I asked the generator for names for a UK accounting practice and it gave me twenty, one of which was "Sage Ledger Consulting". Sage is a FTSE 100 company whose main product is accounting software. The list did not mention trademarks. The run before it, for a bike repair business, had ended with a note telling me to check trademarks before choosing, which is the one piece of genuinely protective advice in the whole output and which that second run simply did not produce. Both runs took under three minutes and both looked equally finished. Telling them apart is the part of naming a business that a generator cannot do for you.
TL;DR: Generating names is fast and reliable in the one way that is easy to measure: twenty names, both times, in 60 and 139 seconds. Everything else varied. One run explained every name and the other explained none, though the prompt asked for reasons both times. A requested keyword appeared in half the names. And the list included a name belonging to a major company in the same field. Use it to get past the blank page, then do the checking yourself. The business name generator runs on the site's own server, so nothing goes to a third-party AI service.
What does a name generator give you?
A list of twenty candidates from a sentence about the business, a style, and optionally a word to work in. That is the whole contract, and the count is the part it keeps.
Two runs against the live tool:
| Input | Time | Output | Names |
|---|---|---|---|
| Mobile bike repair, Playful, no keyword | 60 s | 495 bytes | 20 |
| UK accounting for freelancers, Premium, "ledger" | 139 s | 1,821 bytes | 20 |
Twenty both times, numbered, no duplicates within a run. If what you need is to stop looking at an empty page, that is the job done in about the time it takes to make coffee.
What you have at the end is a list of candidates, not a shortlist. The difference between those two words is the rest of this guide.
Why did one run take more than twice as long?
Because the second one wrote more, and the model writes at a roughly constant rate.
495 bytes in 60 seconds against 1,821 bytes in 139 seconds works out at about 8 and 13 bytes a second, which on a small model running on a CPU-only server is the expected order. The time is the output length, not the difficulty of the request.
Which means a run that comes back quickly has usually given you less, not thought less. The 60 second run was fast because it produced bare names with nothing attached.
Other people using the AI tools matters too. Several requests share one model, so a queue adds to this before any generating starts.
Does it explain the names?
Sometimes, and the prompt asks for it every time.
The instruction sent with both runs ends "For each: the name and a short reason it works." Here is what came back.
Run one, in full for the first few:
1. Steady Rides
2. WheelWise Solutions
3. Quick Cycle
4. HomeHoovely
Run two:
1. **Sterling Ledger:** - Suggests quality & reliability; reflects the core service.
2. **Chartered Ledger:** - Hints at a formal and established business.
3. **Nexus Accounting:** - implies connections and coordination.
Twenty out of twenty explained in the second run, zero out of twenty in the first. Same tool, same prompt, same day.
That is worth knowing before you judge the output. A bare list is not the tool deciding these names need no explanation; it is a different roll of the same dice. If the reasons are what you wanted, run it again.
Does it use the word you asked it to include?
Partly. The keyword is a suggestion to the model, not a filter applied afterwards.
Run two asked for "ledger" and ten of the twenty names contain it: Sterling Ledger, Chartered Ledger, Sentinel Ledger Group, Precision Ledger Solutions, Bluebell Ledger Services, Cardinal Ledger, River Ledger, Sage Ledger Consulting and two more. The other ten do not.
Half is a reasonable outcome, and arguably the better one: a list where all twenty contain the same word is twenty variations on one idea. But if you need every suggestion built around a particular word, this will not give you that, and no amount of insisting in the box will make it a hard constraint.
What goes wrong with the names themselves?
Three distinct failures, and they need different treatment.
Words that are not words. Run one offered "HomeHoovely". It is not a word, it does not mean anything, and it is hard to say and harder to spell, which the prompt specifically asked it to avoid.
Words that are words, used wrongly. Run two offered "Vestigial Accounting", glossed as "subtle and sophisticated". Vestigial means a degenerate remnant of something that used to have a function. It is a confident, fluent, completely wrong use of a real word, and it is the failure that gets through, because it reads as though someone chose it deliberately.
Names that are already somebody's. Covered next, and the one with consequences.
The first two are visible in a careful read. Check every unfamiliar word in a dictionary before you fall in love with it.
Why does it suggest a name that is already taken?
Because nothing in the process checks, and the model's training is exactly what makes it likely.
A name generator produces words that fit the field. Existing companies in that field also have names that fit the field, and the model learned from text in which those companies are discussed. The better the suggestion sounds for your industry, the more likely something in that industry already uses it.
"Sage Ledger Consulting" for a UK accounting practice is the clean example. Sage is one of the best known accounting software brands in the country. The model is not malfunctioning; it produced a highly relevant accounting word, and the reason that word is relevant is the reason it is taken.
So treat every name on the list as unchecked. The tool now says so under the output, every time, in text written by the page rather than requested from the model. That change came out of writing this guide: the reminder used to be part of the prompt, and the system message tells the model to reply with the content and nothing else, so the two instructions competed and the reminder appeared in one of my two runs. A caution that arrives half the time is worse than none, because its absence reads as approval.
What should you check before choosing?
Four things, in this order, because each one is cheaper than the next.
- Search the name plus your industry. Thirty seconds, and it eliminates most of the list.
- Check the domain. The domain search tool does this without the upsell a registrar's search comes with.
- Search the national trademark register. Free public search exists in every jurisdiction that matters: the UK IPO, the USPTO and the EUIPO all offer one. A registered mark in your class is a reason to stop, not a negotiation.
- Check the company register. In the UK, Companies House will refuse a name too similar to an existing one at the point of incorporation, and its rules on same-as and too-like names are stricter than people expect. The counter is a bad moment to find out.
None of this is optional and none of it is slow. The expensive version is discovering it after the signage.
Does the style setting change anything?
Yes, visibly, and more than the keyword does.
Playful produced "Spokes & Smiles", "The Pedal Pal" and "Backyard Bikes". Premium produced "Sterling Ledger", "Ascendancy Accounts" and "Cardinal Ledger". The register is consistent within a run and clearly different between them.
The six options are Modern, Classic, Playful, Premium, Descriptive and Invented word. The one worth thinking hardest about is the last: an invented word is the only category where a trademark search is likely to come back empty, and also the only one where nobody can spell it from hearing it. That is the trade behind every made-up startup name you have ever had to ask someone to repeat.
Why is this tool set to be unpredictable?
Because naming is the one job where a repeatable answer is the wrong answer.
Every AI tool on the site carries a temperature, which is how much randomness the model is allowed. Most sit between 0.5 and 0.8. Four are set to 1.0, the highest value in use, and they are the four naming tools: business names, podcast names, slogans and channel names.
That is deliberate. A grammar checker giving two different answers to the same sentence is broken. A name generator giving the same twenty names every time is useless after the first run, because the second run is what you do when none of the first twenty worked.
It also explains the variance in everything above. The reasons appearing in one run and not the other, the keyword landing in half the names: those are the same setting doing what it is for. Run it three times and treat the union of the lists as the output.
What makes a name easy to say and spell?
Two tests, both of which you can run on a list in a couple of minutes, and neither of which the generator applies despite being asked to.
The radio test. Say the name out loud to someone who has not seen it written, then ask them to write it down. A name that survives this can be given over a phone, heard in a podcast ad, or passed on by a customer to a friend. "Spokes & Smiles" survives it. "HomeHoovely" does not: the listener has to ask how it is spelled, and so does everyone they tell.
The ambiguity test. Look for the places English offers a choice. Is it -ly or -ley, -er or -or, one L or two, "and" or an ampersand, a hyphen or a space? Each fork is a share of people who will type the wrong one and find nothing, or find a competitor. "WheelWise" is one word with a capital in the middle, which is an invitation to type it as two, and a domain has no capitals at all.
The generator was told to produce names that are easy to say and spell. Both runs contained names that fail one test or the other, which is a useful illustration of the general rule: an instruction in a prompt is a preference, not a constraint. Run these tests yourself on whatever survives.
The exception worth naming is the invented word, which fails both tests on purpose. A made-up name is the one most likely to clear a trademark search, and the price of that is teaching everybody how to spell it. That is a marketing budget, not a naming decision.
Should the domain decide the name?
Less than it used to, and more than nothing.
The exact-match .com for any ordinary English phrase has been gone for twenty years, and paying a five-figure sum for one is a decision about marketing spend rather than about naming. What has changed is that nobody types domains any more. Customers search for the business name, and a name that ranks for itself matters more than a name that matches a domain.
What still matters is that there is something sensible to register and that it is not confusable with a competitor's. A workable order of preference:
| Option | Verdict |
|---|---|
| Exact .com, free | Take it |
| Exact .co.uk, .io, .dev or a national domain | Fine, and normal now |
| Name plus a short word, such as "getX" or "Xhq" | Workable, slightly dated |
| Exact .com parked and for sale | Walk away unless the budget is real |
| A near-identical domain to a competitor | Walk away regardless |
Check it before the shortlist, not after, because a name you have already decided on is a name you will overpay for. The domain search tool answers the question without the sense of urgency a registrar's search is designed to create.
How many names do you need to see?
More than twenty, which is why running it again is part of using it.
Of the forty names across my two runs, I would have taken perhaps three to a shortlist. That is a normal hit rate for naming, not a complaint about the tool. The value is in the raw material and the directions it suggests, not in any individual line.
A practical pattern: run it three or four times with different styles, paste everything into one list, cross off anything you cannot spell or say, then start checking what is left. The checking is slower than the generating, which is the right way round.
What it cannot tell you
Whether the name is any good for your business, which is the actual question.
It does not know your market, your competitors, what your customers call the thing you do, whether the name reads badly in another language, whether it sounds like something rude when said quickly, or whether the .com is parked by someone wanting five figures. It has no idea whether "Playful" suits a practice whose clients are worried about their tax return.
It also does not check anything. Not the domain, not the register, not the app stores, not an existing shop two streets away.
What it does is turn a blank page into forty starting points in a few minutes, which is a genuine service and a smaller one than the search volume for this query implies.
Why does this run on the site's own server?
Because the alternative is sending your business idea to a third-party AI service before you have told anyone else.
The description you type here is often the first written statement of a thing that does not exist yet. It goes to a small model running on the site's own hardware, not to an external API, and it is not stored unless you choose to send feedback on the result.
The cost is honest and visible in the numbers above: a small model on a CPU server takes a minute or two, and the output is less polished than a frontier model would produce. The trade is that nobody else sees the idea.
