How to Prompt AI to Name a Product

Ask a model for fifty product names and you get fifty versions of the same name. Six will be a real word plus ly, eight will be two nouns jammed together, and the rest will be Latin roots with the vowels filed off. The model is not being lazy. It is optimising for the average of what naming looks like, and the average is mush.

The fix is three prompts instead of one: constrain, generate with a forced spread, then attack. Roughly fifteen minutes, and it produces a shortlist you can actually defend.

Pass 1: write the brief, do not ask for names

The first prompt produces no names at all. It produces the constraints that make the second prompt work. Most bad naming output traces back to a missing constraint rather than a bad model.

You are helping name a product. Do not suggest any names yet.

Product: a scheduling tool for independent physiotherapists.
Buyer: clinic owners with 1 to 4 practitioners, not technical.
Bought after: a WhatsApp conversation and a 10 minute demo.
Sits next to: paper diaries and a general purpose calendar app.
Tone: calm, competent, slightly clinical. Not playful.
Must survive: being said out loud on the phone to a patient.
Avoid: anything implying medical advice or diagnosis.

Write the naming brief: 6 constraints a candidate name must satisfy,
and 4 things that would disqualify a name immediately.
Be specific enough that two people applying your list would agree.

Read the output before continuing. If a constraint is wrong, fix it here. Everything downstream inherits this, and correcting it later costs another full pass.

Pass 2: generate with a forced spread

This is where most people lose. Asking for more names gets more of the same cluster. Asking for names across named strategies gets genuine variety, because you are making the model move rather than letting it sample from one basin. It is the same trick as getting genuinely different options.

Using the brief above, produce exactly 4 names in each category:

1. Plain descriptive (says what it does, no cleverness)
2. Real English word, borrowed from an unrelated domain
3. Compound of two short real words
4. Invented but pronounceable, no more than 3 syllables
5. Person's name or place name

For each: the name, a 6 word rationale, and the single
strongest objection to it. No duplicates across categories.
Do not rank them.

Two details matter. Withholding the ranking stops the model collapsing toward one favourite and writing the rest as filler. And demanding the strongest objection inline means every candidate arrives pre-criticised, which makes the next pass sharper.

Pass 3: make it attack its own list

Models are agreeable by default and will defend anything they just produced. You have to ask for the opposite explicitly, which is the technique in making AI argue against itself.

Here are the 20 names. You are now a sceptical brand strategist
who thinks most of them are weak.

For each name, answer:
- Say it out loud on a phone call. Does it need spelling out?
- Does it collide with a well known product or company you know of?
- Does it read as a medical claim?
- In 5 years, if the product does more than scheduling, does the name trap it?

Then cut the list to the 5 that survive, and for each write the
one sentence a competitor would use to mock it.

The mockery line is not a joke. If you cannot live with the sentence, the name is wrong, and it is much cheaper to learn that now than after the logo.

What the three passes produce

Running the physiotherapy brief through the structure gives a list that splits by category rather than clustering, which is the whole point. A representative slice of what pass 2 returns:

Pass 3 then removes most of them, and the objections above are why. Notice that the brief constraint about surviving a phone call disqualifies an entire category on its own, which is exactly what a good brief should do. If your pass 3 cuts nothing, pass 1 was too vague.

Naming a feature is not naming a company

The structure scales down. Feature names carry no trademark burden, no domain requirement and no five year horizon, so the third pass collapses to two questions: does it collide with a term already used in the product, and would a new user guess what it does from the name alone.

It also scales up badly. Naming a company involves legal availability across several jurisdictions, a domain strategy and often a professional search. Use the three passes to generate candidates, then hand the shortlist to someone qualified before spending money on it.

The four checks the model cannot do

Everything above is idea generation. None of it is verification, and a model will assert availability with total confidence and no basis. Do these yourself before you commit.

The trademark check is the one people skip and the one that costs money. A model's opinion on whether a mark is registrable is worth nothing, and it will still give you one.

What good output looks like

From a real run of this structure on the physiotherapy example, the surviving shortlist mixed categories rather than clustering: one plain descriptive, two borrowed real words, one compound, one invented. That spread is the signal the process worked. If your final five are all the same shape, pass 2 did not do its job and you should rerun it with the categories weighted differently.

Then stop. Naming has a strong pull toward another round, and the second-best name almost never loses money. Validating the product idea itself is a better use of the next hour.

Where to run the checks

Two registers cover most of what a small product needs. The USPTO trademark search covers United States marks, and TMview searches EU and national registers together. Both are free, both take a few minutes, and both are the sort of thing people skip until a letter arrives.

Search the exact name, then search it without vowels and with an obvious substitution, because confusingly similar is the standard rather than identical. If you find a live mark in your own class, the name is finished. Move on rather than negotiating with yourself about it.

The general habit that makes all of this work is treating the model as an idea generator and yourself as the verifier, which is the thread running through the fundamentals of prompt engineering.

Frequently asked questions

Why not just ask for a hundred names and pick one?

Because volume without spread is repetition. A hundred names from one prompt typically occupy three or four shapes, so you are choosing from four options with extra reading. The category constraint is what buys you range, not the count.

Can AI check whether a name is trademarked?

No, and this is the most dangerous thing it will confidently get wrong. Registers are searchable databases that change daily and are not reliably in any model's training data. Search the registers directly, and take advice if the product matters.

Should I tell the model my competitors' names?

Yes, as an exclusion list rather than inspiration. Given competitor names as context, models drift toward the same phonetics and you end up with a near-clone. Name them and say the output must not rhyme with, alliterate with, or share a root with any of them.

Does this work for naming features rather than products?

Better, actually, because feature names carry fewer constraints and no trademark burden. Drop the last pass to two questions and the whole thing takes five minutes. Generic phrasing is the main risk there, covered in avoiding generic answers.