Most naming projects start with a favourite. Someone on the team already loves one of the names, and the job of a test is to make that name earn its place before anyone pays for the trademark search.
The two projects below show how that works with Boomerang. Both are illustrative examples, composed from typical projects and not a specific client. The names and the results are made up to show the shape of a naming test, not to report one. The arc is the same in both because it is the one we would recommend: a simulation with Silicon Samples to cut the long list, then the shortlist in front of real people with Live Social Samples™, in the same tool and with the same questionnaire.
Same arc, two projects: the simulation cuts the list, people pick the name
Startup on boomerang.ai
- Simulation: names simulated
- People: fielded with people, n = 300
- Decision: into the trademark search
Agency on boomResearch
- Simulation: names simulated, with breakdowns
- People: fielded with people, n = 300
- Decision: recommended to the client
Illustrative projects, composed from typical naming work — not client data.
A founder with fourteen names
Lea runs a four-person company in Basel that sells refill tabs for household cleaner. The product was ready. The name was a placeholder nobody liked, and the list of alternatives on the office whiteboard had grown to fourteen.
She had never commissioned a survey and had no budget line for one. She signed up on boomerang.ai, which is self-service from the first click: you write the questions, choose who answers and start the study yourself.
Step one, the simulation. She put all fourteen names into a single question, asked which name people would most expect to see on a refill cleaner, and simulated Swiss adults aged 18–65+. For an overall result like this, 300–500 simulated respondents is enough, and a run takes 10–30 minutes depending on the size of the population. Results come in live, so she could watch the order settle while the run was still going.
Eleven names had almost no takers and were off the list that afternoon. Three were clearly ahead. The result carried a plausibility score of 6.8. Confidence Guard™, which is in beta but live, scores every simulated result on a 0–10 scale with a separate judge model, and for 6.0–7.9 its advice is short: direction, not decimals. Trust the ranking, not the individual percentage.
That was the honest reading. Refila led, but Klarbe was close enough behind that the order of the two could have been noise.
Step two, people. A name ends up on the packaging and the website, and changing it later costs more than a survey does. So she copied the questionnaire, cut it to the three finalists and sent it to real respondents: n = 300, recruited fresh on social media for this study. Fieldwork averages 7–10 days, depending on sample size and questionnaire complexity, and it is billed per completed answer.
Illustrative numbers, not client data. The share who picked each name as the one they would most expect on a refill cleaner:
The winner held. The percentages did not. In the simulation Refila and Klarbe were almost level; people put Refila well ahead. Pebbl, the name the team had liked least, came last both times.
An agency with a client and a deadline
The second example is a brand agency in Zurich, naming a new savings product for a regional bank. The agency works on boomResearch, which is self-service too: in-app and email support, onboarding on the Pro and Team plans, and a phone hotline with the largest credit pack. What matters for this story is that it carries the same two engines.
The creative team had 22 candidates that survived a first legal check. The client wanted to know whether the name worked across the country and not only in Zurich, so the agency needed breakdowns, which are available by region, age and gender. A breakdown needs a bigger base than a topline, and on the simulation side that means up to 1’000 simulated respondents.
The simulation cut the 22 names to three. This time the plausibility score came back at 8.4, in the top band. Confidence Guard™ describes 8.0–10 as high plausibility: internally consistent and in line with comparable surveys. It then adds a sentence that is easy to skip and shouldn’t be. For decisions with real budget behind them, anchor with human respondents as well.
Where the two simulations landed on Confidence Guard™
- 8.0–10
High plausibility. For decisions with real budget behind them, anchor with human respondents as well.
8.4 Agency, 22 names
- 6.0–7.9
Direction, not decimals. Trust the ranking, not the individual percentage.
6.8 Startup, 14 names
- below 6.0
Go to real people. Say it plainly and put the escalation next to the result.
Both scores are from the illustrative examples. A separate judge model scores plausibility 0–10; it measures internal consistency and fit to comparable surveys, never correctness about the world.
Confidence Guard™ bands, Boomerang Ideas (public wording)
A launch campaign for a bank is that kind of decision. The agency fielded the three names with n = 300 real respondents, same questionnaire, same tool.
Overall, people and simulation agreed on the winner and on the order behind it.
In the deck, every number carried a label saying which engine it came from. If the result is questioned later, nobody has to reconstruct which half was simulated.
What the two projects have in common
In both, the simulation did the part it is good at. It took a list far too long to put in front of people and made it short in an afternoon, without spending anyone’s time on names that were never going to win.
A simulated percentage is a guide to the order, not a forecast of the share.
In both, the ranking at the top held when people answered, and the individual percentages moved. That is what the confidence bands tell you to plan for. A simulated percentage is a guide to the order, not a forecast of the share.
And in both, people added something the simulation had not raised: an association in the comments, a region that leaned the other way. Neither finding changed the winner. Each one changed what the team did next.
Running it yourself
- Put the whole long list into the simulation and ask for a choice between names, not a rating of each one. People are better at preferring than at scoring.
- Read the plausibility score before you read the percentages.
- Take two or three names to real people and reuse the questionnaire, so the two results can be compared line by line.
- If a breakdown matters, size the sample for it up front. Standard breakdowns are by region, age and gender, not by income or customer status.
- Label every result with the engine it came from.
The simulation narrows the list. People confirm the choice before the money is spent on it.
⠿ verschieben
Als anonym