The Talent-Map Inclusion Standard: Count, Flag, or Exclude

Deciding whether a single public profile belongs in a talent map is the step most mapping guides skip. They describe the workflow end to end, then stop at "define criteria" and leave the per-record call to the analyst's gut. This guide is the definition of done for one profile: the scorable bar that lets you grade any candidate as Count, Flag, or Exclude, so two analysts mapping the same segment produce the same population and a coverage number that survives scrutiny. It is for strategy and research teams, talent-intelligence analysts, and operators sizing a market who report coverage to someone who will push back on it.

Why a per-profile bar, not just a workflow

Coverage claims collapse because the inclusion decision is unwritten. The published formula is simple and widely agreed: coverage rate is mapped target organizations or role segments divided by the agreed target universe. The problem is never the division. It is that "mapped" means different things to different analysts, so the numerator is assembled from inconsistent inclusion calls and the whole figure becomes indefensible.

A standard fixes this by making the per-record decision gradeable. Every profile that enters the map gets exactly one of three statuses, and the status is recorded with its reason.

The discipline that makes this work is recording the evidence for inclusion, exclusion, and priority on every record. A map with statuses but no recorded reasons is a map you cannot audit, and an unauditable map is a vanity total. A practical map structure splits records into core targets, adjacent targets, referrals to explore, people needing research, and people excluded with a recorded reason - the last category is as important as the first.

What a profile must carry to earn Count

A profile earns Count only when it clears three independent bars at once: it maps to an in-scope company, it resolves to one real person, and its current-role field is verified inside the review window. Miss any one and the record is Flagged or Excluded. Those three bars decompose into a field set.

Practitioner glossaries converge on a standard set of fields worth capturing: name, job title, current company, previous companies, tenure in role and career, location, publicly available contact details, links to online social profiles, qualifications, languages, and nationality. For a consistent map you also want standardized fields for function, sector, scope, relationship owner, prior contact, restrictions, research status, and evidence strength. Standardized fields are what make the map searchable and reusable later.

Not every field is load-bearing for the Count decision. The acceptance bar is about evidence completeness - mapped people with the minimum verified fields needed for prioritization - not about filling every cell. The three bars below are the ones that decide status.

The fields that make these bars checkable are worth stating as a minimum. Below that line, the record is Flagged, not Counted.

The denominator is the real lever

Coverage is mapped divided by the agreed universe, so the denominator decides the number more than the numerator does. Freeze it before you count. An unfrozen universe lets two analysts report different coverage from the identical list of names, which is exactly the failure a standard exists to prevent.

The universe is almost always bounded by companies, not the whole market. Most talent maps focus on 20 to 50 target companies rather than the full market, and a focused scope of 20 to 40 target companies per role is what actually yields results. That focus is not a shortcut. More names do not mean more hires, and a tight map of forty reachable people beats a list of four hundred you will never contact.

That gap is why the acceptance bar and the coverage expectation have to be set per geography. The depth of a pool is geography-bound, so a "map 200 people" target is near-exhaustive in Germany and a thin slice of the US. Sizing the universe without knowing the real pool depth is guessing.

Figures from Refolk's index for the same Rust title-and-skill filter; the multiple is derived as 604 divided by 96.

Pool depth also concentrates geographically inside each country, which matters when you decide which hubs to work and which employers anchor the list.

Distribution from Refolk's index sample; treat as directional, not absolute, since each query samples a slice rather than the full pool.

Resolving one real person before you count

Count nobody twice and count no bot. Entity resolution - deciding whether two records describe the same human - is the gate between a raw list and a countable population. When there is no shared key, use weighted attribute matching and a hard stopping rule.

The documented confidence floor is clear: do not resolve on fewer than two High-weight attributes. The weighting comes from how discriminating each attribute is.

• Full name (exact): high weight, but discriminative power falls as the name gets more common.
• Date of birth: very high, the single most discriminating biographical attribute.
• Email address: very high, and a unique identifier once verified.

The practical sequence is deterministic first, probabilistic second. Try a deterministic email or SSO match; if two accounts share the same corporate email, merge the nodes. Only then fall back to probabilistic string similarity such as Jaro-Winkler and Levenshtein distance. This is the classical record-linkage problem formalized in the Fellegi-Sunter model of 1969, and open tooling like the Splink library implements exactly this approach.

Cross-source matching is where Refolk removes real friction. Asking Refolk for people who appear on both GitHub and LinkedIn, deduplicated to one person, does the deterministic-then-probabilistic resolution the paragraph above describes without you building a matching pipeline first. The point of the two-High-attribute floor is to stop the match before it guesses.

Count nobody twice and count no bot, or every coverage number you publish is inflated before you start.

Fakes are a measurable tax on any raw count. Platform automation catches crude bots well - LinkedIn reports blocking 94.6% of fake accounts automatically, with the other 5.4% caught by manual review, and it removed over 80.6 million fake accounts at registration in one recent half-year. But automation misses the mid-tier, single-operator fake. An informal consultant audit flagged roughly 8% of 500 connections as fake or hollow, with peers reporting 5 to 12%. That is not a scientific sample, but it sets the order of magnitude: de-duplicating on name will not remove a fabricated profile, so reverse-image the photo and confirm the employer exists before you Count a thin record.

Grading recency: confirmed versus stale

The current-role field is where coverage silently inflates, so grade it, not just the name. Name is permanent while title and company are the highest-decay fields, which means a map "re-verified" on name looks fresh while its most load-bearing field rots.

The decay rate is well benchmarked. B2B contact data decays at roughly 2.1% per month, compounding to 22 to 30% per year, so a quarter of your base is wrong within a year. That tracks the labour market: US median employee tenure is under four years, and 15 to 20% of professionals change roles annually, with some sources putting job changes at 25 to 30% a year. A current-role field untouched for twelve months is not a minor risk. It is more likely wrong than you would accept anywhere else in the map.

The re-confirmation procedure is a match-fetch-parse-refresh loop: match a record to the right profile, fetch the profile, parse it into clean fields, then refresh it before it decays, because cached snapshots go stale as people change jobs and titles.

Thresholds from pool-health benchmarks; the 12-month verification band is the one that directly gates the Count decision.

The grading procedure

Run the profile through eight steps in order. The first step freezes the denominator; the last reports coverage against it. Everything between assigns and records a status per record.

One sequencing note. Some sources argue that internal workforce assessment must precede external market mapping, so you know what you already have before you count what you do not. Market-mapping workflows that start from an external brief skip that step. Either order works; just do not let the internal assessment quietly redefine the frozen universe once counting has begun.

For a time-boxing reference, a focused map of 20 to 30 companies and 100 to 200 named candidates takes a sourcer two to four weeks full time. If your estimate is far below that for a comparable scope, you are probably skipping resolution or recency grading.

Count versus reachable-now

Report two numbers, not one. In-scope count is the full qualifying population; reachable-now count is the subset you can contact today. Collapsing them into a single figure misrepresents the market.

The two populations diverge for a structural reason. There is a real difference between the addressable candidate, a potential person to reach out to who does not necessarily match every keyword, and the ideal candidate who matches all criteria. The strongest people often self-exclude from active channels, so reporting only contactable people understates the true market. The long-list is the full population of qualifying candidates in the mapped market; it is not a shortlist of the easy-to-reach.

Package coverage so the client sees how many qualifying people exist, how they split across target companies, and which people are reachable now versus which need nurturing. That split is a feature of the report, not a caveat buried in a footnote.

How this goes wrong

Most inflated coverage numbers trace to one of seven failure modes. Each has a false positive it produces and a specific check that catches it. Keep this section open while grading; it is the part of the standard that earns its keep.

• Stale current-role passed as confirmed. An analyst Counts a profile whose title changed months ago. Check: reject any current-role field outside the review window. At 2% per month decay, a 12-month-old field is 22 to 30% likely wrong.
• Duplicate counted twice. The same person on GitHub and LinkedIn inflates the count. Check: require resolution to one node and refuse to merge on fewer than two High-weight attributes.
• Fabricated or bot profile Counted. A plausible profile with a stock photo and an empty employer. Check: reverse-image the photo and confirm the employer exists, because platform automation misses mid-tier fakes.
• Scope creep inflates the numerator. Adjacent titles quietly added so coverage looks higher. Check: every Count must map to a company on the frozen in-scope list. Scope creep kills most efforts before they start.
• Reachable-now conflated with in-scope. Only contactable people reported as "the market." Check: report both numbers separately.
• Denominator shrunk to flatter coverage. Dividing by a narrowed universe. Check: freeze the agreed target universe before mapping and never edit it mid-project.
• Thin record graded Count. A profile with a title but no verifiable evidence. Check: enforce the minimum-verified-fields bar before assigning Count.

The verification checklist

Run this before you call coverage done. If any item fails, the number is not yet defensible.

Keeping the standard current

A talent map is a dated snapshot, not a fixed asset, so the standard has to include a re-check cadence or it decays with the data. The mechanism to watch is the 2.1% monthly decay on current-role fields. Set a review window, measure the verified-within-12-months rate against the 40% yellow flag and 20% critical floor, and keep more than 70% of contacts touched within 90 days.

Re-check pool depth too, not just freshness. The 6.3x gap between the US and German Rust pools in Refolk's index is a live figure that moves as hiring moves, so re-run the sizing query when you reopen a segment rather than trusting a cached number. The fastest way to keep both the population and the resolution current is to re-ask for the segment and let the index return deduplicated, freshly indexed people, then re-grade against this standard. The criteria stay fixed; the data underneath them does not.