Building a Living Sector Map for Deal Sourcing

A sector map is the artifact a thesis-driven investor keeps open all week: a maintained list of every company in a defined space, the founders behind each one, how each fits the thesis, and who to call first. This guide is for early-stage investors, platform and talent partners at funds, and angels who source against a stated thesis. It gives you the queries, the fields, the coverage test, and the maintenance loop that turn a one-time landscape slide into a living sourcing tool.

Most published advice on sector mapping stops at "define your thesis and use your network." That produces a static PowerPoint. What you need is a map that provably covers the thesis, tells you honestly where the whitespace is, stays current as founders move, and hands you a ranked contact list on any given Tuesday. That is the job here.

What a complete sector map actually requires

A complete market map is defined by how it was built, not by how many companies it contains. It requires a business-description-based search rather than a list of known competitors, coverage of long-tail private operators, verified ownership data, scale proxies, timing signals, and a live connection to your sourcing workflow.

Row count is a trap. A 2,000-row map can still miss an entire sub-segment if it was built from headlines. The discipline is to assemble from many sources, then test coverage against something you know is complete.

The sources you assemble from are wide: referrals, founder networks, accelerators, universities, communities, market maps, databases, inbound pitches, and targeted outbound. The mistake is to lean on generic large language models, which return the well-known names that show up in prominent industry articles. That gives you a press-ranked map, and press-ranked maps systematically under-cover the exact targets an investor wants, because nearly half of companies that get acquired are bootstrapped at the time, with no VC backing and little coverage.

Assembling the company set from business descriptions

Assemble the company set by searching on what companies do, not on names you already know. Business-description search surfaces the long-tail private operators that named-competitor lists never reach, and every row lands with a source tag so you can audit where coverage is thin.

The named-competitor approach fails on its own terms: it can only find companies adjacent to ones you already named. Business-description search asks the opposite question, "who does this thing," and returns operators you have never heard of. That is where whitespace lives.

Do this across the full source stack in parallel:

• Directories and databases for structured firmographics and stage.
• Accelerator and community rosters for early operators before press exists.
• Registries for freshly incorporated companies inside the boundary.
• Open build signals on GitHub and patent filings for technical founders who have not launched.
• Inbound and referrals, tagged by channel so you can measure source quality later.

Deduplicate on the way in and keep the source tag on every row. When you later find a coverage hole, the source tag tells you which channel to widen. A tool like Refolk does the business-description search directly - you describe the operator profile in plain English and it returns matching people and companies across GitHub, LinkedIn, and the open web, which collapses the manual cross-referencing this step otherwise demands.

Measuring coverage against a known-complete benchmark

Coverage is measured, not asserted, by rebuilding a universe you know is complete and computing recall. Rebuild one Y Combinator batch inside your thesis boundary using only your own queries, then divide the rows you found by the official batch size. That fraction is your recall, and it is the single most honest number your map produces.

The YC directory works as a benchmark because it is known-complete and free. It lists all Y Combinator companies since 2005, filterable by batch and industry, and updates with every new cohort. If your thesis touches a sector YC has funded, you have a ground-truth list to test against.

The critical move is to compute recall per sub-segment, not overall. A map can score 90% recall on funded, famous names and 30% on the unfunded long tail, and the average hides the failure. Segment your recall by stage and by press profile. If recall is high only on the companies that make headlines, your map is press-biased and you need to widen your registry and community sources, not your database subscriptions.

Linking every company to its founders

Every active row needs at least one named, verified founder, because you source people, not logos. Attach founder names and profile links, and for rows you pulled from a registry, cross-reference the officer name to a real profile before you trust it.

Directory-style sources ship founder identity: each company profile carries its founders with LinkedIn and Twitter/X links, plus company social links including GitHub and Crunchbase. Use those as the backbone. For registry-derived rows there is no built-in profile, so you must do the cross-reference yourself: match the officer name on the filing to a person online and confirm they are the operating founder, not a formation agent.

This is also where geography matters more than a map's aesthetics suggest. Founder pools are wildly asymmetric across markets, and a thesis that splits sourcing effort evenly across regions will over-cover thin ones.

Counts from Refolk's index of professional profiles; the ratio column is derived by dividing each count by the Germany count. In Refolk's index, the US financial-services founder pool is 15 times the size of Germany's and about four times the UK's. A "US plus Europe" thesis that allocates equal sourcing hours to each geography is mispriced: effort should track the size of the founder pool, not the symmetry of a slide.

Watching for new company formation inside the boundary

A new-formation watch catches companies before press coverage exists, which is where the timing edge lives. Monitor registries, domains, GitHub, patents, and hiring inside your thesis boundary, and route every hit into the map with a dated first-signal field.

The mechanism is simple and public. Any company that intends to employ people, sign contracts, or accept investment must incorporate, and that incorporation is recorded in a public government registry accessible within days of filing. In the UK, Companies House publishes SH01 share-allotment forms that can reveal stealth-stage funding before any announcement. A researcher who stops publishing papers and files a provisional patent signals a research-to-IP transition that often precedes company formation by three to six months.

Not every signal is a company. A lateral job move is not a startup, and a sabbatical is not a stealth launch. Read a registration or a network signal to confirm before you add a row. A first-time incorporation from someone who held a VP role at a funded startup is a far stronger signal than a generic LLC registration, so weight the person's history, not just the filing.

The step-by-step build

Run the build in order the first time, then let the refresh loop take over. The whole first pass is roughly a working week for an analyst plus a day of partner and reviewer time; the artifact it produces then persists.

There is a legitimate disagreement about order. Thesis-driven funds often front-load a deep white paper before broad list-building: they choose a sub-sector per quarter, pull together a landscape map, run industry calls, and reach out to the ten to fifteen top players. That is a valid variant when the thesis itself is still forming. The version above assumes the thesis is set and the job is coverage.

Scoring thesis fit and ranking who to contact

Thesis fit is scored on fixed dimensions so two people rank the same company the same way. Score alignment across stage, sector, check size, geography, and business model, stack the weighted dimensions into a 100-point rubric, and apply hard disqualifiers on top. High scorers close faster and expand more, so the rank is worth getting right.

The columns you carry make the score auditable. Documented VC deal-flow practice records the company, source, sector, stage, owner, thesis fit, next step, diligence notes, and final decision. Add a last-verified date and a first-signal date, and you can measure pipeline health with qualified rate, first-meeting conversion, diligence conversion, source quality by channel, time in stage, and pass reasons.

Stage (20): Company is at [seed/Series A] per last funding or headcount signal. Yes = 20, adjacent stage = 10, else 0.
Sector (20): Business description matches [defined sub-sector]. Yes = 20, tangential = 10, else 0.
Check size (15): Round size fits [check range]. In range = 15, one band off = 8, else 0.
Geography (15): HQ or founder base inside [geographies]. Yes = 15, else 0.
Business model (15): Model is [B2B SaaS / marketplace / etc]. Yes = 15, hybrid = 8, else 0.
Founder signal (15): At least one founder has [prior exit / senior role at funded startup / patent history]. Yes = 15, else 0.
Disqualifiers (auto-zero): Outside legal/ethical mandate, direct portfolio conflict, or confirmed dead.

Consistency between reviewers is measurable, not a matter of taste. Have two reviewers score a 20-row calibration sample, then compute the inter-rater reliability as an ICC. A documented rubric reached an overall ICC of 0.7, which counts as good agreement on the standard scale: below 0.5 is poor, 0.5 to 0.75 is moderate, 0.75 to 0.9 is good, and above 0.9 is excellent. That result came from concrete, behaviorally anchored criteria. Broad global ratings reproduce poorly, so a yes/no test will out-agree a 1-to-5 quality slider every time.

Keeping the map current: cadence tied to field decay

Refresh cadence is driven by how fast fields decay, and the fastest-decaying fields are the ones about people. Refresh people and status fields monthly, re-verify firmographics quarterly, and run an annual retrospective that connects sourcing data to fund outcomes. This split is the difference between a living map and a stale one.

The decay is uneven, which is the whole point. Company addresses barely move; job titles and roles move constantly. Set your cadence to the fastest-mobility field, not the average.

Job changes drive the most decay. Average job tenure is 2.8 years, so 30 to 40% of contacts change roles annually. A row can look completely live while the founder left months ago, which is why every row carries a last-verified date: anything older than the cadence is presumed wrong until re-checked. This is where the tiered review structure earns its keep - monthly pipeline reviews catch operational drift, quarterly strategy reviews reveal whether the sourcing mix still matches the thesis, and the annual retrospective ties it all to outcomes.

For hot sectors, quarterly-only cleanup is not fast enough: by the time you clean, 15 to 20% has already decayed again. Track the decayed-row rate between refreshes, and if it exceeds your tolerance, shorten the cadence. Manual verification is not a substitute for cadence, either - one exercise reached 91% accuracy but took 143 hours on 10,000 contacts, so verification effort has to be scoped, not applied everywhere at once. Refolk's re-run-the-query model is the cheap version of a monthly refresh: the same plain-English search that built a segment can rebuild it, and the diff shows you who moved.

How this goes wrong

The map fails in predictable ways, and the checks for each are cheap. The weakest point is almost never company coverage - it is the people fields and the method behind the counts.

• LLM-built maps look complete but miss the long tail. Generic models return the names that show up in prominent industry articles. Check: run the YC-batch recall test; if recall is high only on funded or famous names, the map is press-biased.
• Coverage measured by count, not method. A 2,000-row map can miss a whole sub-segment. Check: report segment-level recall, never total count.
• Stale people fields masquerading as current. A row looks live while the founder left months ago. Check: a last-verified date per row; anything older than the cadence is presumed wrong.
• Employment signal mistaken for a new company. A lateral shift or sabbatical read as a stealth launch. Check: require a business registration or network signal before adding a formation row.
• Registry noise. A generic LLC registration scored the same as a founded-startup VP's first incorporation. Check: cross-reference the officer name to a profile before adding.
• Rubric scores that don't agree. Subjective items reproduce poorly. Check: compute ICC on a calibration sample; below 0.5 means rewrite the rubric.
• Quarterly-only cleanup too slow for hot sectors. By cleanup time, 15 to 20% has decayed again. Check: track the decayed-row rate between refreshes and shorten the cadence.

The through-line: companies persist, founders move, and method beats volume. A map that is big and current-looking can be wrong in exactly the places an investor cares about most.

Keeping the artifact alive

The map is only worth building if it stays usable, so the last thing you set up is the discipline that keeps it honest. Before you call the build done, verify the map against the checks that separate a living artifact from a slide.

To keep it current, run the refresh loop on the calendar you set and re-run the recall test each quarter against a fresh benchmark batch, because a map that covered the thesis last quarter can drift as the sector grows. When a founder-status field decays past your tolerance between refreshes, that is the signal to shorten the cadence for that segment, not to clean once and hope. The map earns its keep on the Tuesday you open it, filter to the top-ranked rows inside the boundary, and know exactly who to reach first and why.