You named your candle "Aurora" because the model liked it, and then discovered that nine other candle brands thought the same. Naming with AI is cheap, which is exactly the problem: the model happily gives you fifty names that sound great and have all been used.
The useful work is a process: generate wide, score against criteria you wrote, kill most names on checks that need no lawyer, then do the serious clearance with the right people. This post covers the prompts and a scoring sheet, plus a brand voice sheet prompt. The trademark section is pointers only. I'm not a lawyer, and nothing here is legal advice.
How do you brief the model for product names?
A good brief states constraints the model can't guess. Untested template:
I'm naming a [PRODUCT: e.g. cold-pressed coconut hair oil] for an Indian
D2C brand called [BRAND or "not yet named"].
Audience: [WHO, age, city tier, language they speak at home].
Price position: [budget / mid / premium].
Feeling we want: [3 words].
Feelings we must avoid: [3 words, e.g. "clinical", "cheap"].
Languages: names must be easy to say and spell for people who speak
[Hindi / Tamil / English / etc.]. Flag any name that means something
negative or rude in [LANGUAGES I SPECIFY].
Existing names in our range (do not echo them): [LIST]
Competitor names to stay clear of: [LIST]
Generate 50 names in 5 groups of 10 using these styles:
1 descriptive, 2 invented words, 3 words from [LANGUAGE] with a
meaning I will check, 4 compound words, 5 abstract metaphors.
For each name give: the name, the idea behind it in 6 words, and how it
is pronounced in plain English letters.
Do not claim any name is unused, unique or available.
The last line is deliberate. Models will say "this name is unique" with no basis. Ask for a flat ban.
If you pick words from an Indian language, have a native speaker check meaning, connotation and regional variants. The model can get these wrong, and some words mean different things across regions. The same applies to anything you translate; our translation quality checklist covers how to review it.
How do you score and cut the list?
Cut with criteria, not feelings. Write the scoring sheet before you read the names, so you don't tune it to your favourite. Score each shortlisted name 0 to 2 per row:
| Criterion | 0 | 1 | 2 |
|---|---|---|---|
| Say it over the phone and be understood | Needs spelling | Mostly | Clear |
| Easy to type from memory | Several spellings | One doubt | One obvious spelling |
| Distinct from competitors in your category | Near-identical | Similar | Clearly different |
| Fits the feeling you want | No | Partly | Yes |
| Works across the languages your customers use | Problem | Unclear | Fine |
| Room to grow (not too tied to one product) | Locked in | Some | Flexible |
| Domain / handle workable | None | Compromise | Good |
Don't add the scores up blindly. Treat a 0 on pronunciation or language as a veto. You can ask the model to fill the first drafts of the table as a sanity pass, but check each score yourself; the model will be generous.
Prompt for the cut:
Here are my 12 shortlisted names and my scoring table. For each name,
list the single biggest risk in one sentence: a similar well-known brand
name you are aware of, a pronunciation problem, a negative meaning in
the languages I listed. If you are not sure, write "unsure" rather than
guessing. Do not rate names as legally safe.
"Unsure" is an acceptable answer. Treat anything the model says about existing brands as a lead to check yourself, never as a clearance.
Where do you look before you commit? Trademark pointers
This is the part you shouldn't hand to a chatbot. I'm listing places to start, not telling you what the law requires. On 2026-10-09 the Trade Mark Public Search page (tmrsearch.ipindia.gov.in/tmrpublicsearch) loaded for me, but I did not run a search on it, so I can't describe how it behaves; open it yourself.
- India: the Trademark Registry run by the Controller General of Patents, Designs and Trade Marks has a public search tool online. Search for your exact name, close spellings and phonetic variants, in the class that covers your goods.
- International: WIPO's Global Brand Database lets you search trademark records from many countries. I could only load its verification gateway for this post, so I can't vouch for its coverage; check what it includes.
- Other countries you plan to sell in: each has its own register (the US one is run by the USPTO, for example). Check the relevant ones.
- Common-law use: a name can be in use without being registered. Search the web, marketplaces (Amazon, Flipkart, Nykaa and so on, as relevant), Instagram and app stores.
- Domain and handles: check a registrar for the domain and each platform for the handle. A name where every domain is taken may still be fine, but you should decide that on purpose.
- Then ask a professional. A trademark attorney or registered agent can tell you whether the name is registrable, in which classes, and what the risk is. Models can't do this, and a free search isn't clearance. Search results are incomplete by nature: pending applications, different spellings and similar-sounding marks matter.
I'd also flag one practical point: descriptive names (a name that just says what the product is) are generally harder to protect than invented ones, and a competitor can often use similar words. Whether that matters for you is a question for the attorney.
How do you write a brand voice sheet with AI?
Once you've chosen a name, the next lever is consistency. Feed the model real writing and extract the rules. Untested template:
Below are 5 pieces of our own writing that we think sound right
(product pages, an email, a customer reply). Extract a one-page voice sheet:
1. Five tone words with a one-line explanation each.
2. Average sentence length and how we vary it, estimated from the samples.
3. Ten words or phrases we use often, quoted exactly from the samples.
4. Ten words or phrases a reader would NOT expect from us, marked as
suggestions to ban (clearly labelled as your suggestion, not from samples).
5. Three do/don't pairs showing a flat sentence and a rewrite in our voice.
Quote only from the samples and put quotes in double quotes. If the
samples are too few to tell, say so.
SAMPLES: [PASTE]
Then paste the sheet at the top of your product-description, email and ad prompts. For why this works better than adjectives, see how to prompt AI to sound human, and for a worked product-description setup, ChatGPT prompts for Shopify product descriptions.
How do you generate taglines without clichés?
Give the model the product truth and ban the category clichés. Untested template:
Product: [NAME and one factual sentence]. A true differentiator we can
prove: [ONE FACT YOU CAN SUBSTANTIATE]. Write 20 taglines under 8 words.
Ban these clichés: "pure", "nature's best", "crafted with love",
[YOUR LIST]. Do not make health, income or performance claims.
Do not state anything not in the product truth above.
Taglines that make claims ("clinically proven", "chemical-free") need substantiation, and there are advertising rules for claims. I didn't research India's advertising standards for this post, so check the relevant regulator's guidance or ask a professional before using a claim.
What I'd skip
Skip asking the model whether a name is "available" or "trademarkable". Skip names you can't say aloud in a noisy room. Skip any name where the domain or handle forces an odd spelling your customers will mistype. All prompts here are untested templates and I haven't run them on a live model; the scoring table and the trademark pointers are a process, not clearance, and the legal step belongs to a professional.
