The Talent-Map Completeness Standard for a Named-Candidate List

You just finished a named-candidate talent map and someone is about to ask whether it is done. This guide is the pass/fail rubric for that moment. It is written for strategy and research teams, talent-intelligence analysts, and operators sizing a market, and it gives you fixed criteria - segment coverage, per-person evidence, deduplication, and staleness - so that two analysts grade the same map the same way and a stakeholder cannot dismiss it as "just some names."

Most talent-mapping guides tell you how to build the map and how long it takes. None give a gradeable definition of done. That gap is why completeness gets judged by feel: a full-looking spreadsheet reads as complete even when a fifth of its names are at new companies and its denominator was invented. This standard closes that gap.

What "complete" means for a named-candidate list

A named-candidate talent map is complete when it covers 80 to 90 percent of a segment you defined as observable criteria, when every record carries evidence, when the count is reported as distinct humans, and when the map has an owner and a next-refresh date. Miss any one of those and the correct verdict is incomplete or expired, not "close enough."

The four axes are independent, and a map can pass on three and fail on the fourth. A list with perfect evidence and clean dedup still fails if its coverage is measured against itself. A fully covered, freshly deduped map still fails if it shipped six months ago with no owner. Grade all four, every time.

The three verdicts are deliberately coarse. Complete means all four axes pass. Incomplete means coverage, evidence, or distinctness fails but the map is inside its refresh window. Expired means the map is past its cadence, so its rows can no longer be trusted regardless of how they scored when built. Expired is not a softer form of incomplete; it is a separate gate, because a six-month-old map of senior talent may already have 20 percent of names at new companies.

Coverage: measuring against a real denominator

Coverage is the ratio of mapped target organizations or role segments to the addressable population, and the target for an actionable map is 80 to 90 percent of the defined segment. The number is only meaningful when the denominator - the addressable population - came from somewhere other than the list you are grading.

This is where most maps quietly fail. If you define the segment as "the companies I mapped," coverage is always close to 100 percent, and the grade is a fiction. The denominator has to be estimated independently. The established method for sizing a hidden population is capture-recapture, which estimates population size from the proportion of overlap between independent incomplete lists. No talent-mapping guide I found applies this to candidate maps, so this bridge is load-bearing and not established in the recruiting literature. It is still the only defensible way to put a real number under the fraction.

In practice: run a LinkedIn pass and an independent GitHub or patent or conference pass over the same segment. The overlap between those two lists lets you estimate how many names neither list caught. Low overlap is direct evidence of undercount. High overlap means your independent sources are converging on the same people, which raises confidence that the segment is genuinely small.

The count anchor only becomes a grade once you divide it by the market. Refolk's index makes the point concrete.

Why a fixed name-count target lies

In Refolk's index of professional profiles, the United States returns 65,713 people with "Product Manager" titles and the United Kingdom returns 15,140 - a 4.34x smaller pool. A focused-map target of 100 to 200 named candidates is deep coverage of a narrow UK segment and a thin sample of the US. The exact same list would grade "complete" in one market and "incomplete" in the other. Only the denominator tells you which.

Table C - Pool-size denominator by market (Refolk's index)

Estimating a denominator by hand is the slowest, most error-prone step in this whole procedure, which is where a live index earns its keep. Refolk returns a total count for a described segment, so you can anchor the denominator to a real pool size before you start counting the names you found.

Evidence: what every record must carry

A record counts toward coverage only if it carries a current role and company, an estimated tenure, and a captured source, date, and evidence for each material fact, with an assigned relationship owner and research status. A row that has a title but no source or date is a bare name, and bare names do not count.

No single published source gives a consolidated minimum-viable-record schema, so this is a synthesis of two: the executive-search longlist output (current role, company, estimated tenure, engagement readiness) and the provenance requirement (capture source, date, evidence, and assign an owner). Treat the merged set below as the minimum. Anything less is "identified," not "verified."

name: full name (deduplicated to one distinct human)
current_role: title + company
tenure: estimated years in current role
evidence: prior companies, education, notable work, flight-risk signals
source: where the fact came from (channel, not just "LinkedIn")
captured_on: date the fact was verified
owner: named analyst responsible for this record

The enrichment layer - tenure, prior companies, education, signals of flight risk, notable work - is what turns a name into a candidate a stakeholder can act on. But the provenance layer is what makes the map defensible. When a stakeholder challenges a name, you point at the source and the capture date. When a name has neither, you cannot defend it, and you should not have counted it.

Distinctness: report humans, not rows

The count you hand over must be distinct real people, produced by entity resolution: matching record pairs, then clustering matched pairs so each cluster is one distinct real-world entity. At its core, entity resolution is about accurate counting - is this one person or three?

Deduplication errors run in both directions, and they are asymmetric. A false negative leaves two records for one person, inflating the count. A false positive merges two real people who share a common name, deflating it. Because the errors point opposite ways, a single accuracy number can hide a count that is simultaneously too high and too low. Report precision and recall, not one figure.

The guard against a bad merge is a rule: require a second corroborating attribute before merging two records. Employer plus tenure is a good pair. Two "James Smith" rows at the same company with the same tenure are probably one person; two at different companies are probably two people. Never merge on name alone.

One practical note on cost. Comparing every possible pair scales as n times n minus one over two, so a naive all-pairs comparison becomes expensive fast. For a 100-to-200-name map this is trivial; for a market-wide pass it is not, which is why blocking - only comparing records that already share a key like company or city - matters at scale. Open-source probabilistic record-linkage libraries handle both the clustering and the blocking.

Staleness: relevance expires before contactability

A talent map should be refreshed quarterly at minimum, with monthly updates for the highest-priority roles, and any map past its cadence is graded expired. The reason to gate on this separately is that a stale map looks complete: high row count, high confidence, no timestamps.

The field that defines segment membership decays fastest. Job titles decay about 65.8 percent a year, against a 22.5 percent aggregate contact-record decay - roughly three times faster. So a map can still be reachable while a large share of its records no longer match the segment criteria. This is why aggregate decay figures systematically understate map rot: the aggregate averages in slow-moving fields like company address, while the title that put a person in your segment is the first thing to change.

Table A - Field decay rates, fastest first

A concrete expiry signal: a six-month-old map of senior talent may have 20 percent of names at new companies. Set that against the underlying mechanism - 15 to 20 percent of professionals change jobs annually, and tenure runs shorter in tech - and the quarterly floor stops looking cautious.

One trap on refresh. Re-pulling stale data on cadence adds noise, not freshness; frequent updates without verification can introduce noise. A refresh must re-verify each field against a live source, not re-import the same export with a newer date on it. A new timestamp on an unverified row is a lie about freshness.

How this goes wrong: the failure modes

Every failure mode below produces a map that reads as complete and is not. Each pairs a false positive - the thing that fools the grader - with the check that catches it. This is the part of the standard that does the most work, because a map fails quietly, not loudly.

Self-referential denominator. Defining the segment as only the companies you mapped makes coverage always about 100 percent. The false positive is a map graded 90 percent complete that never estimated the true population. Check: the denominator must come from an independent estimate such as list overlap, not the list itself.

Easiest-profile skew. LinkedIn-visible, active people are over-represented, and those with no public profile - often the highest-value candidates who are not visible through any conventional sourcing channel - are missing. Full coverage of visible profiles then reads as full coverage of the market. Check: low overlap between independent lists (GitHub, patents, conference speakers) signals undercount.

Stale map reads as complete. A six-month map looks full, but around 20 percent of names are wrong. The false positive is high row count, high confidence, no timestamps. Check: a last-verified date per field, and grade expired past the cadence.

Dedup false negative. Name variants of one person counted as two inflate the headcount. Check: entity-resolution clustering before reporting distinct humans.

Dedup false positive. Two real people with a common name merged into one deflate the count. Check: require a second corroborating attribute such as employer plus tenure before merging, and report precision and recall.

Bare name masquerading as verified. A row with a title but no source or date. Check: every record carries source, date, evidence, and owner, or it does not count.

Refresh without verification. Re-pulling stale data on cadence adds noise. Check: a refresh must re-verify against a live source, not re-import the same export.

The procedure: building a gradeable map

Follow these seven steps in order. The first two are where most maps go wrong, because they are the two steps that no existing guide prescribes and the two that make the coverage number defensible. Sources disagree on ordering - some put internal assessment before external mapping, executive-search sources start at external company identification - but all agree the denominator estimate belongs somewhere before you grade, and this standard requires it.

Table B - Scope calibration anchors

Use these anchors to check that your scope is sane before you grade, not to grade. Research points to 5 to 10 critical roles as the right program scope; covering more dilutes the depth and actionability of each segment. A focused map of 20 to 30 companies and 100 to 200 names typically takes a sourcer two to four weeks full-time, and an executive longlist of 20 to 50 names completes in 5 to 10 days. If your map is far outside these bands, something is off with either the segment definition or the effort.

The completeness checklist

Run this before you hand the map over. Every item is checkable, and any single failure sets the verdict. If all items pass and the map is inside its cadence, the verdict is complete. If a coverage, evidence, or distinctness item fails, the verdict is incomplete. If the map is past its cadence, the verdict is expired regardless of the other items.

Keeping the verdict current

A complete verdict has an expiry date, so the work does not end at handover. Assign the cadence and owner as part of the deliverable: quarterly at minimum, monthly for the highest-priority roles. A map with no assigned owner and no refresh cadence is a snapshot that starts going stale the day it is delivered.

Two things make the refresh worth doing rather than performative. First, re-verify against live sources rather than re-importing an old export, because a refresh without verification adds noise. Second, re-verify the title field first, since it decays roughly three times faster than the aggregate record and it is the field that decides segment membership. If titles are current, the map still describes the segment you defined. If they are not, the map describes a segment that no longer exists, no matter how full it looks.

Set a calendar trigger on the next-refresh date and re-run this checklist against the map on that date. If it fails the staleness gate, mark it expired and either refresh or retire it. A map that has quietly passed its cadence is more dangerous than one that is honestly incomplete, because it is still being trusted.