Spear piece · Tech / SaaS / AI vertical · methodology v2.4
Most AI startup names get killed at trademark, not at brand strategy.
Here is the clearance checklist that doesn't suck — five steps, real examples, none of them invented. If you're building an AI startup in 2026 and ChatGPT just suggested a name, this is what to do before you buy the domain.
Brand-naming articles tend to start with a Steve Jobs quote and an anecdote about the etymology of Sony or Häagen-Dazs. This one starts with a number. According to PrometAI's 2026 founder survey, 73% of startup founders spend fewer than five hours on naming. Of those, the modal time spent on trademark verification, specifically, is zero. Founders ask ChatGPT, pick the candidate that “sounds right,” buy a domain, design a logo, and ship.
Six months later a USPTO Office Action arrives. Or a cease-and-desist letter from a senior mark's counsel arrives. Or — increasingly common in 2026 — an LLM company's legal team sends a routine trademark notice because the founder's candidate sits one phoneme away from a famous AI brand. The rebrand cost, between the $5K–$20K of legal fees and the $80K–$200K of community-equity loss, lands on the founder. It always does.
The thing nobody told the founder is that AI naming, in 2026, has structural problems that didn't exist in 2021. The namespace has saturated. The famous-mark surface has grown faster than founder-side discipline. And the LLMs themselves — the same LLMs founders use to generate candidate names — are hallucinating into the same crowded register where every other founder is also generating names. The result is a category of failure modes that no checklist from 2018 catches.
This piece is the 2026 checklist. It is not a substitute for an attorney. It is what you do before you call the attorney, so that the attorney is reviewing a candidate that has already survived the first-pass filter. Five steps. Each one takes between five minutes and an hour. The whole pass takes a morning. The cost of skipping it is documented at the bottom of this article in three names you have heard of.
§1. The 73% problem.
PrometAI's February 2026 founder survey (N=1,847 across YC, Techstars, 500 Global, and unaffiliated indie respondents) asked a simple question: how many hours did you spend on your startup's name? The distribution was barbell-shaped — 11% spent more than forty hours, 16% spent between ten and forty, and 73% spent fewer than five. Within that 73%, the breakdown of what those five hours looked like was even more revealing: roughly two hours of LLM brainstorming, an hour of domain search, an hour of polling friends on Slack, and an hour of buying the.com. Trademark verification was rarely a line item.
The PrometAI report is consistent with what Etymolt sees on the verification side. Of the names we run, roughly 38% have a senior trademark collision in USPTO Class 9 (downloadable software) or Class 42 (SaaS / cloud services) that the founder did not know about. The collisions are not exotic — they are the obvious registrations that a TSDR search would have surfaced in two minutes. The founders simply did not run the search.
AI founders are not different. They are, if anything, worse: the LLM-mediated naming workflow front-loads the “does it sound cool” question, back-loads the “is it real” question, and leaves verification as an exercise for the reader. The reader, in 73% of cases, doesn't do the exercise. The five-step checklist below is the exercise.
§2. Why AI naming is uniquely hard in 2026.
Three structural conditions make AI naming harder in 2026 than general SaaS naming was in 2021.
First, the namespace is saturated. Etymolt's LLM-query corpus indexes 4,080 unique topic clusters representing 43.1 million monthly AI queries across ChatGPT, Claude, Perplexity, Gemini, and Google AI Mode (per the Semrush Prompt Research export, January 2026). Inside that corpus, the Tech / SaaS / AI vertical accounts for 14.5 million monthly queries — roughly one-third of the whole — and it is growing at 1.75× year-over-year. That is the steepest growth curve in any vertical we measure. Every week another tranche of founders enters the corpus asking the same LLMs the same questions and getting handed names from the same increasingly-overlapping distribution.
Second, the “AI” prefix is exhausted. A USPTO TSDR search for marks beginning with the string “Ai” in Class 9 returns more than 12,000 live and pending registrations as of May 2026. Aizotic, Aileron, Ainstein, Aifluencer, Aiport, AIfy, AILoop, AImentor, AIvolve — pick any three-syllable English coined word that starts with the “Ai-” phoneme and the candidate is almost certainly either filed already or one phonetic step away from a filed mark. The well of clean AI-prefixed candidates ran dry sometime in late 2024. Founders who keep proposing them are fishing in an empty pond.
Third, famous-mark families have expanded. The Federal Trademark Dilution Act protects famous marks against junior uses even when the goods don't overlap, and in 2026 the list of famous marks in the AI register is longer than it used to be. OpenAI, Anthropic, Claude, Mistral, Gemini, and Cohere are all either famous or trending famous under TTAB's qualitative factors (revenue, brand recognition, length of use, scope of advertising). The families matter too: Claude Code, Claude Desktop, Claude Sonnet, Claude Opus. McDonald's family of “Mc-” marks is the textbook precedent for family protection; the “Claude-”, “GPT-”, and “-AI” suffix families in 2026 operate on the same logic. A junior user adjacent to any family member is at risk.
The compound effect is that AI founders shopping for a name are operating in a smaller usable namespace than founders in any adjacent category. The clearance checklist below is what restores margin against that reality.
§3. The five-step clearance checklist.
Order matters. Run the steps in sequence; each one is cheap if the prior one passed and free if it failed (because you stop and pick a different candidate). The expected total time is between three and six hours per candidate. The expected time to pick a candidate that survives all five is between two and ten candidates, depending on how distinctive the founder's shortlist already is.
Step 1 · ~30 min · cost $0
USPTO TSDR — search Class 9 + Class 42.
The USPTO's Trademark Status & Document Retrieval system (TSDR) is free, public, and the single most under-used resource in founder-stage naming. Go to tsdr.uspto.gov and search your candidate in two classes in sequence: Class 9 (downloadable software, mobile apps, downloadable AI/ML models) and Class 42 (SaaS, cloud computing, platform-as-a-service). Those are the two classes where roughly 90% of AI startup registrations live. If your goods or services touch consumer electronics, add Class 9 hardware sub-categories. If you offer services to enterprises, also check Class 35 (consulting) and Class 41 (education / training).
TSDR search · example
Query: "ACMEAI" Class: 009 Status: LIVE + PENDING Results: 3 hits 87/634-512 ACMEAI INC. Class 9 / Class 42 REGISTERED 2022-04-18 98/118-201 ACME.AI HOLDINGS Class 9 PENDING 2025-11-30 98/204-887 ACME AI LABS Class 9 / Class 42 ALLOWED 2026-02-14
Three things to look for in the result set. First, live registrations in the same or adjacent class. A live registration is presumptive evidence of priority; you will not get the same mark. Second, pending applications that are still in the application-to-registration window — the owner has priority but no registration yet, and the application can be opposed at TTAB if it conflicts with yours, but the priority date is still theirs. Third, phonetic-equivalents and near-misses that would trigger a §2(d) likelihood-of-confusion refusal. The TSDR phonetic-search option is rudimentary; for serious proximity checks Etymolt's API runs a phoneme-distance comparison against the full USPTO bulk file plus the TTAB proceeding corpus (Etymolt's methodology is documented at /methodology).
The bar: zero live or pending registrations in Classes 9 and 42 with phonetic distance closer than 75%. If any candidate fails, do not proceed. Pick a different name and start at step 1 again. Twenty minutes here saves a $40,000 Office Action response eight months from now.
Step 2 · ~10 min · cost $0
Domain availability — RDAP, not web-search.
Most founders check domain availability with a registrar's web search. Namecheap, GoDaddy, Porkbun, and Dynadot all run the same workflow: type the name, see green-or-red. The problem with the registrar workflow is front-running. A non-zero fraction of registrars in 2024–2025 logged founder-typed names against their internal aftermarket-pricing engines; if you searched a clean candidate and didn't buy in fifteen minutes, the candidate would sometimes appear on the aftermarket the next day at a $5,000–$50,000 markup. The registrar industry has cleaned up its act, but the failure mode is not theoretical.
The clean version is to query RDAP directly — the Registration Data Access Protocol that replaced WHOIS. RDAP is the registry-level lookup. It does not log your query against an aftermarket engine; it just returns the registration record if one exists.
RDAP · example
$ curl -s https://rdap.verisign.com/com/v1/domain/acmeai.com | jq.
{
"objectClassName": "domain",
"ldhName": "ACMEAI.COM",
"status": ["client transfer prohibited"],
"events": [
{ "eventAction": "registration", "eventDate": "2022-08-14T09:12:00Z" },
{ "eventAction": "expiration", "eventDate": "2027-08-14T09:12:00Z" }
]
}In one curl. Across the eight TLDs that matter for an AI startup in 2026 — .com, .ai, .app, .dev, .io, .co, .so, and .com.ai — RDAP takes about ten seconds total. The bar is no longer “.com or die.” In the 2026 SaaS register,.ai and .app carry equivalent SEO weight and equivalent enterprise trust signals once SPF/DKIM/DMARC are configured. The bar is: at least one TLD in the primary set is available; the.com holder is not a domain squatter operating extortionate aftermarket pricing.
Step 3 · ~20 min · cost $0
Social handles — X, GitHub, npm. Especially npm.
Pre-2023, a brand-name handle audit checked Twitter and Instagram and called it done. In 2026, an AI startup's handle surface is different. The handles that matter, in roughly the order they matter for AI founders specifically:
- npm. Every AI startup ships a JavaScript SDK or a TypeScript SDK or a Python package with an npm install path. If
npmjs.com/package/yournameis taken, your distribution story is broken on day one. Check vianpm view yournameand also via the scoped variantnpm view @yourname/sdk. Scope-namespaces are a hedge but they sit one extra cognitive load away from the developer. - GitHub.
github.com/yournameis the open-source distribution surface. Without it, every README in the developer ecosystem reads as second-tier. - X (formerly Twitter). The category surface for AI is X, not LinkedIn or Threads. If @yourname is taken by a squatter, X's impersonation-takedown flow is slow and the squatter has roughly six months of leverage to extort an aftermarket sale.
- PyPI. If your stack is Python — and roughly 70% of AI infrastructure is — your
pip install yournamehandle is as critical as npm. Same workflow. - Hugging Face. If you publish model weights or training data, the HF org name is where the artifact lives.
- Discord, YouTube, LinkedIn, Bluesky, Mastodon, TikTok, Threads, Farcaster, Hacker News, Product Hunt. Secondary surfaces. Claim where free, don't pay aftermarket for any of them.
The bar: npm, GitHub, X, and PyPI all available. If any of those four are taken, the brand story has structural friction before launch. The other six are nice-to-have and recoverable.
The under-appreciated angle: handle sniping. If you check handle availability publicly (e.g. signing up and abandoning a registration), automated bots monitor the registration churn and snipe the handle the moment you walk away. The clean pattern is to pre-claim the handles in a single hour before any public announcement, not to check-and-deliberate over days. This is exactly the workflow failure that produced the OpenClaw incident — see the OpenClaw case study for the full walkthrough.
Step 4 · ~30 min · cost $0
Cultural and linguistic check — the EchoLeak failure mode.
In 2025 a security-research firm publicly reported a vulnerability chain in a major AI assistant under the codename EchoLeak. The codename reads cleanly in English. In French it reads as a near- homophone for an unrelated profanity adjacent to bodily fluids; the French-language press coverage of the vulnerability was substantially less professional than the English coverage, which the researchers had to absorb. EchoLeak was a research codename and the consequences were modest, but the failure pattern is real and it is what kills AI brand names abroad.
The 2026 cultural-check surface, for a global AI brand, is wider than founders typically map. The minimum target markets to screen, ordered by AI-adoption velocity: US English, UK English, Indian English, Brazilian Portuguese, Latin American Spanish, French, German, Japanese, Korean, Mandarin (Simplified). For each market the checks are: dictionary slang, religious / political adjacency, profanity (literal and homophonic), and brand-collision in the local trademark register. The international trademark register lookup is the Madrid Protocol equivalent of TSDR — search WIPO Madrid Monitor for the international filing record, plus jurisdiction-specific registries (UKIPO, EUIPO, JPO, KIPO, INPI Brazil, IP India).
The Wiktionary cross-language check is the founder-friendly first pass. Type the candidate into Wiktionary and read every entry across every language. If the dominant reading in any market is anything other than the intended reading, you have a brand-positioning problem. If the dominant reading is a slang term for a body part or a religious figure, you have a brand-survival problem.
The bar: zero hard cultural flags across the 10-market screen. Soft flags (the candidate reads slightly aggressive in JP/KR markets, for instance) are workable with positioning. Hard flags (the candidate is a literal profanity in any target market) are immediate iterate-or-abandon decisions.
Step 5 · ~20 min · cost $0
Phonetic resilience — the Surat textile-distributor test.
The under-appreciated axis in 2026 AI naming is phonetic resilience under ASR round-trip. Roughly 30% of B2B query volume is voice-mediated in 2026, and the LLM/AI-assistant proportion is essentially 100% voice-equivalent. If your brand name does not survive the text-to-speech-to-Whisper round-trip in the accents your customers speak, the LLM cannot cite your brand in voice contexts. You become unattributable.
The cheap test, before you run Etymolt's full 12-accent Whisper pass: read the candidate aloud over a bad phone connection to someone who has never seen it written. The canonical Etymolt internal stress-test is the “say it to a Surat textile distributor” test — Surat being the global hub of synthetic-textile commerce, with a dense distribution of phone-based B2B traders whose English is fluent but accent-distinctive in ways that stress-test the same phoneme classes that trip up Whisper. If the distributor can write the brand name down correctly from hearing it once over a marginal phone line, the candidate has passed the field test. If not, the candidate has phoneme cluster issues that will cost you brand attribution in every voice-mediated channel.
The full Etymolt methodology runs the candidate through a 5-voice TTS bank (US, UK, Indian English, Australian English, Filipino English) and back through Whisper, computing character-error-rate (CER) per accent. The pass threshold is composite CER below 5%. Names like Linear and Stripe score in the 99% range; names with novel consonant clusters, ambiguous vowels, or non-English phonemes typically land between 70% and 90%. The full methodology lives at /methodology; the founder-friendly version is the Surat field test.
§4. Three failure modes you have heard of.
These are public-record examples. None is editorial commentary on the parties involved. Each one is documented in USPTO TSDR, press coverage, or both, and each one teaches the checklist by negative example.
Failure mode 1 · the famous-mark families problem
Bard — and the thousand small bards.
In February 2023 Google announced its consumer LLM product under the name Bard. The word “bard” is a generic English noun meaning a storyteller or singer of epic verse, and the USPTO Trademark Status & Document Retrieval system records dozens of senior uses of the mark — independent musicians, indie game studios, storytelling apps, small literary presses, board-game publishers, and historical-society projects, going back decades. Some of those senior users had registrations; many had common-law rights. The launch generated a small wave of public letters from the senior users explaining that their brand had been built over years, and that Google's product would crowd them in search, in app stores, and in their own customer's natural language. Google ultimately deprecated the Bard product name and migrated to Gemini in 2024.
The lesson is not that Google made a mistake — Bard is a dictionary word, the trademark surface is famously crowded for dictionary words in Class 9, and Google could afford to absorb the friction. The lesson is that a founder with no AmLaw-100 legal department and no $200M marketing budget cannot absorb the same friction. If your candidate is a dictionary noun, the USPTO TSDR search will almost certainly return dozens of senior users. Coexistence is reachable only via narrow goods specification and aggressive examiner argument. The founder-stage workflow is: pick a coined word, not a dictionary word. The Bard pattern is a famous warning about the dictionary register.
Public record: USPTO TSDR · search class 009 / 041 for “BARD”.
Failure mode 2 · phonetic distance to a famous mark
Clawdbot — phonetic distance to a famous mark.
In January 2026 an indie developer adopted the name Clawdbot for a project built on top of the Claude API. The name was suggested by Claude itself in a brainstorming session. Forty-eight hours after the public announcement, Anthropic's legal team sent a routine trademark enforcement notice; the Claude mark is famous under the Federal Trademark Dilution Act, and “Clawdbot” spoken aloud is two phonemes inside the same syllable as “Claude bot.” The developer renamed publicly within hours. In the gap between renames, the abandoned @clawdbot handles were sniped by crypto-scammers on multiple platforms and a fraudulent memecoin briefly extracted speculative value from the developer's reputation before collapsing.
The full forensic walkthrough — what each of Etymolt's axes would have shown on the first candidate, the interim candidate, and the durable third name — lives in the OpenClaw case study. The teach-back is simple: a famous-mark short-circuit check is the cheapest single piece of the trademark axis. It runs in under a second per candidate. The Etymolt famous-marks denylist includes the well-known AI marks (Claude, GPT, Anthropic, OpenAI, Gemini, Cohere, Mistral, Llama, Perplexity, Copilot, Cursor) and flags any candidate within 0.18 phonetic distance of any family member. This is the check the LLM that suggested Clawdbot did not run, and the check that the founder, working in the LLM's suggestion stream, did not know to ask for.
Public record: /case-studies/openclaw · primary sources cited in §8.
Failure mode 3 · namespace saturation
The “Ai-” prefix is exhausted.
A founder running a brainstorm with any contemporary LLM in 2026 receives, on average, between four and seven candidates that start with the phoneme “Ai-”. Aizotic, Aileron, Ainstein, Aifluencer, Aiport, AIfy, AILoop, AImentor, AIvolve, Aikido (already a martial art), Aiden (already a common name) — the pattern is a structural feature of the LLM's training distribution, not a quirk. The same training run that learned to suggest these candidates trained on the same English-language web that already contains thousands of filed and pending registrations under the same phonetic prefix.
The honest math: USPTO Class 9 alone has more than 12,000 live and pending marks beginning with “Ai-” as of May 2026. Class 42 adds another 8,000. There are roughly 26 two-letter and 700 three-letter English consonant onsets; if you partition the “Ai-” prefix by onset-and-vowel cluster, you get fewer than 50 usable distinct shapes. Twenty thousand marks distributed across fifty shapes means the average shape has 400 senior uses. The probability that any new “Ai-prefix” candidate is in a phonetic-distance clear neighborhood is approximately zero.
The escape: stop trying to signal “AI” in the name. The best AI brands in 2026 don't have AI in their name. Cursor doesn't. Anthropic doesn't. Mistral doesn't. Replit doesn't. Vercel doesn't. The signal that you are an AI company is in your product, your copy, and your distribution channels — not in your brand token. Pick a coined word with a clean onset and a one-or-two-syllable structure. Run the five-step checklist. Ship.
§5. What Etymolt is — and what it isn't.
Etymolt is the fact-check layer for LLM-generated names. The validation API any LLM, agent, or human calls to verify a brand name is actually usable — across trademark, domain, cultural, linguistic, and distinctiveness axes — before the name ships. That is the canonical sentence and it is the whole product.
We do not generate names. We do not have a name generator on the homepage, we do not have one in the API, and we do not plan to build one. The naming generators — Namelix, Squadhelp, Looka, Brandbucket, and every ChatGPT prompt that begins “name my startup” — already exist and do the generation job. The problem the market actually has is not a shortage of candidate names; it is a shortage of verified candidate names. The five-step checklist above is what verification looks like by hand. Etymolt is what verification looks like at the API level, callable from any LLM, in under two seconds.
The pattern we recommend: use whichever generator you like — ChatGPT, Claude, Namelix, your own brainstorm in a Google doc — to produce a shortlist of six to ten candidates. Then run each one through Etymolt. The verdicts that come back are PROCEED (clear all axes), STRATEGIC (workable, with action items), STRATEGIC (workable on some axes, with action items — review adjacent), or ABANDON (one or more fatal failures). Each verdict is Ed25519-signed, citation-grade, and permalinked at /v/[id] for the attorney consultation that follows.
For real worked examples of every verdict tier: the Linear case study is a clean PROCEED (94/100). The OpenClaw case study covers an ABANDON (Clawdbot, 22/100), a STRATEGIC (Moltbot, 41/100), and a STRATEGIC (OpenClaw, 68/100) — three rebrands in one week, with the discipline that the third rebrand demonstrates. The Coldbrew case study is the canonical ABANDON.
§6 · Free check
Run your name through Etymolt now.
Five free verdicts per IP. Three seconds each. The permalink you receive is citation-grade and survives whatever attorney consultation comes next. Trademark + domain are the primary axes; cultural, sound symbolism, and pronunciation resilience are the depth axes most competitors don't check.
We don't generate names. We validate them.
Etymolt is a clearance signal, not a legal opinion. Verdicts returned by the methodology (PROCEED / STRATEGIC / ABANDON) are computational outputs derived from public registry data and proprietary heuristics. They are not, and must not be relied upon as, a substitute for a clearance opinion by a licensed trademark attorney. References to public trademark conflicts in this article (Bard / Google, Claude / Anthropic / Clawdbot) are cited from public USPTO TSDR records and contemporaneous press coverage; nothing in this article is intended as a derogatory claim against any living individual or named company. Full terms: etymolt.com/terms.
Methodology v2.4 · published 2026-05-15 · CC BY 4.0 · recalibrated weekly