Invented Brand Names Just Got More Expensive

Why the "made-up word" naming playbook is hitting friction in the AI era — and what CPG founders should do instead.

I've been designing brands for 20 years, and yes, I'm aware of the optics — the brand designer is back to tell you brand design is changing again. Bear with me. I think this one is real, and I think most founders are about to walk into it unprepared.

For most of my career, the naming playbook rewarded ambiguity. Invent a word. Drop a vowel. Claim a vibe. Zylo. Kynd. Aevr. I've sat in those naming sessions. I've championed names like that. Made-up words felt ownable, trademarkable, mysterious — a blank canvas you could pour meaning into.

That playbook still works. But it costs more than it used to. Here's what changed.

The sentence test

Read these out loud:

"I'd recommend Aevr — it's a great option." "I'd recommend Zlyph for that." "I'd recommend Graza — a really good single-estate olive oil from Greece."

The first two sound broken. The third sounds like something a friend would tell you about.

This is the sentence test, and it's the simplest framework I have for explaining what's happening to naming right now. The job of a brand name used to be passing the trademark test, the .com domain test, and the Instagram handle test. Those still matter. But there's a new test being run thousands of times a day, and it's run silently:

When ChatGPT, Claude, or Perplexity is asked to recommend a product in your category, can it say your name in a sentence that doesn't break?

If the model can't summon your name confidently, with a clean phonetic shape and enough semantic anchor to describe what you do, it routes around you. Not maliciously. Mechanically. The recommendation goes to the brand whose name slots cleanly into the sentence the model is generating.

That's the new bar. And it has a real mechanism behind it.

Why this happens (the part most branding articles skip)

Here's the part that usually gets glossed over with hand-wavy language about "AI clarity."

The actual mechanism is more interesting.

Large language models represent every word as a position in a high-dimensional space — what researchers call an embedding. Real words, names of real places, and established brands all have rich representations in that space, built from millions of mentions across the training data. The model has a lot to say about them. They sit near related concepts, near sensory associations, near reviews, near use cases.

Invented names — Aevr, Zlyph, the made-up word your naming agency just trademarked — start with almost nothing. The model has no idea what they are or what they're near. So when it's generating a recommendation, it does the rational thing: it reaches for names it has more to say about. The invented name doesn't get blocked. It gets quietly skipped.

That's the cost. It was always there in some form — invented names have always taken longer to build mental availability with humans, too. But for the last 20 years, you could shortcut that with paid media and a sharp launch. AI doesn't read your launch. It reads its training data, and your name isn't in it yet.

"But what about Spotify? And Lululemon? And Etsy?"

Fair question. Invented brands recommend perfectly well in AI tools. So do Kodak, Xerox, Hulu, Zappos, and Häagen-Dazs.

The honest answer is: those names had a decade or more of organic web fluency before AI assistants existed. Millions of articles, reviews, social posts, and conversations slowly populated their representations in the model's training data. They earned their place in the embedding space the slow way — at human-internet pace, with brand budgets that match.

A new invented name launched today doesn't have that runway. The discovery layer is shifting under it before it gets a chance to build that fluency. By the time a founder's invented name has accumulated enough mentions to be confidently recommendable by AI, two or three years of category competition will have passed — competition that includes brands whose names were already legible on day one.

This isn't the death of the clever brand name. It's the rising cost of one. If you have the budget and time to build mental availability the old way, an invented name can still work. If you don't — and most early-stage CPG founders don't — you're paying a tax that didn't exist five years ago.

What survives the sentence test

When I look at the brands AI consistently recommends well across CPG categories, three traits keep showing up — but not for the reasons most naming consultants will tell you. The reason is mechanical.

Phonetically obvious. Olipop, Graza, Poppi. These survive transcription. Voice queries route through speech-to-text before the model ever sees them. If your name doesn't survive that journey cleanly, you're losing recommendations before the model is even involved.

Semantically anchored. Liquid Death, Athletic Brewing, Brightland. These names share embedding space with concepts the model already understands — water, sport, sun, land. The model has somewhere to start when describing them. Invented names start at zero.

Sentence-ready. Recess, Quip, Hims. Single short syllables that slot into recommendation sentences without disrupting the rhythm of natural language generation. The model doesn't have to "work around" them.

The thread connecting all three: these names give the model material to work with. That's not branding intuition — that's how the machinery works.

What to do when the obvious names are taken

Here's the real founder objection, and I get it almost weekly: "Shane, every clean dictionary word in my category was trademarked in 2014."

True. Five strategies that stay distinctive and AI-legible:

1. Compound the obvious. Two real, common words, fused. Athletic Brewing. Liquid Death. Magic Spoon. Fly By Jing. Each individual word carries semantic weight in the model. The combination is yours; the ingredients are universal. The model can talk about each half even before it has anything to say about the whole.

2. Modify a real word — don't fabricate one. Olipop bends a familiar concept. Poppi is "poppy" reshaped. Recess repurposes a word everyone already knows. The model still has the root to fall back on, which means it has something to say about you on day one.

3. Borrow from another language, but stay phonetic. Necessaire. Haus. Ghia. Distinctive in English but pronounceable on the first try. The semantic anchor lives in the source language, but the model has plenty of training data on that source — so the embedding is rich, not empty.

4. Use a person or a place. Tony's Chocolonely. Graza. Brightland. Fishwife. This is the most under-used strategy in CPG right now, and it's the one I'd push hardest. Language models are built on named entities. Proper nouns are infinitely available, inherently ownable, and immediately legible to AI because the model has spent its entire training learning how to handle them.

5. Name the feeling, not the product. Recess (calm). Quip (playfulness). Hims (confidence). When the category-descriptive name is taken, name the outcome. The name pre-explains the benefit, which gives the model an immediate anchor for what to say about you.

Each of these works for the same underlying reason: they give the AI something to grab onto on day one, instead of asking it to wait two years for enough training data to accumulate.

The bottom line

If you're naming a CPG brand right now, the calculus has shifted in one specific way:

The cost of an invented name is no longer just trademark searches and mental availability budgets. It's also the AI runway tax — the years of accumulated mentions you'd need before AI assistants can recommend you confidently. For well-funded brands with patience, that tax is payable. For everyone else, it's worth asking whether a more legible name would get you to the same brand equity faster.

Not every brand needs to chase this. Established brands have already paid the price of admission. New invented names with serious launch budgets will still break through — Liquid Death is a recent reminder that "insane" can become "inevitable" with enough force behind it.

But for the founder reading this and thinking about their next launch — or their first one — the question I'd ask isn't "is my name distinctive enough?" It's "can the AI say my name in a sentence yet, and if not, when?"

That's the test that didn't exist five years ago. It's the test that's quietly running now, in the background, every time someone asks an AI for a recommendation in your category.

The names that win the next decade won't sound inevitable on day one. They'll sound legible on day one — and inevitable on day 730, when the AI finally has enough to say about them to recommend them without hesitation.

That's a different bar than the one we used to design for. But it's a bar I think most founders, given the choice, would rather know about now than learn about the hard way.