Using ChatGPT and Claude for Brand Naming

Large language models are high-volume, low-trust naming engines. They pattern-match training data—Stripe, Notion, Figma—then recombine morphemes (-ify, -ly, Get-). Useful for divergent brainstorming; dangerous as final arbiter without verification gates.

Anthropic and OpenAI document no trademark clearance, domain availability, or cross-lingual taboo checking in base models. Your job is the constraint layer.

Receipt #1: What LLMs actually optimize for

Models default to:

• Familiar morphemes — -ify, -io, Meta-, Neo- (overrepresented in training corpora)
• English phonotactics — mostly fluent, occasionally Xylqr-grade clusters
• Semantic plausibility — names that sound like startups, not legal clearance

NameMesh ground truth: 36% of ~5,500 YC companies use brandable coinages—LLMs mimic this distribution but cannot verify availability. Atom marketplace: 250,000+ vetted names at $2K–$8 figures exist because generation ≠ clearance.

Measured failure rate in practice: Teams running unconstrained 50-name prompts report ~70–85% rejection after domain + USPTO + fluency filters (agency benchmarks cited in Namedrop vs Atom comparisons: human contests $199–$1,299 vs AI bulk at $9–$20 with mandatory verification).

The constraint-first prompt architecture

Layer 1: Strategic brief (non-negotiable context)

Company: B2B API observability for Series A infra teams
Avoid: literal "monitor", "API", "observ"
Phonetic target: voiceless stops + front vowels (speed/precision)
Length: 5–8 letters, single word
TLD priority: .com exact match, else .ai
Negative: no hyphens, no Get/Try prefixes (78% consumer distrust, Atom 2024)

Layer 2: Output schema (forces auditability)

Request structured output:

| Column | Purpose | | --- | --- | | Name | Candidate string | | Syllables | Fluency check | | Phonetic rationale | Map to Lowrey/Shrum vowel research | | Domain guess | LLM often wrong—flag for API check | | Trademark risk flag | Self-critique pass | | Kill reason | Pre-fill if obvious failure |

Layer 3: Adversarial second pass

Prompt: "Act as trademark examiner + radio test host. Kill any name a stranger couldn't spell after hearing once. Cite homophones and negative connotations in Spanish, Mandarin, Hindi."

Claude and GPT-4-class models perform better on critique than generation—run generate → critique → regenerate from survivors.

Verification pipeline (automated + human)

Gate 1: Processing fluency (Alter/Oppenheimer protocol)

Read each survivor aloud to 3 people unfamiliar with the project. Any hesitation = reject. PNAS 2006: fluency affects trust judgments and even short-term financial perception.

Gate 2: Domain availability (API, not LLM)

Bulk-check .com, .ai, .io via registrar API or NewName.ai bulk search. Identity Digital: 88% exact-match on non-traditional TLDs for funded startups when .com taken.

Never trust model claims of "available"—hallucination rate spikes on niche strings.

Gate 3: Trademark pre-screen

USPTO TESS + EUIPO eSearch + WIPO Global Brand Database. LLMs miss similar marks in Class 42/9. Budget $2K–$5K attorney clearance before final selection—not optional for funded startups.

Gate 4: Social handles

Consistent @name on X, LinkedIn, GitHub. Atom: 77% trust damage when same name exists on different extension/handle.

Gate 5: Consumer panel (5–10 ICP members)

Blind rank on memorability and category fit—not "which do you like" (biased toward familiar morphemes).

ChatGPT vs Claude: practical differences (2026)

| Dimension | ChatGPT (GPT-4o+) | Claude (3.5/4-class) | | --- | --- | --- | | Volume | High creativity, more -ify drift | Slightly more conservative coinages | | Structure | Strong with JSON mode | Strong with artifact/table requests | | Self-critique | Good adversarial pass | Often more thorough on homophones | | Best use | Divergent bursts (100 names) | Critique + rationale documentation |

Workflow: Claude critique → ChatGPT divergent refill → human fluency gate → API verification.

Prompt templates that work

Template A: Phonetic-engineered batch

Generate 30 coined names for {category}. Each must use {consonant profile} and {vowel profile} per phonetic symbolism research. Output JSON array with fields: name, syllableCount, phoneticRationale, potentialConfusion. Exclude dictionary words and existing Fortune 500 names.

Template B: Compound descriptive (middle path)

Generate 20 two-morpheme compounds like ShipBob or SkillSync for {workflow}. Max 10 letters. No hyphens. Each morpheme must be pronounceable in English and Mandarin romanization.

Template C: Upgrade path naming

We're on GetProduct.ai post-seed. Generate 10 cleaner upgrade targets (Product.ai taken—suggest alternatives). Flag Atom 78% distrust for Get/Try prefixes.

Failure modes to hard-block

| LLM output pattern | Why reject | | --- | --- | | Noxa, Toxa | Negative phonetic echo (noxious/toxic) | | Get{Name}.ai | Atom 78% add-on distrust | | {Keyword}Hub, {Keyword}ly | 2015 SaaS template; crowded USPTO classes | | Unpronounceable clusters | Alter/Oppenheimer disfluency penalty | | "Available!" without API check | Hallucination |

China workflow add-on

• Run pinyin + tone check for unintended Mandarin homophones (ma/mā/mǎ/mà)
• Search CNIPA trademark database—USPTO clear ≠ China clear
• Test 2–4 character Han script adaptation—domain may be pinyin while brand is 汉字
• WeChat mini-program name independent of domain—align both before PR

Bottom line

ChatGPT and Claude are naming accelerators, not naming authorities. The winning stack:

1. Constraint-first prompts with phonetic targets
2. Adversarial critique pass
3. API verification (domain, trademark, handles)
4. Human fluency gate (radio test)
5. Attorney clearance on final 1–3

LLMs cut brainstorming from weeks to hours. Verification still separates $12 registrations from `$12K rebrand mistakes.

For systematic generation with built-in checks, see AI-powered domain generation. For phonetic strategy, see science of naming. For domain alignment, see brand name vs domain name.