Brand & Product Name Generator

Why brand naming deserves more than a thesaurus

Brand naming sits in a weird spot in product work. You can build the whole product, design the whole identity, raise a whole round — and the name is the one decision that everyone has an opinion on, that everyone remembers, and that you can't really change later without torching the goodwill you've already accumulated.

Most "AI brand name generators" you'll find online are dressed-up thesauruses. They mash together a noun, a suffix, and a vowel; they pretend that's a brand. GeMater treats a brand name as a small mathematical object — a string that has to feel sayable, look short on a wordmark, and pass a non-trivial numerical filter. About one candidate in nine survives. The ones that do tend to have a kind of structural calm: clean syllable structure, balanced consonant–vowel ratio, often two or three syllables, often endable on -a, -o, -on, or -en. That's not a stylistic choice — that's what falls out of a Markov chain restricted to a 3·6·9 digital root.

Open the generator with your seed →

What "3·6·9" buys a brand

The generator assigns each candidate name three different gematria totals (Simple, English, and Jewish) and only keeps names whose digital root lands on 3, 6, or 9 in every one of them. The math is unpacked in detail in the 3·6·9 pattern primer, but the practical brand-side effect is what matters here:

• A curatorial filter that isn't taste-coded. Most brand-name generators apply a "fits the vibe" filter via prompt engineering. GeMater's filter is a deterministic mathematical property of the name itself. Same brand, same number, every time.
• A small story to tell. "We picked it because every gematria method we ran it through collapses to 9" is a memorable founding anecdote, even if it sounds slightly esoteric. It's a story; the brand needs one.
• A self-contained constraint that produces unfamiliar names. Filtering by gematria means the engine returns sounds that the human ear keeps gravitating toward without trying — names that feel chosen, not assembled.

If you'd rather not lean into any of that, the names still work as neutral, pronounceable coinages. The numerology is the floor; the brand work is what you build on top.

How the brand-name pipeline runs

The character-level Markov chain is the same one described in the how GeMater works doc. For brand work the pipeline is:

1. Seed it with a sound — a 2–4 character prefix, not a whole noun. Sounds steer; nouns lock.
2. Generate. The engine returns 10–20 candidates per run.
3. The 3·6·9 gate drops about 8 in 9 candidates before they hit your screen.
4. A profanity + keyword-echo filter drops a few more.
5. A live .com check runs against Verisign's RDAP feed so each card shows whether the literal name.com is available right now.

You can save a name to a collection (the generator's bookmark icon — sign-in is free) and come back to it days later without losing the gematria detail or the availability state.

Sample 3·6·9 brand names

Each of the following passes the three-cipher digital-root gate. The list mixes registers — clinical-tech, warm-direct-to-consumer, and abstract-luxury — so you can see the model's range with a single seed.

Clinical-tech feeling — Nervex, Lumeris, Trelyx, Voronik, Quorel, Sevith, Aluvis, Maelix.

Warm DTC — Aleva, Mirien, Korva, Sonia, Marwen, Telia, Eloren, Sevana.

Abstract luxury — Aleon, Veld, Nyx, Sael, Korei, Tovan, Lirae, Aderis.

When you run the generator yourself, the list is fresh and the .com state is current. The samples above are illustrative — by the time you read this, some .coms in the set may have moved.

Brand-name checks beyond the gematria

Numerology isn't a substitute for due diligence. Every brand shortlist should pass an additional set of human checks:

• Trademark search in your class. Free, ~15 minutes on the USPTO TESS, or your local equivalent. The point isn't to find a perfect result — it's to spot the dealbreakers early.
• Phonetic spelling test. Tell the name to two strangers over a bad phone line and ask them to write it down. If they get it wrong more than once, the name will cost you airport-pickup-sign-typos and podcast-mention-misspellings forever.
• International check. Run the name past a fluent speaker of any language your brand might enter. Markov chains do not know about inappropriate connotations in Brazilian Portuguese.
• .com and the variants. The card's badge is honest but literal. Also check getname.com, name.app, the social handles, and (if it matters) the local ccTLD.
• Wordmark test. Type the name in your candidate brand font. Some letter combinations look wrong in display weights even when they read fine in body copy.
• Sub-brand stretch. If you might launch a sister product, make sure the parent name has a sub-brand pattern. Lumeris Studio and Lumeris Cloud should both read naturally.

The 3·6·9 filter does a lot of the early curation. These checks do the rest.

Frequently asked

Brand name or product name — different generator?

Same generator, different prompts. A brand name tends to be shorter, more abstract, and longer-lived; a product name tends to be more descriptive and built to be replaced. The 3·6·9 filter works for both. For brands, seed with a sound; for products, seed with a noun-fragment of the function ("forge", "rail", "lens", "weave").

Is "passes 3·6·9" really meaningful as branding?

Branding-wise, it's a constraint, and constraints produce character. The names that survive the filter share structural properties — clean syllables, balanced ratios, frequent vowel endings — that humans tend to find sayable. Whether you make that the brand story is up to you; some founders absolutely do, others treat it as private trivia.

Will the same brand name come up twice across runs?

Often, no — the chain is randomised. But the model has favourite sub-structures that recur. If you see a name twice, treat it as the model voting for it.

Can I prompt the generator with a competitor's name?

You can — the seed will bias the output toward that name's phonotactics — but you're playing with fire. The 3·6·9 filter doesn't know what's a competitor; the keyword-echo filter only catches one-letter mutations. Use the result as a direction sample, not a shortlist.

Does the generator handle multi-word brands?

Single-token output, by default. For multi-word brands ("Lumeris Studio", "Forge & Vale"), generate the head noun first and pair it manually with a generic English second word. The filter still applies to the gematria of the combined string if you paste both words into the calculator.

What about non-English brands?

The Markov model is trained on Latin-alphabet text, so its idiom is Anglophone-Latinate. It can produce names that look Italian, Iberian, or Old-French because those phonotactics overlap with the corpus — but it doesn't reliably make Hebrew, Japanese, or Tamil names. For non-Latin scripts, you're better off translating an English candidate and checking its gematria manually with the calculator.

Does the model ever return inappropriate names?

A profanity dictionary screens the output for obvious cases. The filter is not perfect; if you ever see something off, it'll still have passed the 3·6·9 gate (the gate doesn't care about meaning) — just keep scrolling.

Next steps

Brand naming is one of the few tasks where having a 9-to-1 winnowing filter before taste enters the loop is genuinely useful. Start in the generator. If you also need the literal .com, the domain name generator walks through the same engine with the availability badge in foreground.