# Sourcing Companies Before the Round Is Announced

*You will be able to find pre-seed and seed companies weeks before they surface in any funding database, and reach the founder before hire number ten.*

- Canonical URL: https://www.refolk.ai/guides/sourcing-before-round-announced
- Pillar: Investing and deal sourcing
- Format: Playbook
- Published: 2026-08-15
- Last reviewed: 2026-08-15
- Reading time: 15 min

This is a playbook for finding pre-seed and seed companies weeks before they appear in any funding database. It is written for early-stage investors, platform and talent partners at funds, and angels who want proprietary deal flow instead of the shared flow a warm intro produces. Follow it start to finish and you will have a monitoring system that surfaces founders while they are still in stealth, plus a scoring bar that keeps the noise out.

The premise is simple and it is the whole reason this guide exists: funding databases are late by construction. By the time a company shows up in one, the founder has already chosen a lead investor and filled the first ten hires. The job is not to read the news faster. It is to detect the traces a company leaves before it makes any news at all.

## Why funding databases are structurally late

Funding databases are late because the legal event they key on fires after the round is already done. In the US, the primary trigger is Form D, which must be filed within 15 calendar days after the first sale of securities. The filing then becomes a public EDGAR record that investors and press read during diligence. So even in the best case, the database reflects a decision that closed at least two weeks earlier.

It gets worse for anyone relying on that data. Startups may lawfully skip Form D to stay in stealth, using less-public state notices such as California's 25102(f), which are more likely to fly under the radar. So a pipeline built on funding events is not merely delayed. It is blind to a real slice of the market on purpose.

**15 - Calendar days after the first sale of securities before Form D is due**

The primary US funding-event signal arrives after the round has already closed, and some founders never file it at all.

The insight to internalize is that this is not a lag you can wait out. It is a floor. No tool keyed to funding events can beat it, which is why the load-bearing signals in this playbook are people-level and registry-level rather than money-level.

> **Rule:** Never build the pipeline on EDGAR alone
>
> Form D arrives 15 days after the first sale, and stealth companies may never file it. Treat funding filings as confirmation of a signal you already had, never as the leading edge.

## Which signals reveal a company before the money moves

The signals that surface a company early are people-level and registry-level. A profile updated to "stealth" or to a vague title like "building something new" is the most explicit signal that someone is starting a company, and everything else corroborates or contradicts it.

Even companies operating in strict stealth leave observable traces, because certain things are not optional. Legal formation happens in public. A 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. Technical founders are usually writing code, and that activity is partially observable on the public GitHub graph. Infrastructure requires domain names. Deep tech ventures file patents. None of these are things a founder can suppress without also failing to build the company.

Hiring compounds all of it. When someone posts a role like "Founding Engineer," "First Designer," or "Early GTM Hire," they are assembling a team, which means a founder is active and growing. And a quieter version of the same signal: when a senior engineer leaves a top AI lab with no publicly announced next role, that departure is often the first indication a stealth company is forming.

#### The signal layers, outermost first

1. **Self-declared** - Title changed to "stealth" or "building something new"
2. **Hiring** - "Founding Engineer / First Designer / Early GTM" posts at tiny companies
3. **People movement** - Senior departure from a target company with no next role listed
4. **Registry and code** - New incorporation, domain registration, GitHub activity, patent filing

*You read from the most explicit self-declared signal down to the quiet infrastructure traces that a founder cannot suppress.*

Here is what each signal proves and how it lies. A stealth title proves intent to declare, but consultants and between-jobs operators use the same label. A departure with no next role suggests a build, but it can be a sabbatical, a layoff, or an unannounced move to another big company. A founding-engineer post proves a team is forming, unless a recruiting agency posted it for a client that is already well-funded. A registered domain proves someone reserved a name, unless it is parked, squatted, or hidden behind privacy WHOIS. Every signal has a false positive, which is why no single one is enough.

## How much lead time you get, and what the real deadline is

The window is weeks to months, and it closes at hiring, not at announcement. By the time a stealth startup surfaces through a launch or a funding event, the founder has already chosen their lead investor and locked in their first ten hires. The teams that won those allocations reached the founder weeks or months earlier.

That reframes the entire job. The deadline is not the announcement date. It is hire number ten. You are racing the founding roster, not the press release. This is why hiring signals matter so much: they are both a detection tool and a countdown clock. When a company you flagged pre-round starts adding headcount fast, that confirms your early signal was real, because startups raising Seed through Series B add 30 to 60% more headcount within six months of the raise. If a company you flagged before the round begins that surge, it validates the flag and dates exactly how early you were.

> The deadline is not the announcement. It is hire number ten, and the roster fills faster than any database updates.

## Sizing the pool: where the stealth founders are

The stealth-founder population is smaller and more concentrated than the noise suggests, and geography collapses it faster than sector does. In Refolk's index of professional profiles, roughly 4,891 US-based founders carry a stealth signal, and the top employer label is literally "Stealth Startup," concentrated in San Francisco and New York. The UK pool is far smaller.

| Country | Founders with stealth signal | Top hub | Share vs US |
|---|---|---|---|
| United States | 4,891 | SF / New York | 1.00 |
| United Kingdom | 1,088 | London | 0.22 |

The derived ratio is the useful part: the US pool is about 4.5 times the UK pool, and both concentrate in one or two cities. A two-city US filter removes most of the population before you spend any analyst time. That is the cheapest cut you can make.

**4.5x - How much larger the US stealth-founder pool is than the UK's, in Refolk's index**

Geography is the sharpest filter available; a two-city cut removes most of the population before any manual review.

Sector, by contrast, barely differentiates anymore. AI took 41.7% of all seed capital in 2025, and recent YC batches ran 77 to 80% AI. "AI startup" is the default, not an edge. What differentiates is provenance: which lab or company the founder left. "Ex-AI-lab engineer in SF" is a filter. "AI startup" is a description of nearly everything.

This is exactly the kind of population question where a keyword-graph index earns its place. Instead of scraping and reconciling profiles yourself, you can ask [Refolk](/) for a role, a signal, and a place in plain English and get the resolved people back.

I ran this search: `Engineers who left OpenAI, Anthropic, or Google DeepMind in the last six months with no current employer listed` - [see the full result list](https://www.refolk.ai/s/z77kzgyzrm).

*Returns named engineers whose most recent public role at a top AI lab has ended with no listed successor, the classic pre-stealth departure signal.*

## The market context that sets your monitoring width

Seed is a breadth game, so your monitoring has to be wide even though most of what you track will never progress. Seed is about 40% of new funding events but only 9.4% of the cash, and only 15.4% of the 2022 seed cohort raised a Series A within two years. You are watching a large, shallow field where the base rate of progression is low.

| Metric | Value | Period |
|---|---|---|
| US pre-seed cash raised | $822M | Q2 2025 |
| Median seed pre-money | $16M | Q3 2025 |
| Seed share of deal count / of cash | 40% / 9.4% | Q3 2025 |
| Median seed post-money | $24M | Q4 2025 |

The takeaway from these benchmarks is not the dollar figures on their own. It is the shape: many small events, few of which grow. Median seed post-money hit a record $24M in Q4 2025, up from $18M a year earlier, and median pre-money hit a record $16M in Q3, so valuations are climbing even as progression rates fall. That combination is why sourcing width matters more than sourcing precision at this stage. You cannot pick your way to good returns from a narrow funnel, which is the point behind the a16z line that success in venture is "10% about picking, and 90% about sourcing the right deals." Roughly 45% of new VC deals still originate from existing networks, and meeting-to-investment conversion runs below 1%, so a wide, monitored top of funnel is not optional.

## The procedure

This is the end-to-end method, in order, with who does each stage and what done looks like. Run steps one through four to stand up the system, then run five through eight as a weekly and per-batch cadence.

#### Sourcing before the round is announced

1. **Define the thesis and filter** - Write a one-page target: sector, geography, and founder provenance, such as ex-AI-lab engineers in San Francisco. Done means a filter specific enough that sector alone does not carry it, since AI is the crowded default. Analyst or partner, about half a day.
2. **Wire up people-level monitors** - Build saved searches for title changes to "Stealth" or "building something new" and for departures from target companies with no listed next role. Done means both monitors are running on real profiles. Platform or data associate, 1 to 2 days.
3. **Add registry and code monitors** - Add watchlists for new incorporations, newly registered domains, patents, and GitHub activity, since incorporation is public within days of filing. Done means each is live. Data associate, 1 to 2 days.
4. **Add hiring monitors** - Set alerts on "Founding Engineer," "First Designer," and "Early GTM" posts at sub-five-person companies. Done means alerts fire, while you accept the best founding hires are placed before the role is posted. Associate, ongoing.
5. **Score and dedupe weekly** - Require two or more independent corroborating signals per candidate before outreach, such as title change plus registered domain, or departure plus a co-hire cluster. Done means every live candidate clears the bar and duplicates are merged. Associate, 2 to 3 hours weekly.
6. **Reach the founder before hire ten** - Open with a first meeting that offers value, not a term sheet: introductions, customer referrals, deck feedback. Done means a real conversation started before the roster fills. Partner, over weeks.
7. **Log everything in a CRM** - Record every company, signal, source, and pass rationale from day one. Done means the pipeline is fully reconstructable and the database is becoming a competitive asset. Whole team, from day one.
8. **Sweep accelerator batches as a backstop** - Capture accelerator batch lists before demo day as a completeness check. Done means the batch is logged and treated as competitive, not proprietary, flow. Associate, per batch.

One honest note on ordering: sources disagree. Signal-detection vendors put monitoring first, while traditional VC guides put relationship-building and accelerators first. This playbook leads with monitoring because the deadline is hire ten, and monitoring is what buys you the weeks. But the relationship step is where deals are actually won, so do not treat step six as an afterthought.

#### From raw signal to a founder meeting

1. **Detect** - People, registry, and code monitors surface a raw candidate
2. **Corroborate** - Require a second independent signal before scoring
3. **Score and dedupe** - Rank weekly, merge duplicates, drop single-signal noise
4. **Reach out** - Partner opens with value before hire ten
5. **Log** - Every company, signal, and pass rationale into the CRM

*Each stage narrows the population and adds corroboration, so only twice-confirmed candidates reach a partner's time.*

## How this goes wrong: the false positives

Most of the value in this playbook is knowing which signals lie and how to catch them. Every early signal has a plausible false positive, and acting on a single signal is the fastest way to waste a partner's calendar. Below are the failure modes to screen for before anything reaches outreach.

| Signal | Looks like a startup | But could be | Confirm with |
|---|---|---|---|
| "Stealth" title | New company forming | Fractional exec, consultant, between jobs | Co-hire cluster, domain, or incorporation |
| Departure, no next role | Founder building quietly | Sabbatical, layoff, unannounced big-co move | GitHub commits, domain, or direct outreach |
| "Founding Engineer" post | Team assembling | Agency posting for a funded client | Company headcount and founder identity |
| New domain registration | Infrastructure being set up | Parked, squatted, or privacy-hidden domain | A matching team signal |

Two more failures deserve their own callouts because they cost the most time.

The first is treating Form D as a leading signal. It is not. It arrives 15 days after the first sale, the round is already done, and some stealth companies never file at all. Building the pipeline on EDGAR is building it on the one signal guaranteed to be late.

The second is a query-construction error that will silently gut your results. A literal, stacked keyword phrase under-counts by orders of magnitude.

> **Watch out:** Exact-phrase over-filtering hides the population
>
> A literal query for "stealth artificial intelligence" in Refolk's index returned only about 6 people, versus 4,891 for "stealth" alone. Filter on skills and title fields, not stacked literal phrases, or you will under-count by orders of magnitude and conclude the pool is empty when it is not.

The accelerator trap is the last one. YC/accelerator lists get mistaken for proprietary deal flow, but most companies exit stealth right before demo day, so by the time the batch list is complete the deal is competitive. A newsletter reconstructed recent batches before demo day, and the numbers show how large and public this flow already is.

| Batch | Companies | Founders |
|---|---|---|
| YC F24 | ~100 | 202 |
| YC W25 | ~172 | 375 |
| YC X25 (Spring) | ~145 | 305 |

Use these as a completeness backstop against your own monitoring, never as a source of edge. If a company only shows up when the batch list drops, you were late.

> **Tip:** Score before you reach out, not after
>
> Require two independent corroborating signals per candidate every week: title plus domain, or departure plus a co-hire cluster. This one rule eliminates most of the false positives above before a partner ever spends time on them.

## Before you call the system live

Run this check before you trust the monitoring to feed real outreach. Each item is a specific state to verify, not a topic to think about.

#### System readiness check

- [ ] The thesis is written on one page with sector, geography, and founder provenance, and provenance is doing the differentiating work rather than sector.
- [ ] A two-city or single-region geographic filter is applied, since geography collapses the pool faster than sector.
- [ ] People-level monitors fire on both stealth-title changes and role-less departures from named target companies.
- [ ] Registry, domain, patent, and GitHub monitors are live and returning at least one real hit each.
- [ ] Hiring alerts cover "Founding Engineer," "First Designer," and "Early GTM" at sub-five-person companies.
- [ ] The scoring rule enforces two-plus independent corroborating signals before any candidate reaches outreach.
- [ ] Queries use skill and title fields, not stacked literal phrases, so the population is not silently under-counted.
- [ ] The CRM captures every company, signal, source, and pass rationale from day one.

## Keeping the system current

A sourcing system decays because the signals move. To keep it working, re-check the parts that are time-sensitive rather than the numbers themselves. The mechanisms are what matter.

Re-derive your geographic and provenance filters as the population shifts. The stealth-founder pool concentration is stable enough to plan around, but the specific labs and companies worth watching change as talent moves. When a new lab spins out engineers, add it to the departure monitor. When a hub cools, reweight your two-city filter.

Watch your own confirmation loop. The cleanest validation that your early signal was real is the post-raise hiring surge: companies add 30 to 60% headcount within six months of raising. Track which flagged companies enter that surge and how many weeks before the round you first saw them. That number is your actual lead time, and it is the only honest scorecard for whether the system is beating the databases.

Finally, keep the outreach honest. Proprietary deal flow is widely claimed and rarely achieved, because most network-based sourcing produces well-positioned shared flow, not genuinely proprietary flow. The only way to earn the real version is to reach founders before any process starts and offer them genuine value: introductions, customer referrals, feedback on their decks. Those relationships compound over two to three years into a referral network that generates proprietary deal flow. The monitoring buys you the weeks. What you do with them is what builds the moat.

## Frequently asked questions

### How much lead time can I actually get before a round is announced?

Weeks to months, and the window closes at hiring rather than at announcement. By the time a stealth company surfaces through a launch or funding event, the founder has usually already chosen a lead investor and locked in the first ten hires. The investors who won the allocation reached the founder weeks or months earlier, off people-level and registry signals rather than funding data.

### Why are funding databases always late?

Because the primary US legal trigger is a lagging one. Form D must be filed within 15 calendar days after the first sale of securities, and it becomes a public EDGAR record only then, so the round is already done before the filing appears. Some stealth founders skip Form D lawfully using less-public state notices, so a pipeline keyed to funding events is structurally weeks behind and sometimes blind entirely.

### Is a LinkedIn 'stealth' title enough to act on?

No. A single stealth signal is one of the most common false positives in this work. Fractional executives, consultants, and between-jobs operators all self-label stealth. Require a second independent signal before outreach: a co-hire cluster, a registered domain, or an incorporation. The scoring rule of two-plus corroborating signals exists specifically to filter this noise out.

### Are accelerator batch lists proprietary deal flow?

No, treat them as a competitive backstop. Most companies exit stealth right before demo day, so by the time a batch list is complete the deal is already visible to every fund watching. A newsletter reconstructed YC W25 at about 172 companies and 375 founders before demo day, which is useful for completeness but not an edge. Real proprietary flow comes from reaching founders earlier through people-level signals.

### Does filtering by sector help narrow the pool?

Barely, anymore. AI took 41.7% of seed capital in 2025 and made up 77 to 80% of recent YC batches, so 'AI startup' is the default rather than an edge. Geography collapses noise far faster: in Refolk's index the US stealth-founder pool is about 4.5 times the UK's and concentrates in two cities. The real differentiator is founder provenance, meaning which lab or company the person left.

---

*From the Refolk guide library. I revise these guides rather than replacing them, so the current version is always at https://www.refolk.ai/guides/sourcing-before-round-announced*