conversion, messages sent become reply rate.

Small samples

An offer acceptance rate of 50% from 2 offers is 1 decline. My rule: never show a rate without its n, and never act on a rate built from fewer than about 20 events. Pool similar roles or wait a quarter instead.

Bias in historical data

Past hires reflect past decisions, including the biased ones. A model, a scoring rule or a recruiter's instinct trained on who got hired before will repeat those patterns. Check pass rates by group at every stage, not only at the end. Ranking that learns from your shortlists will learn your habits too, which is one more reason to audit those shortlists. Our guide to AI bias in hiring covers how to run that check, and The Cognitive has not published a bias audit.

Optimising one number

Push time to first interview alone and the bar quietly drops. Pair every speed metric with a quality metric and review them on the same page, so a gain on one that costs the other is visible the week it happens.

Someone else's benchmark

Benchmarks come from other teams' stage names and other roles. They are useful context and a poor target. Our funnel conversion rates guide dates and sources every figure it shows, and your own trailing 12 months, split by role family, is the comparison that fits.

Where does The Cognitive produce recruitment data?

The Cognitive produces sourcing data and per-candidate interview reports. It does not replace your ATS reporting or a recruitment analytics tool. Here is what it records, so you know what you can pull from it:

• Sourcing yield per search. Each search ends with "Search complete" and the number of candidates found, and every person you Shortlist or Hide is recorded against the role. Shortlisted ÷ found is your yield for that search. Names show on the cards; contact details stay hidden until you reveal them.
• Contact reveal outcomes. Revealing an email costs 5 credits and a phone number 10, charged only when a value comes back. When nothing comes back, the card says "No email available" or "No phone available", so your reach rate per role is countable.
• Replies by intent. On the Sourcing Pro plan, email and SMS sequences go out from platform mailboxes, and a reply or opt-out on either channel stops both. Replies are labelled Interested, Needs info, Question, Not interested or Opted out.
• Interview stage and dates. Each invited candidate carries a status (Ready to Invite, Invite Sent, Scheduled, Completed, Pending Review, Approved, No Show) and the app records when the invite went out, when the interview was scheduled and when it was completed.
• Scored reports. 1 to 5 per criterion, a weighted score out of 100 and a suggested verdict, on a rubric fixed per role. The read-only share link carries scores, written feedback and the recording, not the transcript.

What it does not do matters just as much. It is not an ATS: no job posting, no offers and no funnel analytics. It does not search your ATS history, so the rediscovered candidates Gem counts still live in your own system. It connects to 60+ ATS platforms, and for candidates imported from your ATS it writes back a note with the score, a summary and the report link. People sourced inside The Cognitive are not written into the ATS automatically, which is why the tagging rule above matters. AI Sourcing starts at $49/month and AI Interview at $99/month, on monthly plans you can cancel anytime.

New accounts get a free trial with 100 sourcing credits.

Who is The Cognitive built for?

It is built for recruiters and founders who want better data at the top of the funnel: a search they can check, a shortlist that teaches the ranking, and interview scores they can compare across every candidate for a role. It suits teams that fill roles by sourcing rather than waiting for applicants, and teams whose interviewers are the bottleneck.

• Keep your ATS as the system of record for stages, offers and source of hire.
• Keep your analytics tool or ATS reports for funnel charts across every role.
• Pair The Cognitive with them for sourcing yield, reach, replies and structured interview scores, and tag what you bring across.

Measure the top of your funnel on one role Write the brief in plain English, see sourcing yield per search, and interview the people you shortlist live with AI. Start free

How should you start with data-driven recruitment this month?

Start with one open role for 4 weeks, not the whole team. Week 1, lock the source list and the stage names, and write the 9 definitions on one page. Week 2, run your searches and log yield per search before any outreach. Week 3, send outreach and track reply rate by message and by channel. Week 4, read the numbers with their n beside them and make one change: a new brief, a new message or a new channel. Then do it again on the next role.

If you want sourcing and interview data from the same place while you do it, sign up for The Cognitive and run that first role there.

Run one role the data-driven way Search ~900M public profiles from a plain-English brief, reveal contacts only when you need them, and get a scored report for every interview. Start free

Sources

• Gem, "Key takeaways from the 2026 Recruiting Benchmarks Report", published December 2025: gem.com/blog, read on 4 October 2026.
• Patrick Gavan O'Shea and Adrienne Fox Luscombe, "Is cognitive ability the best predictor of job performance?", SIOP TIP, 6 April 2024: siop.org, read on 4 October 2026.
• Sackett, Zhang, Berry and Lievens (2022), "Revisiting meta-analytic estimates of validity in personnel selection", Journal of Applied Psychology 107(11), 2040-2068: PubMed record, read on 4 October 2026.
• The Cognitive, pricing: thecognitive.io/pricing, read on 4 October 2026.

All posts · thecognitive.io