Refolk
FrameworkProcess, data, and compliance

The Duplicate Person-Pair Verdict: Merge, Link, Keep Separate, or Escalate

Score any flagged pair of person records, assign a verdict a second reviewer would match, and merge without losing an opt-out, objection, notice date, or source.

21 min readLast reviewed November 18, 2026Read as Markdown

Key takeaways

  • Salesforce, HubSpot, Greenhouse, Lever, and JazzHR all document merges as impossible to undo, so a false merge is the costlier error and doubt should resolve to Link or Keep separate.
  • A pair earns Merge only when two identifiers captured independently agree at the strong rungs and all four trap checks pass: role inbox, same-name colleague, recycled identifier, and job move.
  • Email and phone evidence decays with age: Yahoo and GMX treat addresses as inactive after 12 months, Google and Microsoft hold inactive accounts for two years, and 35 million US phone numbers sit disconnected.
  • HubSpot keeps the most marketable status on merge, and the primary email's subscriptions govern sends, so an opt-out held only on the non-primary record can silently stop applying.
  • Suppression and objection fields take the most restrictive value, data-acquired and notice dates take the earliest, and origin takes the earliest record's value, whichever record is primary.
  • Refolk's index shows 195 recruiting-ops profiles in the US and 27 in the UK, so second review has to be reserved for Escalate verdicts and merges that change compliance state.

This is the working standard for a standing queue of flagged "possible duplicate" person records, whether that means candidates in an ATS or contacts in a CRM, reviewed one pair at a time, every week. It is for recruiting ops, rev ops, and whoever has to answer for how the data was gathered. It gives you an evidence ladder, four trap checks, four verdicts a second reviewer should reproduce, and a per-field survivorship table that keeps opt-outs, objections, and notice dates from getting weaker in a merge.

Why is a false merge the costlier error?

Every major ATS and CRM documents merging as irreversible, and each one settles conflicts in favour of the primary record. A wrong merge therefore destroys data you cannot get back. A missed duplicate costs one more look next week. That difference sets the default for this whole standard: when in doubt, do not combine the records.

First, two terms. The primary record (Salesforce calls it the master) is the record that survives the merge. Survivorship is the rule that decides which value each field keeps when the two records disagree.

This is what each tool does by default:

ToolConflict defaultUndoDocumented exception
SalesforceMaster supersedesIrreversible; 15-day Recycle Bin (third-party source)None found
HubSpotPrimary prioritizedNoneLegal basis: most recent kept; marketing status: most marketable
GreenhousePrimary; auto-merge keeps oldestNoneNone found
LeverPrimary preferredNoneNone found
RecruiteeNot retrievedNot retrievedOrigin and creation date from oldest profile

Read the exceptions column carefully, because that is where compliance state leaks out. Several behaviours bear directly on the pair in front of you:

  • Salesforce combines up to three records into one and deletes the others. In Apex, the master's field values always supersede the values on the records being merged in. The 15-day Recycle Bin window comes from a third-party source, not from Salesforce's own merge documentation, so do not plan your recovery around it.
  • HubSpot creates a new record with a new ID for the merged result. After the merge, check anything keyed to the old IDs, such as sync mappings, exports, and lists kept outside the CRM.
  • Gem automatically sets every application the merged candidate had to Rejected. If a merge goes wrong, recovery means deleting the people and adding them back from your own snapshot.
  • JazzHR keeps the older profile's workflow stage when two candidates on the same job are merged.
  • Lever supports bulk merge from list view. That is convenient, and it skips every check in this guide.
A false merge cannot be undone. A missed duplicate costs one more look next week.

How do I score the identity evidence on a pair?

Mark every identifier the two records share as agree, disagree, or missing. Then place each agreement on a four-rung ladder according to what it proves on its own. The pair's evidence band comes from counting agreements that were captured independently at the top two rungs, not from how many fields happen to match.

You cannot hand this job to the duplicate flag, because the tools disagree about what counts as a match. Greenhouse flags a duplicate on the same email address, phone number, LinkedIn profile URL, or name. Gem's merge prompt requires matching first and last names and the primary email to appear on the other profile. HubSpot's sync to Salesforce deduplicates by email address only, and same-name records or other matching fields are not treated as duplicates. Three tools give you three definitions of a match, and none of them looks at compliance state.

RungIdentifiersWhat it provesHow it lies
3: AnchorSame public profile URL, captured separatelyOne account holderURL pasted into both records by one sourcer
2: Strong, decaysPersonal email; mobile numberSame holder when capturedAddress or number recycled to a new person
1: SupportingFull name; employer; work domain; locationConsistent with one personColleagues, common names, job moves
0: Not evidenceRole inbox; switchboard; name aloneNothing about the individualMerges different people into one

Turning rungs into a band

The bands follow the Fellegi-Sunter model, the 1969 record-linkage model that uses two thresholds to split match weights into three groups: links, non-links, and possible links. No published source gives universal numeric bands for matching people, so I count agreements instead of computing weights:

  • Strong: at least two independent agreements at rung 2 or 3, and no disagreement at rung 3.
  • Possible: exactly one agreement at rung 2 or 3, or several agreements at rung 1 only.
  • Weak: agreements at rung 0 only, or a disagreement at rung 3 (two distinct, live public profiles).

"Independent" means each value was captured separately. If one sourcer pasted the same profile URL into both records, that counts as one observation, not two.

Disagreeing emails do not show that the records belong to two people. People keep both personal and work addresses, and Greenhouse's flag is known to miss candidates who applied with a personal email one time and a work email another. Score email and phone disagreements as neutral.

Decay: hard identifiers go soft with age

An email address or phone number identifies whoever held it when it was captured, not whoever holds it now. Downgrade a rung-2 agreement to rung 1 when the activity on the two records is further apart than the provider's recycling window (see the table in the next section). I found no published reassignment window for phone numbers. Set one locally, write it down, and apply it the same way every week.

Which trap checks catch the classic false matches?

Four checks catch most of the false matches and missed matches in a queue of person pairs: a shared role inbox, a same-name colleague, a recycled identifier, and a job move. Run all four on every pair before you assign a verdict. Log each one as passed or failed, with a one-line reason.

Role inbox

Look at the local part of the email, meaning everything before the @. If it is a function word such as hiring, sales, careers, or talent, the address belongs to a team, and matching on it merges different people into one record.

  • Passes: the local part names a person.
  • How it lies: an alias that looks personal but forwards to a shared mailbox. You only catch it when the record's activity clearly comes from more than one person.
  • Limit: no public rate exists for this failure. Your own audit sample is the only measurement you will get.

Same-name colleague

The same full name at the same employer means two colleagues until something else agrees. HubSpot's sync does not treat a name match as a duplicate, and neither should you.

  • Passes: a second, independent identifier at rung 2 or 3 agrees.
  • Fails: name and employer are the only things that agree.

Recycled email or number

An identifier that was once reliable becomes unreliable after a provider reassigns it.

IdentifierWindow or scaleSource
Yahoo emailInactive at 12 monthsSummary of Yahoo help
Google / Microsoft emailInactive accounts held two yearsSame summary
GMX / WEB.DE email12 monthsSame summary
US mobile numbers35M disconnected; 171 of 259 sampled hijack-vulnerablePrinceton study
  • Passes: the activity timelines on both records fit one person, with no dormant gap longer than the provider's window between the last activity on one record and the first activity on the other.
  • Fails: record A went quiet for longer than the window, and then record B appeared with the same address or number. Downgrade that agreement to rung 1 and score the band again.

Job move

This check catches the opposite error, a missed match. A different employer and a different work email can still belong to the same person.

  • Passes: a public career history shows the move happening between the dates on the two records.
  • Fails: the history shows two different people who share a name.

Finding that history used to mean a separate search on each source. With Refolk, you ask in plain English for the person by name and previous employer, and you get their public profile from LinkedIn, GitHub, and the open web together. That usually tells you whether the move is real before you open anything else.

Which verdict does the pair get?

Assign the verdict from the evidence band and the trap checks:

  • Merge when the evidence is strong and all four checks pass.
  • Link when the evidence is possible and the checks pass.
  • Keep separate when the evidence is weak or a check shows two people.
  • Escalate when strong evidence runs into a failed or unresolved check.

Two reviewers who apply the ladder and the checks to the same pair should reach the same verdict.

Person-pair verdict matrix

A trap check failed or is unresolvedAll trap checks passed
Keep separate
A trap shows two people; log the failed check so the flag does not come back next week
Escalate
The evidence says one person and a trap says otherwise; send to a second reviewer
Link
Plausibly one person, not proven; cross-reference both records and do not combine them
Merge
Snapshot both records, set survivorship by hand, then merge
Weak or possible evidenceStrong evidence
The verdict comes from two readings: how strong the identity evidence is, and whether every trap check passed.

What each verdict commits you to:

  • Merge: combine the two into one record using the survivorship table below. This cannot be undone.
  • Link: leave both records intact and add a cross-reference on each one that names the other record's ID and the date it was reviewed. This is the possible-link band from Fellegi-Sunter, held where it is instead of being forced into a decision. Look at the pair again when a new identifier arrives.
  • Keep separate: log the pair as reviewed and not a duplicate, and name the check that failed. Where your tool lets you, dismiss the flag so the pair does not come back every week.
  • Escalate: send the pair to a second reviewer, or to the privacy owner if a rights request is involved. Nothing is combined until that person decides.

Overrides that beat the matrix

Some conditions change the verdict whatever the matrix says:

  1. Active applications on both records: Link, not Merge, until one of them closes. Gem rejects the merged candidate's applications, and JazzHR keeps the older profile's stage.
  2. Suppression, objection, or notice fields differ: Merge only with a second approval and survivorship set by hand.
  3. An open objection or rectification request on either record: Escalate. The ICO expects a response to a rectification request within one month, and merging in the middle of one blurs what was actually corrected.
  4. Referral or sourced credit on the record that would not be primary: Escalate or Link. One practitioner source puts it plainly: "Duplicate detection and merges often prioritize 'one person record' over 'clean attribution.'"

Sizing the review band to the people you have

A second reviewer on every pair is the safest approach, and almost nobody can staff it.

195
People with recruiting or talent ops titles in the US, in Refolk's index
The same index holds only 27 in the UK, so most teams that run this queue have one or two people to do it.
SegmentCountDerived
Recruiting / Talent Ops, US1957.2x the UK
Recruiting / Talent Ops, UK27Baseline
Revenue Ops, US3471.78x US Recruiting Ops

These counts come from title matching in Refolk's index, so they are lower bounds: variant titles are missed. Even so, the shape is clear. Recruiting ops is a thin function, and the Fellegi-Sunter design aims to keep clerical review to a minimum for exactly this reason. One practitioner source recommends two steward approvals for merges whose business impact is above a set threshold. In this standard, that threshold is concrete: a second reviewer sees every Escalate, and every Merge that changes suppression, objection, notice, or attribution state. Every other pair gets one reviewer and a logged reason.

If the review band still outgrows your staff, the fix is another reviewer, not a looser band. Refolk can find one in a single request, for example a peer at a company that runs the same ATS, who can also sanity-check your house rules.

Step-by-step procedure for one flagged pair

Work each pair through the same eight steps, in order, and do not skip the snapshot. Each step has a done-state that you can check in the verdict log.

From flag to verified merge

  1. Pull the pair into the queue
    Log both record IDs, the rule or tool that flagged the pair, and the date it was flagged. HubSpot compares property values daily, so expect the queue to refill continuously.
  2. Score every shared identifier
    Mark each identifier as agree, disagree, or missing, place each agreement on the ladder, and apply decay to email and phone agreements. Done when the pair has a band.
  3. Run the four trap checks
    Check for a role inbox, a same-name colleague, a recycled identifier, and a job move. Done when each is logged as passed or failed with a reason.
  4. Assign the verdict
    Apply the matrix, then the overrides. Done when Merge, Link, Keep separate, or Escalate is logged with a one-line reason.
  5. Get a second review where required
    Every Escalate, and every Merge that changes compliance or attribution state, goes to a second reviewer who scores the pair without seeing the first verdict.
  6. Snapshot both records
    Capture both records in full, including the activity feed, applications, stages, and compliance fields. Gem's own recovery advice is a screenshot of details and activity feed taken before deletion.
  7. Choose the primary and set survivorship
    Fill the survivorship map and set each winning value on the primary before merging, because the master's values supersede the others.
  8. Merge and verify
    Execute the merge, then re-check subscription status, objection flags, dates, origin, and stage by hand. Log the new record ID if the tool issues one.

Watch the ordering problem. Greenhouse auto-merge keeps the oldest profile as primary, and HubSpot's bulk merge lets you choose the primary by engagement or creation date. Either way, the system can choose the primary at step 7 before anyone has done step 2. If auto-merge is on, treat every auto-merged pair as already past step 7, and audit it against the survivorship map after the fact.

Per-pair verdict log entry
Record A ID:                 Record B ID:
Flagged on:                  Flag source:
IDENTIFIERS (agree / disagree / missing, rung, decayed y/n)
  Public profile URL:
  Personal email:
  Mobile number:
  Work email:
  Name + employer:
BAND: strong / possible / weak
TRAP CHECKS (pass / fail, one-line reason)
  Role inbox:
  Same-name colleague:
  Recycled identifier:
  Job move:
OVERRIDES: active apps on both / compliance fields differ / open rights request / attribution on non-primary / none
VERDICT: MERGE / LINK / KEEP SEPARATE / ESCALATE
Reason:
Reviewer 1:                  Reviewer 2 (if required):
Snapshot location:
Surviving record ID after merge:

One entry per flagged pair; keep it beside the record IDs so a second reviewer can re-score blind.

Survivorship rules: which value wins in each field

Descriptive fields such as title and employer take the most recent verified value. Suppression and objection fields always take the most restrictive value, notice and data-acquired dates take the earliest, and origin takes the earliest record's value. Never accept the tool's primary-wins default for any of these.

Survivorship precedence

  1. Suppression and objection
    Opt-outs, unsubscribes, objection flags; most restrictive wins
  2. Notice and legal basis
    Data-acquired date and first notice date (earliest wins); legal basis reviewed by hand
  3. Origin and attribution
    Earliest record's origin, sourcedBy, and referral credit; never blanked
  4. Pipeline state
    The live application's stage, not the older profile's
  5. Descriptive fields
    Title, employer, location; most recent verified value, history kept
Resolve the outer layers first; a field in an outer layer is never overwritten by a choice made for an inner one.
FieldSurvivesWhy
Opt-out / unsubscribeMost restrictive of the twoHubSpot keeps the most marketable status
Objection flagSet if either record has itICO: never delete from a suppression list
Data-acquired and notice datesEarliestThe notice clock runs from first acquisition
Legal basisReviewed by handHubSpot keeps the most recent value
Origin and source creditEarliest record's valueKeeps sourced-before-applied history
Pipeline stageThe live application's stageJazzHR keeps the older stage

The reasoning behind each compliance row:

  • Objection. Under the ICO's guidance, people have an absolute right to object to processing for direct marketing. An objection that disappears in a merge has been overridden in practice.
  • Notice dates. When you did not collect data from the person directly, they get privacy information no later than one month after you obtained it. A sourced record that predates the application started that clock, so the earliest data-acquired date governs. If that date is more than a month old and neither record shows a notice sent, escalate the pair, because the notice is overdue.
  • Legal basis. HubSpot keeps the most recent value from both records, which may not be the one that applies. Read both and decide.
  • Origin. In Lever, origin belongs to the opportunity, and each opportunity has exactly one: Applied, Sourced, Referred, University, Internal, or Agency. Lever's sourcedBy field is null when the opportunity was not sourced, so a null there means "not sourced", not "missing data". Recruitee combines all source tags but keeps origin and creation date from the first profile created. That behaviour matches this rule.
  • Rectification. If the person corrected a value on one record, that value survives regardless of which record is primary. I found no public regulator guidance on keeping an audit trail for rectification during a merge, so record it in your own verdict log.

To execute the merge, set every winning value on the primary record first. In Salesforce Apex, you keep a merged field value by setting it on the master before the merge runs. After the merge, update the merged contact's subscription status by hand, whatever the tool claims to have done.

Per-pair survivorship map
FIELD                       | RECORD A | RECORD B | SURVIVES | RULE
Email opt-out / unsubscribe |          |          |          | most restrictive
Objection flag              |          |          |          | set if either has it
Data-acquired date          |          |          |          | earliest
Privacy notice sent date    |          |          |          | earliest; note both
Legal basis                 |          |          |          | reviewed by hand
Origin / sourcedBy          |          |          |          | earliest record
Referral credit             |          |          |          | keep; never blank
Live application + stage    |          |          |          | live application's stage
Primary email               |          |          |          | the address carrying any opt-out
Title / employer            |          |          |          | most recent verified; keep history

Fill every row before merging; the SURVIVES column is what you set on the primary record first.

How this goes wrong: failure modes and false positives

Most bad merges in a person-pair queue come from seven repeatable mistakes. Each one has a check you can run before you merge. Read this section before you trust any band or verdict.

  1. Treating email as a hard key. A recycled address joins two different people. Check whether the provider has a recycling window of 12 months or more, and whether the activity dates on both records fit one person.
  2. Same name, same employer. These are colleagues, not one person. HubSpot's sync does not treat a name match as proof, and most tools agree. Require a second, independent identifier.
  3. Shared role inbox. Addresses like hiring@ or sales@ merge different people into one record. No public rate exists for this. Check whether the local part is a function word.
  4. Primary wins on opt-outs. The non-primary record's opt-out stops governing sends because the primary email's subscription statuses apply. Compare the suppression fields on both records before you choose the primary.
  5. Lost attribution. Referral credit disappears because the merge favours one person record over clean attribution. Check source, origin, and referrer on both records, and set them on the primary.
  6. Pipeline side effects. Gem rejects applications on the merged profile, and JazzHR keeps the older profile's workflow stage. Check active applications and stages before merging. "Oldest wins" protects origin but not stage: when a sourced record predates an inbound application, the origin comes out right and the live stage can come out wrong.
  7. Matching that is too narrow. Greenhouse's flags catch exact email matches but miss candidates who applied once with a personal email and once with a work email. The queue shows you some of your duplicates, not all of them. Expect missed matches to surface later as job-move pairs.

There is one false positive about the method itself. A clean-looking band does not mean a clean merge. A pair can score strong, pass all four traps, and still destroy an objection flag if you let the tool choose survivorship. Banding answers "is this one person". Only the survivorship map answers "what will this person's record say afterwards".

Keeping the queue honest week to week

A standing queue drifts in two ways: your false-merge rate creeps up as the backlog pushes people toward bulk merge, and vendors change how merging behaves. Measure the first from your own sample and re-test the second whenever release notes mention merging.

  • Measure your own false-merge rate. Each week, pull a sample of completed merges and have someone who did not make them re-score them blind from the snapshots. One practitioner blog suggests pausing auto-merge when the false-merge rate goes above 0.2%. That is a blog figure, not a benchmark, so use it only until you have a rate of your own.
  • Do not let the backlog choose the method. Bulk merge in Lever and HubSpot skips the trap checks and lets the system pick the primary. If the queue is growing, add review capacity or tighten the flag rules. Do not batch-merge.
  • Re-test merge behaviour when vendors change it. HubSpot has already published a changelog entry changing how merges work for contacts and companies, which shows these defaults are not fixed. Keep one sandbox pair with an opt-out on the non-primary record and an active application on each record. Merge a copy whenever release notes mention merging, and compare the result against the default-behaviour table at the top of this guide.
  • Revisit Linked pairs. Each week, re-score any Linked pair that has gained a new identifier since it was last reviewed.

Before you close a pair

  • Both record IDs, the flag source, and the flag date are in the verdict log.
  • Every shared identifier is marked agree, disagree, or missing, with a rung.
  • Every email and phone agreement was checked against the provider's recycling window.
  • All four trap checks are logged as passed or failed, with a reason.
  • The verdict and a one-line reason are logged, along with any override that applied.
  • A second reviewer approved every merge that changes suppression, objection, notice, or attribution state.
  • Both records were snapshotted, including the activity feed and applications, before the merge.
  • The surviving record carries the most restrictive opt-out and any objection flag from either record.
  • The surviving record carries the earliest data-acquired date and the earliest notice date.
  • Subscription status, origin, and live application stage were re-checked by hand after the merge.
  • The new record ID is logged, and anything keyed to the old IDs has been updated.

Questions practitioners ask

How do I tell if two candidate records are the same person?

Two records are the same person when at least two identifiers captured independently agree, such as a public profile URL and a personal email, and none of the four trap checks fails. A name match alone proves nothing, and HubSpot's sync to Salesforce does not treat same-name records as duplicates either. Before you count an email or phone agreement, check it against the provider's recycling window, because an address or number can pass to a new holder.

Should I merge duplicate candidates who both have active applications?

No. Link them and hold the merge until one application closes. Gem automatically sets the merged candidate's applications to Rejected, and JazzHR keeps the older profile's workflow stage when two candidates on the same job are merged. Either behaviour can quietly drop a live candidate out of a pipeline. Check applications and stages on both records before you choose a verdict, not after.

Which opt-out survives when I merge two contacts?

Set it by hand, because the tool's default may not keep the opt-out. HubSpot keeps the most marketable status, and the primary email address's subscription types become the contact's, so an opt-out held only on the other address stops governing sends. Make the surviving record carry the most restrictive status of the two, then re-check subscription status after the merge. Salesforce's handling of its email opt-out field was not confirmed, so test it on a sandbox pair.

Can I undo a merge in Salesforce, HubSpot, or Greenhouse?

No. Salesforce, HubSpot, Greenhouse, and Lever all document merges as impossible to undo. A third-party source says the non-master Salesforce records sit in the Recycle Bin for 15 days, but that comes from outside Salesforce's merge documentation, so do not rely on it. Your only dependable recovery is the snapshot you took of both records before merging.

Should auto-merge be turned on for candidate duplicates?

Only if you audit its output. Greenhouse auto-merge keeps the oldest profile as primary, which means the system picks the primary before anyone has checked for an opt-out or an active application on the newer record. If you do use it, re-score a sample of auto-merged pairs every week and track your own false-merge rate. One practitioner blog suggests pausing auto-merge when that rate exceeds 0.2%, but that is not a benchmark.

What happens to source and origin when two candidate profiles merge?

It depends on the tool, so set it deliberately. Recruitee combines all source tags but keeps origin and creation date from the first profile created. Lever gives each opportunity exactly one origin, and its sourcedBy field is null for opportunities that were not sourced. Keep the earliest record's origin so that a sourced-before-applied history survives, and never let referral credit go blank.

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