Skip to main content
Builder InsightsNovember 13, 20265 min read

We Tested Spark on 50 Business Briefs. Here Are the Results.

The recorded results behind Spark Beta: 50 synthetic briefs, three runs each, the failures we fixed and the limits the headline cannot show.

2 sources and further readingHow we prepare our guides

A naming tool can look convincing when you show only its best suggestion. We wanted a more useful view of Spark before opening its beta: could it keep producing a usable set when the business was specific, category words were banned, or the description said what the company did not do? We wrote 50 synthetic briefs across 16 industries and ran each with three fixed seeds. This article reports that development test, including the results that make the headline less tidy.

Spark test notebook: 50 briefs, 150 recorded runs, with results open for inspection.
A development test, not an independent ranking or a guarantee of name quality.

The current recorded version, Spark 3.3.5, passes the declared structure rules on all 50 briefs in all three runs. The complete notebook shows every brief and output. That does not mean every suggestion is a strong brand. It means each run retained at least four distinct names, respected its avoid list, avoided the specified filler roots and passed our review for obvious reversed grammar. Preference, distinctiveness in the market and availability are separate questions.

What was in the test?

Forty briefs described positive offers: revision software, bookkeeping, physical goods, logistics, local services, healthcare support, agriculture and more. Ten were negative controls. Those added explicit exclusions or sharp distinctions between the provider and its customers. Software for a veterinary practice, for example, should not turn into a pet-food brand merely because the description mentions animals. The test asks whether the engine keeps track of the offer rather than collecting every noun it sees.

We froze the briefs, settings and three seeds before the first capture. Each run requested a candidate pool, applied the existing eligibility filter, and kept at most the first eight eligible results. We did not pick the nicest seed for the headline. A brief counted as passing only if all three runs met the same four criteria. The original threshold was 35 passing concepts out of 50; the current result exceeds it.

Current recorded results, 13 November 2026

Measure
Briefs / runs
Recorded result
50 / 150
Measure
Briefs passing all three runs
Recorded result
50 / 50
Measure
Negative controls
Recorded result
10 briefs / 30 runs
Measure
Runs with eight names
Recorded result
133
Measure
Runs with seven names
Recorded result
5
Measure
Runs with six names
Recorded result
12
Measure
Empty sets
Recorded result
0
Measure
Displayed name occurrences
Recorded result
1,171; not 1,171 globally unique names

The first result was 33 out of 50

The earlier 3.3.2 capture passed 33 concepts. Sixteen concepts had at least one run with an obvious grammar problem, and one fell below four displayed names. Additional semantic review found four negative controls that drifted towards a denied activity or the customer’s sector. Those failures were useful: they exposed gaps that a gallery of favourite outputs would have hidden.

The fixes focused on interpreting the provider’s actual job, keeping negation active across lists of excluded services, and adding bounded vocabulary for missing tasks. Appointment coordination, bookkeeping, interview handoffs, speech practice and home maintenance needed more precise associations. Filtering also became stricter about sentence fragments and reversed constructions. Some sets became shorter because unsuitable candidates were removed rather than replaced with generic words.

We then reran the same frozen cases. That is a regression test: it checks whether the changes address known failures without breaking earlier cases. It is not an unseen holdout. Once an engine has been improved using a test set, a better score on that set cannot establish how well it will generalise to all future businesses. The public notebook therefore calls these development results, and keeps the actual outputs available for inspection.

Read a few directions before reading the score

Four recorded names from run one of selected cases

Offer
Refrigerated delivery planning software
Naming directions
MilePlan, StopSync, MapLink, FleetLoop
Offer
Electrical apprenticeship revision flashcards
Naming directions
NoteCue, TopicDeck, LessonLoop, MemoryStep
Offer
Interview coordination for charities
Naming directions
TeamLink, RolePlan, SlotHandoff, PanelSync
Offer
Refillable mineral-salt deodorant
Naming directions
SaltEase, DaySkin, StickGlide, RefillCare
Offer
Spare-change investing
Naming directions
SliceShare, SumGather, RoundReserve, SpareStep

These examples were selected to illustrate different offers, not declared the best brands in a ranking. MilePlan makes a task visible; NoteCue points towards recall; SaltEase takes a material and comfort direction. You may prefer a less literal or more unusual name. That disagreement is valuable because a structure test has no reliable way to decide what your audience should find compelling.

The complete set also includes more awkward suggestions, such as GraspLearn for tutoring. Adjacent businesses can receive overlapping names: appointment software for repair shops and veterinary practices uses a similar coordination vocabulary. This can be a sign of the engine recognising the same underlying task, but it also limits distinctiveness. We have retained those examples instead of presenting only the most flattering cases.

What a pass does not tell you

The grammar review was performed internally with assisted checks, not by a blind panel of founders. A name that avoids a reversed construction can still be dull, too literal, difficult to say in another language or poorly suited to a brand’s tone. The test records output consistency and constraint handling; it does not measure conversion, memorability in real customers or whether anyone would pay for the name.

No domain, company or trade mark lookup was part of these offline runs. A candidate is an idea, not a claim of exclusivity. Spark did not call a generative AI service for this test; that says nothing about the ordinary hosting cost of running a website. We are also not using local test timings as a promise about response speed on every connection or device.

How to use the evidence for your own shortlist

Find a test case near your actual offer and read the brief before judging its names. Check the avoid words, then compare all three runs. Notice whether the same roots dominate, whether an important difference disappears, and whether you could explain a candidate without repeating the tool’s rationale. A useful next step is to test your own wording and see whether the set moves in the direction you intended.

NamoLux combines Name Sprint, Bulk Check, and Founder Signal with the evidence founders need to compare a shortlist.

You can download the recorded JSON data, including the briefs, exclusions, seeds and displayed outputs. The release snapshot is dated so future engine changes do not silently rewrite this result. For the five additional cases involving conflicting positioning, multiple services and unusual niches, continue to our micro-niche test report.

✨

Try Spark free with a few keywords and compare the directions it suggests.

Try Spark

Sources and further reading

Related Articles

Builder Insights6 min read

How Spark’s Public Naming Beta Worked

A practical first session with Spark: write a focused brief, explore naming directions, save favourites and decide when a report is useful.

Read article

Bring the candidates into one view.

Check domains and compare the strongest options together.

Check a shortlist