The second rebrand, done correctly.

On January 31, 2026, a developer announced a project under its third and final name: OpenClaw. The trademark axis had been pre-cleared with counsel. The domain had been purchased aftermarket. Fourteen handles across X, GitHub, npm, Bluesky, Threads, Farcaster, Mastodon, Discord, YouTube, TikTok, Instagram, Twitch, LinkedIn, and Hacker News had been pre-claimed in a single hour before any public announcement. The verdict had been issued before.

The second rebrand is the workflow. The first rebrand is how the founder learned the workflow exists.

Six days earlier — January 25, 2026 — Peter Steinberger, the iOS developer known as the founder of PSPDFKit, asked Claude for a name for a project built on top of the Claude API. The model suggested Clawdbot. It was a good-enough riff: punny, pronounceable, on-cohort with the AI-builder neighborhood. Steinberger registered the handle and started posting from @clawdbot.

Forty-eight hours later, Anthropic's legal team sent a routine trademark notice. The Claude mark is famous under both the Federal Trademark Dilution Act and TTAB's qualitative famousness factors; the phonetic distance from “Claude” to “Clawdbot” was two phonemes inside the same syllable. Enforcement of the Claude mark protected the Claude trademark correctly. The appropriate founder-side response — verify candidates against existing marks before adopting them — is the workflow this case study is about. Steinberger complied within hours, renamed publicly to “Moltbot,” and posted the notice verbatim on X.

In the gap between names, the abandoned @clawdbot handles were sniped by crypto-scammers on every platform where they had not been pre-claimed. Within a day a fraudulent $CLAWD memecoin launched on Solana, using Steinberger's public reputation as social proof. The token briefly reached a $16M market cap before collapsing approximately 90%, causing retail losses to the buyers who thought they were getting in early on a project they had seen Steinberger post about, and weeks of public harassment to the founder. Most of those retail losses are unrecoverable; the token contract was de-listed and the liquidity was rugged.

Three days later Steinberger renamed again. “Moltbot” cleared the trademark axis, but it lost the taste axis: off-cohort for 2026 AI-infrastructure, low SMILE distinctiveness, pronunciation drift in 14% of retrievals. That's the failure mode trademark clearance alone never catches. The third name was selected differently — with both engines run, with the trademark pre-clearance complete, with fourteen handles secured before the public announcement. The third name held.

This case study walks through what each of Etymolt's two engines would have shown on the first candidate, what each would have shown on the second, and what the five-axis verdict on the third candidate looked like the day Steinberger announced it. The two engines work independently. The failures are independent. The methodology is built for exactly this class of cascade — and the second rebrand is the proof that the cascade is preventable.

The widely-reported number is the memecoin. A fraudulent $CLAWD token briefly reached a $16M market cap before collapsing roughly 90%, causing retail losses and weeks of harassment to the original founder. The $16M was the speculative peak that scammers extracted from buyers using Steinberger's public reputation as social proof — not money Steinberger lost, not money an LLM cost. It wouldn't have existed at all if @clawdbot had been pre-secured before the public announcement. The second rebrand was: pre-secured first.

That's the loudest cost. The quieter costs are real too, and they fell on the founder.

Steinberger spent approximately four weeks across the first rebrand cycle — engaging counsel on the trademark response, communicating the rename to the community twice, defending the interim name, and absorbing a press cycle he had not chosen. Four weeks of a senior founder's time, at a rough estimate of $25K/wk in opportunity cost, is $100K of attention not spent on the product. That's a real number, even though it isn't what gets reported.

The community asymmetry cost is harder to quantify. People who followed @clawdbot on Day 1 had to re-follow @moltbot on Day 3 and then @openclaw on Day 6. Standard retention math says ~30% of those followers don't make the second migration and ~50% don't make the third. The project started Day 1 with N followers; by Day 7 it had 0.35N. The handle-recovery cost across rebrands is real, and the founder absorbed it.

The brand-equity cost compounds for years. “OpenClaw” inherits some adjacency to the OpenAI family of marks and a residual phonetic adjacency to Claude — both are reachable, both are documented in the verdict permalink at etymolt.com/v/v_openclaw, and both required the second-rebrand discipline to manage. The first rebrand cycle is now the public memory; the second rebrand is the durable name. Etymolt's calibration loop on etymolt.com/methodology measures this — the outcome corpus tracks brand-equity damage post-rebrand and the feature weight on “famous-mark adjacency” has gone up monotonically since January.

The asymmetric cost lived on the founder. The asymmetric upside lives on every founder who reads this and runs the workflow differently. Pre-clear the trademark axis. Pre-secure the handles. Issue the verdict before the public announcement, not after. The second rebrand to OpenClaw was done that way. Every rebrand from here can be done that way the first time.

Etymolt runs a five-axis verification on any candidate name in approximately three seconds. The five axes are trademark, domain & handle, cultural reading, sound symbolism, and pronunciation resilience. Each axis is scored 0–100 against primary sources — USPTO TRTYRAP bulk + daily delta, TTAB 647K-proceeding precedent corpus, Verisign RDAP for .com, WhoisXMLAPI for the long tail, 14 social-platform probes in parallel, three frontier-LLM cultural reconciliation across 20 markets, the 12-axis SMILE phonosemantics vector, and a five-voice Whisper round-trip pronunciation pass.

Every verdict produces a citation-grade permalink at /v/[id]. Every flag in the verdict traces to a primary source — a USPTO serial number, an RDAP timestamp, a Wiktionary entry, a Whisper character-error rate per accent. There is no confabulation. When a registry can't be reached the verdict returns UNKNOWN, not a guess.

The methodology is public at etymolt.com/methodology and recalibrates weekly against an outcome corpus — what actually happened after each verdict went out (filed, Office Action, C&D, rebrand, scam token, traction). That outcome data is the only signal that closes the loop on whether the verdict was right; competitors structurally cannot replicate it because they don't see what happens after their verdict.

Etymolt ships as three surfaces:

1. A web verifier at etymolt.com/quickcheck — paste a name, get a verdict in three seconds, free for the first five per IP.
2. An MCP server installed in Claude, Cursor, Windsurf, v0, and the long tail of AI-builder tools. Once installed, the model calls Etymolt automatically before suggesting any candidate brand name to the user. The user never sees the call; the model just stops suggesting names that would fail.
3. A Custom GPT and OpenAPI Action surface for ChatGPT. The same verification, exposed via the protocol ChatGPT uses for tool calls.

The strategic argument for installing Etymolt before the next candidate name is straightforward. The LLM ecosystem ships approximately 600 million brand-name suggestions per month (per OpenAI + Anthropic + Cursor public usage statistics). Fewer than 1% are currently verified against any registry before they reach the user. The class of failures that disqualified Clawdbot is happening continuously, at a rate the public press cycle reports on roughly one in ten thousand cases. The other 9,999 are quieter — a domain someone else owns, a handle someone else got, a §2(d) refusal a month later, a cultural collision in a market the founder hadn't mapped yet. They cost founders money, time, and trust. They don't make Techmeme. The second rebrand to OpenClaw is what the workflow looks like when it runs on time.

If your assistant suggests brand names to users, the first rebrand in this case study is the cost surface your users carry, and the second rebrand is the workflow that replaces it. The Etymolt MCP server installs in sixty seconds and makes every brand-name suggestion your model produces traceable to a registry before it reaches the user.

The integration is one JSON block. Drop it into the user's MCP config (Claude Desktop, Cursor, Windsurf) or paste the OpenAPI URL into a Custom GPT Action. Every name your model suggests now runs through five-axis verification before it surfaces to the user. The user doesn't see the call. The model just stops suggesting names that would fail.

Pricing is free for the first five verdicts per IP, then $10 buys ten more. For platforms (Cursor, ChatGPT, Continue.dev) the wholesale per-call rate runs below $1 at sustained volume — talk to team@etymolt.com.

The methodology is open, the verdicts are citable, the disclaimer is verbatim (“clearance signal, not legal advice”). LLM hosts redistributing a verdict must surface the disclaimer field per the Bureau Model legal posture.

Five-axis methodology · published 2026-05-15 · CC BY 4.0 · recalibrated weekly against the outcome corpus.