Why Name Generators All Sound the Same (We Fixed Ours)

A name checker that also proposes names lives or dies on the quality of those proposals, so it was important we got it right. We thought we did, until an anonymous user stomped on us with the most mundane search query.

That one query sent us back to the start, and it began with a question we thought we had already settled.

Everyone Knows that Naming is Hard, But Why?

There is an ancient proverb that 'fine things are very difficult' to know about, and

it certainly isn't easy to get to know about names (Cratylus 384b).

As limitless as the human mind is, one of its biggest fears is naming something wrong, especially when that name sticks forever. Yet as humans, we have the opportunity to become a nomenclator at least once in our lifetime, and because we sometimes understand the importance of a name, we give it our every ounce of reasoning such that we get stuck. Stuck like soaked bread trying to walk down a muddy path.

And with all those six packs of effort, we often miss the point. Before we sink our teeth into this, let us backtrack a base ten digit.

Naming is hard because there is a complex psychology to it and a good part of it works against the person doing the naming.

Let's start with memory.

Proper names are arbitrary by nature, since they tell you very little about the thing they point to. Researchers asked people to name cartoon and comic-strip characters and found that memory blocks happened more often with arbitrary names than with descriptive ones (Brédart & Valentine, 1998).

A proper name often says nothing, giving the mind very little to hold on to, and that is exactly what a freshly invented brand name is.

Then comes the ear.

Before anyone reads a word of your explanation, the sound of the name has already started talking. If I were given all the ports Raymond Reddington owned, it wouldn't hold how many times I have heard people say "I hear how your name sounds."

Well, not to me since I have a simple common name. More on this soon.

Richard Klink showed that the sounds inside a brand name can communicate things like size, speed, and weight, even without any marketing around it (Klink, 2000). Alter and Oppenheimer found that companies with easy-to-pronounce names outperformed those with awkward ones on the stock market in the short term (2006).

So a name is judged by how it feels in the mouth before it is judged by what it means. Yeah, we are still discussing the WHY behind "Naming is hard."

Finally comes the crowd.

When many candidate names compete for the same object, people spread out across them, and the stronger the competition, the wider the spread (Gualdoni et al., 2023). Two reasonable people can look at the same thing and honestly land on different names, and nobody is wrong. Which is how a team ends up stuck.

So, the memory, the ear, and the crowd. Yet there is a trap sitting on top of all three: Once you know what a name is supposed to mean, you cannot hear it the way a stranger does.

In a well-known experiment called Tappers and Listeners conducted by Elizabeth Newton, participants were assigned to roles of "tappers" or "listeners." Tappers tapped the rhythm of well-known songs (like "Happy Birthday") on a table, while listeners guessed the song.

Tappers predicted that listeners would guess the songs correctly 50% of the time. Yet, out of 120 songs tapped, listeners guessed only 3 correctly, a success rate of 2.5% (or 1 in 40).

Curse of knowledge was a term coined in 1989 by Colin Camerer, George Loewenstein, and Martin Weber in this article. Elizabeth's experiment demonstrated this discrepancy. The discrepancy where those with prior knowledge (tappers) struggle to imagine what it is like not to know it, leading them to overestimate the clarity of their communication.

Naming is a similar experiment, with you tapping the name and your target audience listening. You are set on choosing a name that your audience can quickly associate with your product, but it is supposed to be a proper name, and proper names are arbitrary. Because you already know what the name is meant to stand for, you cannot hear it the way a stranger does, so you are likely to overestimate how much of that meaning reaches them. And since the name becomes your product’s identity in the long run, you want to get it right. That was the point we missed, and a stranger’s search query was about to show us.

What in The Name of Agnes is This?

In our product, everything downstream after proposing names is verification. The domains, the handles, the AI associations are all verification. Even though it is important because individuals and businesses want to know how common their names are, whether it has the available domains, we banked on name generation more. We are not just a name checker. We propose too. And to be of value to you, what we propose has to be better than whatever you would come up with.

Hence, generation is the creative act people are actually buying, and it happens in the first ten seconds, while the user is deciding whether this product is smart or generic. So when we audited the pipeline, the question was not only "are these names available" but "does this feel like a naming agency or like autocomplete."

But what led to this audit? Oh yeah... Let's talk about the stranger's search query.

What in the name of the daughter of Elizabeth Keen is this?

The Stranger's Search Query

If you have used our product, you can already tell what is wrong with the query. That we generate names for ideas, not babies. Yet, this query unlocked a very interesting use case and made us reflect on how we could do better.

Anonymous searches are always public, so I could see the search result for this query. They were disappointing, mainly because the query was a wrong use of our product. But, it made me rethink the wall of prohibitions we had set in the name generation pipeline.

The pipeline was one model call with a system prompt that was mostly a list of things the model was not allowed to do. No dictionary words. Five to ten letters. No get, try, or use prefixes. No -ly, -ify, -hub, -lab, -base, -stack, and more suffixes. No two common English words glued together. No hyphens or numbers.

These rules were intentional. Usually language models naming a product will suggest useless names like "GetHub", "SmartFlow", and "AI Notes Pro" pretty much every time, because those are the statistical average of every product name in their training data. In fact, you can easily tell names that were generated with AI these days similar by staring at the combination. Hence, the prohibitions were a tourniquet.

The trouble is that the cure became the disease. Strip out every dictionary word, every compound, every recognizable suffix, and what is left? Invented five-to-ten-letter strings. Run the pipeline on a lawyer tool or a dog food brand and you get Veltrix, Moxen, and Pultra for both. The lists were available but characterless, and a characterless list is indistinguishable from every other generator on the planet (errrhh... I meant Internet).

When I said the pipeline was just one model call with a system prompt of prohibitions, I oversimplified. There is also a second-round retry: after the model returned up to forty names, we screened each one's .com through RDAP, fed the taken ones back as "do not suggest anything close to these", and asked again.

It was clever, though shallow.

Basically, there were two (2) specific problems:

1. The prompt banned personality, which was the cause of the embarrassing names generated for that stranger's search query.
2. Category and style were decorative as they were never sent to the model.

We Redesigned the Names Generation Pipeline

When we embarked on this journey, we kept asking:

Our name generation pipeline had felt like an oracle before the redesign. Oracles are single sources of truth. Hence, when they fail, everything else that depends on them does too.

To treat generation like a small creative agency, we first sharpen the brief. For those who have been following this blog, you can confirm that sharpening the brief (the search query) is a query rewriting technique.

We distilled the raw brief into a positioning sentence plus audience, tone, and category, and stored it on the search. This finally puts the stored category field and any style preference to work, and it gives every downstream call a crisp brief instead of a rambling paragraph.

Next, we introduced four naming studios. Instead of one prompt that must do everything, we introduced four specialized prompts that share a set of positive-framed rules, things like "every name must be pronounceable and letters only," and then take distinct strategies:

• Coined roots, fresh invented syllables
• Blends and re-spellings of two evocative roots, the Notion and Vimeo construction
• Classical roots, Latin, Greek, and Sanskrit, translated or lightly altered
• Foreign real words that are easy for English speakers to say and type

These naming studios run in parallel. The beauty of this approach is that each studio's strategy supplies the character that the old rules destroyed. Also, smaller niche prompts generally produce better results than larger ones trying to do everything at once.

Aside from the studios, we also introduced the editor and every studio's raw output passes through it. It reads the brief, the candidates, and the names already kept, then drops near-duplicates and names that collide with famous brands. Something called semantic dedupe that the old pipeline never had.

We still screened each one's .com availability, but we ensured we streamed the updates to the frontend this time. Each studio pipeline invents, get judged, gets screened, and writes its survivors to the database immediately, in parallel with the other studios. Since the results page polls the database, names start appearing while the progress bar is still moving.

Isn't This More Expensive than The Oracle?

Haha! Nope! In fact, it was about 1.3 to 1.5 times cheaper. Look at the table below and compare for yourself.

	Before	After
LLM calls per search	1 to 2 big ones	9 small ones
Names per call	Up to 40 in one response	3 to 7 per studio
Strategies	1 generic	4 distinct
Near-duplicate handling	String match only	Semantic judge
First names on screen	15 to 40 seconds	About 5 seconds
Metadata used	Category ignored	Sharpened into every prompt
Worst-case cost	Roughly 2 to 3 cents	Roughly 1.5 to 2 cents

The new pipeline makes about nine times as many calls and costs less. The reason is output tokens, which are the expensive direction. The old call asked for forty names and their rationales in a single response, which is roughly one and a half to two thousand (1500-2000) output tokens in one shot. The new pipeline splits the same total work across nine calls that each return a couple hundred tokens, and output does not scale with the number of calls the way a single giant response does.

Though, making nine parallel calls per search means four times the burst against our model provider.

Here is a screenshot of the kind of results I was getting back with the new pipeline on the search query:

A platform to lean Igbo creatively using Nigerian and Igbo culture and roots as backbone.

I hope you can spot the creativity it exudes. Until next time. Peace.

References

Alter, A. L., & Oppenheimer, D. M. (2006). Predicting short-term stock fluctuations by using processing fluency. Proceedings of the National Academy of Sciences, 103(24), 9369-9372. https://doi.org/10.1073/pnas.0601071103

Brédart, S., & Valentine, T. (1998). Descriptiveness and proper name retrieval. Memory, 6(2), 199-206. https://orbi.uliege.be/handle/2268/13755

Gualdoni, E., Brochhagen, T., Mädebach, A., & Boleda, G. (2023). What's in a name? A large-scale computational study on how competition between names affects naming variation. Journal of Memory and Language, 133, 104459. https://doi.org/10.1016/j.jml.2023.104459

Klink, R. R. (2000). Creating brand names with meaning: The use of sound symbolism. Marketing Letters, 11(1), 5-20. https://doi.org/10.1023/A:1008184423824

The Decision Lab. (n.d.). Curse of knowledge. https://www.thedecisionlab.com/reference-guide/management/curse-of-knowledge