Learn · Founder story
Why We Built Etymolt — A Founder's USPTO Office Action Story
The envelope arrived on a Tuesday in November 2025. Week 16 of the previous brand. I could see the USPTO return address through the window and knew, in the way you know things at 8:17 in the morning before your coffee has done anything, exactly what was inside. I sat down on the front step and opened it there.
It was an Office Action under §2(d) — likelihood of confusion with a senior mark in the same Nice class. The senior mark had been registered five years earlier, by an operator I'd never heard of, in a sub-vertical I didn't consider competitive with what I was building. The USPTO examiner did not consider those distinctions relevant. Same class. Same phonetic cluster. Refusal.
We had been live for eleven weeks. The product had a domain, a logo, a Series A pitch deck, three customer-facing announcements, and a media-relations push scheduled for the following Monday. I cancelled the Monday push from the front step. I called my attorney. The first conversation, which lasted nine minutes, cost $437. The full legal response — including the office-action reply, the eventual decision to rebrand rather than fight, and the brand-identity redo — came in at $40,217. Eleven weeks of go-to-market momentum vanished. The Series A timing slid by a quarter.
I had run a USPTO check before launch. I had searched the wrong Nice class. The mark that killed us was in Class 9. I had checked Class 35.
That was the first realization, and it stung. Trademark search is presented to founders as a 30-minute task you do once. It's actually a four-hour task, if you do it right, and the four hours are the difference between checking Class 35 (which is what every founder defaults to) and checking the actual Nice class that governs your product. Most founders don't have four hours on a Tuesday in week 14 of a launch. Most founders do it the way I did — 30 minutes, the wrong class, a clean read, and a $40K bill four months later.
The second realization: the LLM had picked the name
When I went back through the naming history — I keep notes, badly, but I keep them — the name that killed us came out of a Claude session in week 4. I'd asked for fifty candidates against a brief. Claude returned fifty. I narrowed to six finalists. Five of those six came from the Claude list.
None of the six had been checked against any register at the point Claude returned them. Claude doesn't check; that's not what an LLM does. Claude produces plausible-sounding names against a brief and presents them with the same confidence it uses for factual questions. There's no warning label. There's no flagging of likely conflicts. The output is a list of strings; the verification is the founder's problem.
I'd done my Level 1 TESS check on the finalist I'd chosen. I'd done it in the wrong class. I'd felt the warm glow of due diligence completed. And I had been wrong, and I'd eaten the $40K to find out.
The third realization: trademark wasn't the only failure mode
When we picked the replacement brand, I ran the trademark check carefully. The right Nice class. TESS, TTAB, and the Madrid register. The candidate cleared. The relief lasted three days, until I tried to register the social handles.
@[brand] on GitHub had been held since 2009 by an account that had posted twice and was dormant. @[brand] on X had been held since 2007 by a different dormant account. Both were unretrievable. I petitioned GitHub Support. The petition was rejected in twelve days. We shipped under @[brand]hq on every platform and accepted the dilution forever.
Two months later, a friend launching a consumer product asked me to look at his shortlist. One of the candidates was a clean English word that, in Brazilian Portuguese, was a regional slur. He'd planned a Latin American launch in year two. He'd also been working with Claude, which had returned the candidate as a fluent suggestion against his brief. Neither of us had thought to check the cultural axis. The check exists — Wiktionary, Urban Dictionary, indexed Reddit threads — but it takes 90 minutes per market, and you have to know which markets matter. He pivoted.
That was the moment the system problem became visible. Trademark wasn't the only failure mode. Handle was a failure mode. Cultural was a failure mode. Sound symbolism — whether the name even feels like the brand it's naming — was a failure mode that nobody was checking. Pronunciation in noisy environments, in accents you don't have, was a failure mode that nobody was checking. The other four failure modes killed names at rates that, taken together, exceeded the trademark rate.
The five-axis vision
Every naming tool I could find checked one axis (trademark) and called the candidate cleared. Some checked zero — they returned candidates the way Claude does, without verification. None of them checked all five axes in a single call. None of them traced the flags back to record numbers. None of them returned a verdict in three seconds.
The product spec wrote itself: a single API call, POST /v1/verify, that runs every axis in parallel and returns a verdict with every flag traced to a record number. Sub-three-second p95. Citation-grade output. Five free per IP so any founder could check their candidate on a Friday night without giving up their email. The full methodology public, so the work is auditable and the calibration is honest. The legal posture explicit: Bureau Model. Clearance signal, not legal advice. We surface evidence; we're not a law firm. We complement attorneys; we don't replace them.
Building Etymolt
The Office Action arrived in November 2025. The first commit on the Etymolt repository was December 7, 2025. The MVP — single API call, two axes (USPTO +.com), a working verdict — was six weeks later, on January 18, 2026. The first paying customer, a YC W26 founder running through his own naming-storming weekend, verified twelve candidates in 48 hours. He picked the one that scored 91/100 across all five axes and shipped under it.
The methodology — five axes, twenty markets, twelve accents, every flag citation — took longer than the API. The cultural corpus took six weeks of cleanup. The pronunciation pipeline (TTS→Whisper round-trip) took four weeks of accent calibration. The sound-symbolism scoring engine, calibrated against the Kawahara & Shinohara corpus, took eight weeks. We shipped the public methodology page in May 2026, the same week as the production API.
Where we are now: Fremont, CA, May 2026, public launch. The homepage runs a live verifier. The methodology is citable and recalibrated weekly. The first five verifies per IP are free. Every verdict gets a permalink at etymolt.com/v/[id]. The Bureau Model posture is the legal floor: we surface signals; we're not a law firm. Attorneys partner with us — we save them prep time, they keep the opinion letter and the regulated work. The founder bio is at /about.
If you're reading this on a Friday night
If you're a founder reading this on a Friday night, naming-storming for Monday's pitch — verify before you commit. Five free, no signup. The verdict comes back in under three seconds. Every flag traces to a record number. If the verdict is PROCEED, you have evidence to show investors. If the verdict is STRATEGIC, you have a punch list of what to clean up. If the verdict is ABANDON, you find out before you pay for the domain.
The week-16 envelope cost me $40,217 and a quarter of momentum. The five-axis verdict would have cost me three seconds and zero dollars. I wish I'd had it. Etymolt is that wish, productized.
Related reading
- How to Check If a Brand Name Is Trademarked — the four-level ladder for the trademark axis. Free, freemium, API, attorney.
- The Five Ways a Brand Name Dies — the frequency distribution of 400 names, axis by axis.
- About Etymolt — the company, the posture, the operators behind it.
Take the next step
Verify your candidate before you commit.
Five free clearances per IP. No signup. Every flag traces to a record number. Three seconds, end to end.
Verify a name