Refolk
PlaybookRecruiting and sourcing

The Referral Mining Session Playbook: Past Teams to a Credited Slate

Pre-build 30 to 60 former colleagues, run a 20-minute session, and leave with a deduped slate where every name is rated, routed, and credit-logged before contact.

24 min readLast reviewed November 7, 2026Read as Markdown

Key takeaways

  • In Ashby's data, referred candidates received offers at a 7.3% rate in 2024 against 0.2% for inbound applicants, but the referral offer rate was 12.1% in 2021, so the advantage is real and shrinking.
  • A pre-built list removes the employee's recall step but not their screening step, so the per-name rating from the session is the asset and the list is only scaffolding.
  • Overlapping tenure at a large employer does not prove two people worked together, so record 'worked directly' as its own field, separate from the overlap evidence.
  • No published referral policy covers names a recruiter put in front of the employee, and every policy settles disputes by timestamp, so credit status must be decided and logged before anyone is contacted.
  • In Beaman et al., men were as connected to qualified women as to qualified men, yet 77% chose to refer men; a targeted ask in a Behavioural Insights Team trial doubled the number of women referred.
  • Under GDPR Article 14(3), a pre-built list starts a notice clock of at most one month from the day the data is obtained, so date every list and send notice no later than first contact.

This playbook is for in-house recruiters, sourcers, and founders who have one hard-to-fill role and one teammate who came from the right companies. It turns that teammate's past teams into named, rated candidates, so you are not relying on whoever they happen to remember. You pre-build a list of 30 to 60 likely former colleagues from public evidence and run a 20-minute session. You leave with a deduped slate, and before anyone is contacted, every name on it has a rating, an agreed outreach route, and a referral-credit status logged in the ATS.

Why run a referral mining session instead of waiting for referrals?

Referrals convert far better than other channels, but a passive program only captures the people an employee thinks of on their own. A referral mining session is a scheduled meeting where a recruiter walks one employee through a prepared list of people they probably worked with. The employee only has to recognise names instead of recalling them, and their judgement still decides who goes forward.

Ashby's Talent Trends data is the most detailed public view of the gap between channels. Rows 2 and 3 come from Ashby via secondary summaries, and row 5 is my own arithmetic on row 2.

MetricReferralInboundSourced
Share of hires (Ashby, to Jun 2025)~18%43 to 52%~16%
Offer rate, 20247.3%0.2%n/a
Application to interview~40%~12%~8%
Top quality rating36%28%32%
Offer rate vs inbound (derived)36.5x1xn/a

The "top quality rating" row is the share of hires rated "Greatly Exceeds Expectations" in Ashby's quality-of-hire data. What the table shows is that the referrer has already screened the candidate before the recruiter sees them, and that is the whole mechanism. A pre-built list takes over the recall step but not the screening step. That is why this playbook treats the employee's rating as the asset and the list as scaffolding.

The list removes the recall step. It does not remove the screening step, so the rating is the asset.

Keep four limits in mind when you read those numbers:

  • The edge is shrinking. The referral offer rate fell from 12.1% in 2021 to 7.3% in 2024. Over the same period the inbound rate fell from 0.7% to 0.2%.
  • The share figures conflict. Ashby's inbound report has referrals steady at about 18% of hires. Its later productivity report says referrals and sourced hires each fell to about 17%, with inbound reaching 52% by Q1 2026. Treat 17 to 18% as the range.
  • Startups see fewer referral hires. Ashby puts referrals at 15% of startup hires against 18% across all company sizes.
  • Retention is not established. Some sources assert that referred hires stay longer, but I found no named benchmark with a referral retention figure. If retention is part of your internal case, pull it from your own ATS.

A secondary summary of Gem's 2026 benchmarks reports that employee referrals convert at 11x the inbound applicant rate. That points the same way, but I have not seen the primary figure.

Which employee and which role should you pick?

Pick one employee whose past employers match the companies you would target for this role, and one requisition that inbound is not filling. You are done when you have a named employee, one req, and the employee's agreement to a 20-minute session. I budget about 15 minutes for this with the hiring manager, but that is my working estimate, not a measured one.

Score the candidate referrer on these criteria:

  • Employer match. Their past employers overlap your target company list. This tells you their network points at the right places. It misleads when they spent six months there or worked in an unrelated function.
  • Function match. They worked in or beside the function you are hiring for. A backend engineer from the right company knows backend engineers there. They may know almost no one in its ML org.
  • Overlap depth. They stayed long enough at the target employer to have shared teams, projects, and reviews. Short stints produce lists the employee cannot rate.
  • Willingness and clarity. They agreed before you built anything, and they have seen the credit rule (see the credit section below). An employee who learns the credit rule after rating 40 names feels misled.
  • Rotation. You have not already mined this person's network for a similar role in the last few months.

Choosing the referrer controls the diversity of the slate, not just its size. Rubineau and Fernandez's tipping-point model places the lever in how much different groups refer, not in how homophilous their networks are. Their finding is that "if a firm can get its under-represented group to refer more, referral recruitment can be made neutral to job segregation, or even integrative." If you always mine the same profile of employee, you will get the same profile of slate.

On timing, Peter Kazanjy observes that each new hire brings "~500 new contacts to the table to mine and recruit out of." Nobody has published a measurement of whether a session in someone's first few weeks yields more than one with a long-tenured employee. Log the referrer's tenure at each session and compare after a few rounds.

If several employees share an ex-employer, consider the group-first order Patrick Burke uses, as described in First Round Review: a group referral session followed by individual debriefs. One shared list saves build time, but rate each name one-to-one, because ratings drift when colleagues are listening.

How do you build the "did you work with these people" list?

Build the list from four kinds of public evidence, ranked by how strongly each implies that two people actually worked together, and aim for 30 to 60 names. No published study gives a minimum overlap threshold, so the thresholds below are Refolk's working method, not research.

Evidence that two people worked together

  1. Same employer, overlapping dates
    Co-presence only; weakest at large employers
  2. Same org unit or office
    Same function, building, or metro during the overlap
  3. Same named team
    Both profiles name the team, product, or group
  4. Shared artifact
    Co-authored paper or co-committed repository
Each layer inward is narrower and more likely to mean the employee actually worked with the person, so the outer layer needs a second signal before it goes on the list.

For each signal, here is what it proves and how it misleads:

  • Shared artifact (co-authored paper, co-committed repository). This proves both people put their names on the same piece of work. It misleads on papers with long author lists and on large repositories where hundreds of contributors never interacted. Count the authors or contributors before trusting it.
  • Same named team. This proves both profiles say they were in the same group. It misleads when the team name is generic, such as "Platform" or "Infrastructure", and covers hundreds of people.
  • Same org unit or office. This proves both people sat in the same function or location at the same time. It misleads at companies whose functions span many sites.
  • Same employer, overlapping dates. This proves co-presence and nothing more. At a large employer it is mostly noise, and it is the main source of "colleagues" the employee has never met.

Large employers come up again and again in any talent pool, which is why employer-and-dates matches fill a list so quickly.

10,178
Entry-level Machine Learning Engineer profiles in the United States, in Refolk's index
In a top-employer sample of 25 from that pool, Meta accounted for 3 and Intel for 2, and 6 of the 25 were in the San Francisco Bay Area.

When one employer recurs like that, a date-overlap search on it returns plenty of people who were in the building and almost nobody who shared a team. These are my inclusion thresholds:

  1. Shared artifact: include at any length of overlap, provided the paper or repository is small enough that contributors plausibly interacted.
  2. Same named team: include at 6 months or more of overlap.
  3. Same employer and dates only: include only at 12 months or more of overlap, in the same function and the same metro, and cap these at a third of the list.

Tune these on your own hit rate, which is the share of listed names the employee says they worked with directly. If that share falls well below half, tighten the third rule first.

Building this by hand means cross-referencing employment dates, team names, locations, and contributor graphs for every candidate name. Refolk takes that as a plain-English request: former employer, a date window, and current location in one ask, run across LinkedIn, the public GitHub graph, and the open web.

There is one real disagreement about order. Kazanjy has the employee "log in to LinkedIn, and do a 'first degree network' search" during the session, so the employee leads. This playbook has the recruiter pre-build the list first. My position is to do both in that order. The pre-built list catches people the employee never connected with online and puts evidence next to every name. The employee's own search catches people who leave no public trace. Spend most of the session on the list and keep the last five minutes for open recall.

Date the sheet on the day you build it. That date starts a legal notice clock, covered in the last section.

The referral mining session, step by step

The session itself is short. Most of the time goes into steps 2 and 3 beforehand and step 5 afterwards. SocialTalent's definition of referral sourcing frames the ask as "Can we spend 20 minutes walking through your network?", and the procedure below fills those 20 minutes with a prepared list instead of a blank page.

From one employee to a credited slate

  1. Pick the employee and the role
    Choose one employee whose past employers match the targets and one requisition inbound is not filling, and get their agreement after they have seen the credit rule.
  2. Pre-build 30 to 60 likely former colleagues
    Pull names from shared artifacts, shared teams, shared org units, and overlapping dates, recording the evidence type, in a dated working sheet with no ATS records yet.
  3. Dedupe the list against the ATS
    Check every name against the ATS before the session and tag it new or in pipeline with the date of last contact, flagging anyone inside the 6 to 12 month eligibility window.
  4. Run the 20-minute session
    Ask for each name whether they worked directly together and whether the employee would hire them, then use the last minutes for open recall of names the list missed.
  5. Log rating, relationship, and credit
    Record worked-directly, how they know each other and for how long, the would-hire rating, closeness, and the credit status, all timestamped in the ATS.
  6. Check the slate for gender skew
    Compare the gender make-up of the would-hire slate with the role's candidate pool and make a specific targeted ask if it is skewed.
  7. Choose the route and send notice
    Agree per name between a warm intro and recruiter outreach, and send a first message that states the data source, no later than one month after the list was built.

Set expectations for yield before you start. Kazanjy's employee-led version gives a rough baseline:

Employee-led referral yield (Kazanjy)

  1. Contacts a new hire brings
    ~500

    Kazanjy's estimate per new hire

  2. First-degree connections reviewed
    a couple hundred

    The employee's own LinkedIn search

  3. Names picked
    10 to 20

    5 to 10% of those reviewed, derived

Kazanjy's heuristic works out to roughly 10 to 20 names per employee, a derived figure that a pre-built list should beat on recall but not on screening.

Here is how I split the 20 minutes:

  1. Minutes 0 to 2. Restate the role in one sentence, then the credit rule in one sentence. Tell the employee that "never met" is a useful answer.
  2. Minutes 2 to 15. Go through the list quickly, at about 15 to 25 seconds a name. Ask two questions only: "Did you work with this person directly?" and "Would you hire them for this role?" Anything longer goes into a follow-up note.
  3. Minutes 15 to 20. Open recall ("who did the list miss?"), then the skew check and any targeted ask described below.

Names already tagged "in pipeline" stay visible during the pass, marked as such, so the employee can add context without expecting credit.

How do you rate and log each name?

Log each name as separate fields: worked directly (yes or no), how the two know each other and for how long, a would-hire rating, and closeness, all timestamped. Keeping them separate is what makes the slate auditable later, because evidence, relationship, and judgement each fail in different ways.

The relationship fields follow referral-policy templates, which recommend recording "how they know each other (e.g., former colleague, friend, university classmate, ex-manager, professional network) and how long they have known each other."

Session log, one row per name
List built on: YYYY-MM-DD
Referrer (employee):
Requisition:
Name:
Public evidence: shared artifact / same team / same org unit / employer and dates only
Overlap: employer, start month, end month, months overlapped
Worked directly: yes / no
How they know each other: former colleague / friend / university classmate / ex-manager / professional network
How long they have known each other:
Would hire: 3 hire again / 2 would interview / 1 would not vouch / 0 never met
Closeness: close friend / friendly / professional only
ATS status before session: new / in pipeline (date of last contact)
Credit status: employee referral / recruiter-sourced / ineligible / pending
Outreach route: warm intro by employee / recruiter outreach, may name employee / recruiter outreach, do not name employee
Logged at (timestamp):

Use these as column headers in the working sheet; copy each confirmed row into the ATS before the session ends.

"Would hire" and "closeness" are separate on purpose. A close friend can be rated 1, and a professional-only contact can be rated 3. If every 3 on a slate is also a close friend, the ratings are measuring affection rather than work.

The decision for each name comes from crossing the strength of the evidence with the employee's rating:

What to do with each rated name

Would hire againWould not vouch
Recall find
Confirm they worked directly, then treat as a referral like any other
Core slate
Credit to the employee, agree the route, contact first
Overlap false positive
Drop from the list and count it against your thresholds
Known, not vouched
No credit and no use of the employee's name; contact cold only on a strong profile
Employer and dates onlyShared artifact or team
The strength of the overlap evidence says how likely the name is a real colleague, and the employee's rating says whether that colleague belongs on the slate.

The bottom-left quadrant is feedback, not waste. Each name the employee has never met tells you which inclusion rule let it in.

Who gets referral credit when the recruiter surfaced the name?

No published referral policy covers names a recruiter put in front of the employee, so you need your own written rule. Show it to the employee before the session and log a credit status for every name before anyone is contacted. Published policies assume the employee came up with the name, and all of them settle disputes by timestamp. A pre-built list breaks that assumption.

This is what the published policies do cover:

PolicyRecency exclusionDuplicate ruleValidity
Lattice template6 to 12 monthsFirst timestampn/a
Covenant Health6 months, same positionFirst submissionn/a
PeopleStrong template6 months after exit (ex-employees)n/a3 months
Hyringn/aFirst in ATS; some split within 24 to 48 hoursn/a

Lattice's template says that a candidate who has "already applied or been in contact with our recruiting team within the last 6 - 12 months" is not eligible. Covenant Health states that "Recruiting Team Members are not eligible to receive a referral bonus." Put those two together and you have the trap. If the sourcer logs the list to the ATS first, a literal first-timestamp rule gives credit to recruiting, and recruiting cannot be paid it. The employee who actually vouched gets nothing, and nobody intended that outcome.

This is the rule I use. It is my method, not a published standard:

  1. Employee referral, credited. The employee worked directly with the person, rates them 2 or 3, and agrees to the outreach route. The recruiter surfacing the name does not reduce the employee's credit, because the vouching is what earns the referral premium.
  2. Recruiter-sourced, no credit. The employee recognises the name but did not work directly with them, or rates them 1. They may still be contacted cold, but not in the employee's name.
  3. Ineligible. The name was in contact with recruiting inside your recency window, typically 6 to 12 months. Tag it during the pre-session dedupe so the employee knows before rating it.
  4. Duplicates across employees. If two employees in separate sessions vouch for the same person, the first timestamp wins, as in the Lattice and Covenant rules. Hyring notes that some companies split the bonus when submissions land within 24 to 48 hours. Choose one approach and write it down.
  5. Expiry. Give credit a validity period. PeopleStrong's template uses 3 months from submission. Without one, a name rated once can claim credit for a hire two years later.

How do you check the slate for homophily skew?

Compare the gender make-up of the "would hire" slate with the candidate pool for the role. If the slate is skewed, make a specific, targeted ask before the session ends. A list built from one employee's past teams carries that network's skew, and the research shows that most of the skew comes from who people choose to refer, not from who they know.

StudySettingSame-gender referralIntervention result
Hederos et al.Swedish business school, n=453Men 75%, women 71%n/a
Beaman, Keleher, MagruderMalawi enumerator jobsMen 77%, women 43%Men referred qualified women when required to
Behavioural Insights TeamUK Ministry of Defencen/a2x as many women referred

The Beaman finding is the one to act on. Men in that study "were just as likely to be connected to qualified women as they were to qualified men," yet "77 percent of whom systematically referred other men when given the choice." Only 30% of referrals were women, against 38% of applicants. The qualified women were already in their networks. What changed the outcome was being asked for them.

2x
Women referred after a targeted ask, versus control
In the Behavioural Insights Team trial, an explicit, targeted referral request doubled the number of women referred for a vacancy.

A pre-built list makes the targeted ask concrete. You are not asking "do you know any women engineers?" in the abstract. You can point at specific names:

  1. Before the session, count the gender split of the list and of the wider pool for the role. In Refolk, a pool query such as "women machine learning engineers who previously worked at Apple, US-based" gives a comparison set for that slice. Inferring gender from public profiles is imperfect, so record your counts as estimates.
  2. At minute 15, compare the "would hire" slate with the list. If women on the list were rated less often than men with similar evidence, you have a choice effect, not a reach effect.
  3. Ask a specific question: "Of the people you worked directly with on that team, which women would you hire for this role?" Then go back through the women on the list who were skipped quickly.
  4. Record the counts and the ask in the session log, whether or not the slate changed.

How does referral mining go wrong?

Most of the damage happens before or after the session itself: a list that confuses co-presence with collaboration, a dedupe done too late, a credit rule that misfires, a skew nobody counted, or a notice clock nobody started. Each failure below includes the check that catches it.

1. Tenure overlap read as "worked together"

At an employer the size of Meta, an overlapping-dates match gives you a "colleague" the employee may never have met. The list looks strong and the session turns into a string of "no idea who that is." Check: ask "did you work with this person directly?" for every name and record the answer as its own field, separate from the overlap evidence. Track how often employer-and-dates matches come back as "never met".

2. Dedupe done after the session

The employee rates 40 names, assumes credit on all of them, and then learns that eight were contacted by recruiting last quarter. Under a 6 to 12 month recency rule those eight are ineligible, and the employee feels the session was bait. Check: dedupe against the ATS before the session and show the "in pipeline" tags during the pass.

3. The recruiter's list wins the timestamp

If the sourcer loads the pre-built list into the ATS as candidates, the recruiter's timestamp comes first on every name. A literal first-timestamp rule then hands credit to recruiting, and policies like Covenant's bar recruiters from receiving it. Check: keep the list in a working sheet until the session, and decide credit status per name before anyone is contacted.

4. Homophily carried over from the list

The list reflects the employee's past teams, so a skewed team produces a skewed list, and then rating by choice skews the slate further. Check: compare the slate's gender share with the list and the pool every time, and make the targeted ask while the employee is still in the session.

5. A list left past the notice clock

A pre-built list that sits unused for more than a month triggers the GDPR Article 14 notice deadline even if no outreach has happened. The clock runs from when you obtained the data, not from when you contact anyone. Check: date every list on the day it is built, and send notice or close the search within a month.

6. Ratings that measure friendship

In the referral literature, "one half of individuals refer a close friend." A 3 from a close friend and a 3 from a former tech lead do not carry the same information. Check: keep "would hire" and "closeness" in separate fields, and treat a slate where every 3 is a close friend as a slate that needs a second opinion.

7. A zero from a seniority filter

When I ran Machine Learning Engineer profiles through the "Senior" band in Refolk's index, both the United States and the United Kingdom returned 0. The entry-level band for the United States holds 10,178. A real absence of senior ML engineers is not plausible, so this looks like a filter artifact. Check: treat any zero from a seniority or title filter as a search problem until you have rerun the query with other seniority labels or with the title alone.

How do you choose the outreach route, send notice, and keep the slate current?

For each name on the slate, agree one of two routes: a warm intro from the employee, or recruiter outreach that may or may not mention the employee. Whichever route you choose, the first message carries the source notice and goes out no later than one month after the list was built. Under GDPR Article 14(3), notice is due "within a reasonable period after obtaining the personal data, but at the latest within one month." If you use the data to contact the person, notice is due "at the latest at the time of the first communication." The notice must "state the source of the data and whether it came from publicly accessible sources." This describes the mechanism and is not legal advice, so check your wording with counsel.

Choose the route per name:

  • Warm intro by the employee. Use it for names rated 3 where the employee agrees to send the intro themselves. You still owe the source notice in your first direct message, because you built the list from public data.
  • Recruiter outreach, may name the employee. Use it when the employee rated 2 or 3 but would rather not write the intro. Get explicit permission to use their name, and log it.
  • Recruiter outreach, do not name the employee. Use it for recruiter-sourced names and for anyone the employee asked you to keep them out of.
First recruiter message with source notice
Subject: <Role> at <Company>

Hi <first name>,

<Employee> and I were talking about the people they most valued working with at <past employer>, and they asked me to reach out to you directly.

I am hiring a <role> to <one-line description of the problem the role owns>. Based on your work on <specific team, project, paper, or repository>, I think it is worth a 20-minute conversation.

Where your details came from: I found your profile through publicly accessible sources (<public professional profile, public code contributions, published papers>), and <Employee> confirmed they worked with you. I hold your name, current role, and work history for this search only. Reply "remove" and I will delete them.

Would <two specific time slots> work for a short call?

<Your name>

Delete the second paragraph when the employee has not agreed to be named; keep the source paragraph in every version.

To keep the work current, close each session with the checks below. After three or four sessions, review your thresholds against the hit rate. A rule that keeps producing "never met" names should be tightened. Expire credited names on the validity period you set. Re-mine an employee's network only for a different role, or after several months have passed.

Before the slate goes to outreach

  • The working sheet is dated, and the date is less than one month old or notice has already gone out.
  • Every name was checked against the ATS before the session and tagged new or in pipeline with a last-contact date.
  • Every name has "worked directly" recorded as a separate field from its overlap evidence.
  • Every name has a would-hire rating recorded separately from closeness.
  • Every name has a credit status of employee referral, recruiter-sourced, or ineligible, timestamped in the ATS, and none reads pending.
  • The employee has seen the credit outcome for each name they rated.
  • The slate's gender share has been compared with the list and the pool, and any targeted ask is logged.
  • Each name has an agreed route, and the employee's permission is logged wherever their name will be used.
  • Every first message states the data source and whether it came from publicly accessible sources.

Questions practitioners ask

How many candidates should one referral mining session produce?

Expect roughly 10 to 20 usable names per employee if you follow Peter Kazanjy's heuristic that people pick 5 to 10% of a couple hundred first-degree connections. That is a derived figure, not a measured benchmark. With a pre-built list of 30 to 60 names, the 'would hire' slate is usually a fraction of the list, because many overlap matches turn out to be people the employee never worked with. Track your own hit rate across sessions and set expectations from that.

Who gets referral credit when the recruiter found the name first?

No published referral policy settles this, so you have to write your own rule and log each name's status before anyone is contacted. My rule: the employee gets credit when they worked directly with the person, rate them 'would hire', and agree to the outreach route. Names they only recognise count as recruiter-sourced. Keep pre-built names out of the ATS until the session so a first-timestamp rule does not hand credit to recruiting by accident.

Should the employee search their own LinkedIn connections, or should the recruiter build the list?

Do both, with the recruiter-built list first. Kazanjy runs it employee-led, with the employee doing a first-degree LinkedIn search during the session. A pre-built list catches people the employee never connected with and shows public evidence for each name. The employee's own search catches people who leave no public trace. Spend most of the session on the pre-built list and keep the last five minutes for open recall.

How long can I keep a pre-built referral list before I have to notify the people on it?

Under GDPR Article 14(3), at most one month from the day you obtained the data, or earlier if you contact them first. The notice is due no later than the first communication. It must also say where the data came from and whether it came from publicly accessible sources. Date every list when you build it, and check your notice wording with counsel, because this guide describes the mechanism and is not legal advice.

Do employee referrals still outperform other recruiting channels?

Yes, but by less than before. Ashby's data shows a 7.3% offer rate for referred candidates in 2024 against 0.2% for inbound, down from 12.1% in 2021. Referrals are about 17 to 18% of hires, depending on which Ashby report you read. I could not find a public benchmark for referral retention, so if retention is part of your case, measure it in your own ATS.

When is the best time to run a referral session with a new hire?

Nobody has established this publicly. No source I found measures whether sessions in a new hire's first few weeks yield more than sessions with long-tenured staff. Kazanjy notes that each new hire brings about 500 new contacts to mine, which argues for running a session soon after someone joins. Log the employee's tenure at each session and compare hit rates after a few rounds.

Try it on the search you came here for

Stop building boolean strings. Just describe the person.

Type one sentence. I plan the search, read GitHub, public LinkedIn and Crunchbase records, and the open web as it is right now, and hand back a ranked list with the reason next to every name.

  1. 01Describe them

    One plain sentence. Role, city, stack, stage, whatever matters to you.

  2. 02I read the web live

    GitHub, public LinkedIn and Crunchbase records, the open web. Not a database that went stale last quarter.

  3. 03You read the shortlist

    Ranked, with the reasoning under every name. Open a profile, ask a follow-up, narrow it down.

  • No boolean, no filters, no seat to buy. One box.
  • Read at search time, so a profile updated yesterday counts today.
  • Every step visible as it runs, every name with its reason.

500 free credits on sign-up. No card, no demo call. See real searches.

Read next