Original research and operator guide · 7 min read
A YC directory row is not a prospect until it passes the identity gate
A reproducible nine-cohort YC snapshot shows when a listed website stays on the same host, moves within the same domain, or needs a manual identity check after a cross-domain redirect.
Cofounder of Lazyweb (a16z Speedrun). Previously grew subscriptions and built native ads at Duolingo. Writes First1000 (90,000+ subscribers).
Published · Updated
A directory row is not a prospect. It is a lead to reconcile.
That sounds fussy until the website in the row quietly redirects somewhere else. Then a founder can end up comparing the wrong product, sending outreach to a stale company name, or treating a rebrand, acquisition, product split, and dead link as the same thing.
I built a small status gate from the current public YC directory. The point is not to decide which startups are good. The point is to decide which rows are safe to carry into a prospect list or peer set without hiding uncertainty.
The finding
The saved snapshot contains 1,586 current YC directory rows across nine cohorts, from Summer 2024 through Summer 2026. A guarded raw-HTTP observation on September 19, 2026 got a 2xx HTML homepage for 1,515 rows. Seventy-one rows remain unknown for page-level checks: 54 had no HTTP status and 17 returned a non-2xx status.
Among the 1,515 observed pages:
| Page relationship | Rows | Share of observed pages | Gate status |
|---|---|---|---|
| Listed and final host match | 1,476 | 97.4% | Continue to source review |
| Host changes, same registrable domain | 5 | 0.3% | Continue, record the subdomain |
| Cross-registrable-domain redirect | 34 | 2.2% | Manual identity check |
The 39 host changes are the useful part. A host change is not automatically a conflict. fifthdoor.com ending on play.fifthdoor.com is a different host but the same registrable domain. A cross-domain change is a stronger reason to pause because the old directory row and the current page no longer share the same domain boundary.
I then made one bounded primary check on those 39 rows. The current page title, meta description, and first eight h1/h2 headings were fetched on September 21, with response bodies discarded after hashing. Twenty-nine of the 34 cross-domain rows still contained a non-generic token from the YC row name in that page material. Five did not.
Those five rows form a manual queue. The study does not claim they are acquisitions.
- Hey Revia lists
heyrevia.ai; the saved redirect ended at OpenLoop Health, whose title names OpenLoop Health. - Midship ended at Optro, whose title says “AI-Powered GRC Software.”
- Storia AI ended at a Substack page, which does not identify the page as Storia in its title.
- Maive ended at Ladder.
- Permitify ended at Pangea, an AI meeting-notes product for the built world.
Investigate the row before using it. Preserve it until reconciled instead of silently replacing the old name or throwing the row away and pretending the denominator was clean.
There are also legitimate-looking domain changes that show why a URL heuristic should stay a gate rather than a verdict. Wardstone’s YC row listed wardstone.us and the observed page was wardstone.space; the current title still says Wardstone. Kontigo’s row moved from kontigo.lat to kontigo.com, whose title still says Kontigo. YouShift’s row moved from you-shift.com to youshift.com, with the current title still naming YouShift.
Those pages provide a same-brand signal while leaving ownership, a closed acquisition, and exact current operation unverified.
The gate I would use
1. Freeze the row identity.
Keep the YC company ID, exact row URL, cohort, row name, listed website, and the snapshot date together. Do not join companies by a fuzzy name. If the directory says “Midship” and the current page says “Optro,” that is a reconciliation task, not permission to rewrite the row.
2. Separate page reachability from identity.
A 200 HTML response answers one narrow question: did a public page respond to the guarded request? It does not answer whether the company is active, whether the page is the right product, or whether anyone buys it. The 71 unknown rows stay unknown. They are not failures and should not be silently removed from the denominator.
3. Compare hosts before reading the copy.
An exact host match is the low-friction state. A subdomain on the same registrable domain is a small continuity change worth recording. A cross-registrable-domain redirect gets a manual_identity_check status. This is the point where a prospecting system should stop auto-enriching the row.
4. Read the current first-party identity signal.
Check the page title and primary heading, then compare the product boundary with the YC one-liner. A matching name is a useful signal. A different name is a reason to look for a first-party announcement, current company page, or other public explanation. A missing token in raw HTML can also reflect JavaScript-only content or an extraction limit, so treat it as escalation rather than an identity conflict. Do not infer an acquisition or rebrand from a redirect alone.
5. Preserve the result as a status, not a cleanup.
I would use three labels in a working list:
- Green: page observed and host identity is continuous enough to proceed to ordinary source review.
- Yellow: cross-domain redirect or a current page that names a different surface; hold for manual reconciliation.
- Gray: no usable page observation; keep the row, but do not pretend the website was checked.
The labels describe what the researcher can safely do next. They do not grade the company.
Why the cohort table matters
The cross-domain queue is not evenly distributed in this snapshot. Summer 2024 contributes 11 rows, Winter 2025 contributes 8, Fall 2025 contributes 4, Summer 2025 contributes 4, Spring 2026 contributes 3, Spring 2025 contributes 2, Fall 2024 contributes 1, Summer 2026 contributes 1, and Winter 2026 contributes none. The same table also shows the unknown rows, so a large “clean” cohort does not accidentally mean “every company checked.”
| Cohort | Rows | 2xx HTML | Unknown | Cross-domain queue |
|---|---|---|---|---|
| Summer 2024 | 248 | 229 | 19 | 11 |
| Fall 2024 | 94 | 85 | 9 | 1 |
| Winter 2025 | 165 | 156 | 9 | 8 |
| Spring 2025 | 143 | 138 | 5 | 2 |
| Summer 2025 | 166 | 159 | 7 | 4 |
| Fall 2025 | 146 | 138 | 8 | 4 |
| Winter 2026 | 199 | 191 | 8 | 0 |
| Spring 2026 | 193 | 188 | 5 | 3 |
| Summer 2026 | 232 | 231 | 1 | 1 |
This remains a current-directory observation rather than a market-prevalence estimate. Cohorts differ in age, public-site state, and how their companies describe themselves. The table tells you where the manual queue was in this edition of the snapshot; it does not explain why.
The operator decision
If I were building a founder prospect list from an accelerator directory, I would never pass the raw export straight into outreach. I would run the gate first, carry the gray and yellow states visibly, and only then add buyer fit, product boundaries, pricing, or research notes.
That order matters. A polished profile of the wrong company is worse than an incomplete profile that says “identity unresolved.” The first one creates false confidence and contaminates every later count. The second one gives the operator a short queue to resolve.
The full recount is standard-library code over the public compact URL/status dataset linked in the methods receipt. It checks all 1,586 rows, keeps the nine cohort denominators, and treats missing page observations as unknown. The 39 flagged pages were the only pages re-fetched for the bounded primary check. The original roster and raw page bodies are not redistributed; no login, form, product execution, or private data was involved.
Before asking what a directory company sells, verify that the public page still belongs in the same row.
Methods and limits
The YC snapshot was captured September 18, 2026 from the rendered public directory’s batch-filtered responses. It covers Summer 2024, Fall 2024, Winter 2025, Spring 2025, Summer 2025, Fall 2025, Winter 2026, Spring 2026, and Summer 2026. The homepage inventory was observed September 19 with credential-free guarded GETs, robots checks, a two-million-byte body limit, and no reference-page fetching. The 39 saved host-change rows received a second credential-free homepage and robots.txt check on September 21; only public title/headings/meta description and a response hash were retained.
“Same registrable domain” uses a small explicit suffix rule in the recount script. “Name signal” means a non-generic token from the YC row name appeared in the verified page title, meta description, or first eight headings. Neither rule proves legal identity. No claim is made about employment, demographics, customer adoption, product execution, or causal market change. Page content and redirects can change after these observation dates.
Seven of the 39 primary checks were repeated later on September 21 to clarify robots-policy evidence, including one redirected path. All seven retained the same title, headings and meta description, and the 29/5 cross-domain name-signal partition was unchanged. The downloadable primary packet records both observation periods.
About the author
Cofounder of Lazyweb (a16z Speedrun). Previously grew subscriptions and built native ads at Duolingo. Writes First1000 (90,000+ subscribers).
I cofound Lazyweb and publish Mudpie. This is an owner-written publication, not an independent testing organization. Research notes distinguish observations, sourced reporting and editorial judgment.
First1000 ↗ · X ↗