Established instruments. Deterministic analysis. Simulated respondents. Directional output, fast. This page is the long version, including the parts that don't flatter us.
How it works
Not a survey, a decision. "Should we launch this in Saudi?" "Which of these three concepts?" "What will people pay?" VETT works out what needs to be measured.
Every mission maps to an established research method with canonical question wording. Van Westendorp for price sensitivity. MaxDiff for feature trade-offs. Kano for satisfaction drivers. Exposed and control cells for campaign lift. Structural validators run at generation, before anything is fielded: an unbalanced MaxDiff set or a broken Kano pair is rejected and rebuilt, and if it still fails it comes back to you to fix rather than going out quietly.
Market, demographics, category behaviour, and any screening criteria you set. Screeners are enforced at generation, so the sample you specified is the sample that answers: no drop-off, no quota chasing, no fieldwork delay.
Every respondent completes the full instrument individually and gives its reasoning for each answer. On missions up to 50 respondents that reasoning is stored next to the response, and any question in your report opens the actual reasoning behind a sample of the answers. Most research never shows you that.
Every survey figure in your report is computed by code. NPS, top-two-box intent, optimal price point, demand index, segment sizes, lift: calculated by the same statistical routines a research analyst would apply. The language model writes the narrative around those numbers. It does not produce them.
Real artifacts from a completed VETT mission, not a mockup
In: the instrument
The grooming brands I use reflect my personal image and say something meaningful about who I am.
Canonical wording, as generated
Van Westendorp
At what price would [product] be SO EXPENSIVE you would not consider buying it?
Kano (functional)
How would you feel if [feature] WAS in the product?
MaxDiff
Of these 4 features, which is MOST important to you, and which is LEAST important?
NPS
How likely are you to recommend [brand] to a friend or colleague?
Out: the computed panel
n = 60Segmentation gate cleared at n = 50. Clustering ran. Key dimension: Status orientation.
n = 43 · signature vs sample mean
n = 17 · signature vs sample mean
Question wording is quoted from the instrument generator. The output panel is the stored analysis of a completed VETT reference mission (audience profiling, n = 60), run by us, not by a customer. Figures are reproduced as computed: k-means clustering with a fixed seed, so the same responses give the same segments every time.
If research is your job, you'll want to know exactly what this is. Here it is plainly.
| What you're testing | Method |
|---|---|
| Price sensitivity | Van Westendorp PSM, with Gabor-Granger follow-up |
| Feature and benefit trade-offs | MaxDiff, balanced-set validated |
| Satisfaction drivers | Kano, functional and dysfunctional pairs, Sauerwein evaluation |
| Loyalty and advocacy | NPS, CSAT, CES, standard wording |
| Campaign effect | Exposed vs control cells, balanced 50/50 split |
| Audience structure | k-means clustering, seeded and reproducible |
Wording follows the canonical form for each instrument. These aren't improvised questions with a methodology label attached afterwards. Structural validators run at generation: an unbalanced MaxDiff set or a broken Kano pair is rejected and rebuilt, and if the rebuild still fails, the instrument comes back to you to fix rather than going out quietly.
For every survey metric, the language model computes nothing that appears as a figure. Deterministic analysis modules take the raw response set and produce the metrics. Every raw response is stored. You can trace any survey number back to the answers underneath it, down to the individual respondent.
Ask your current research supplier for that.
Where the model does touch a number
Two figures in the product are model-produced, and it would be dishonest to let the sentence above imply otherwise:
Everything else in the table above is computed.
When the sample cannot support a headline figure, VETT does not show one. Pricing, feature roadmap, and market entry are hard gated at 30 responses. Segmentation is hard gated at 50. Below its gate the headline number is withheld outright, on the results page and in the PDF, PowerPoint and Excel exports alike, and the report states the sample size and the threshold it did not reach.
Withheld means the figure, not the evidence. The response distributions, per-question breakdowns, respondent profiles and verbatims are all still there. What is removed is the single confident number a reader would otherwise quote, because that is the part a small sample cannot support. Below 50 responses VETT also will not cluster, and you get an aggregate profile rather than segments it cannot stand behind.
Outside those four methods, a mission under 30 responses keeps its numbers and carries a directional note explaining that the read is strong on ranking and consensus but indicative on magnitudes. We say which of the two you are looking at rather than treating every small sample the same way.
Most tools will show you a number regardless of whether it means anything, and say nothing. VETT will tell you the sample is too small and show you nothing. That constraint is in the product, not in a footnote.
Established instruments. Deterministic analysis. Simulated respondents. Directional output, fast.
VETT models how a specified audience would respond. It isn't a human panel and doesn't pretend to be one. What it gives you is a structured, statistically disciplined read on a decision, at a cost and speed that makes testing routine instead of exceptional.
"Simulated respondents aren't real people."
Correct, and we say so on every export. The relevant question isn't whether it equals a panel, it's what you do today instead. For most decisions, the honest answer is a meeting, an opinion, and a guess. VETT replaces the guess, not the panel.
"Couldn't I just ask ChatGPT?"
You'd get prose. You wouldn't get a validated instrument, a specified and screened sample, per-respondent reasoning, deterministic statistics, significance gating, or a report you can hand to a board. The model is one component here, and on the survey metrics it's the component that doesn't touch the numbers.
"How do I know the AI isn't making the numbers up?"
Because on the survey metrics it never sees them. Responses go into statistical code; the model writes narrative from the output. Run the same responses through the analysis twice and you get identical figures. The two places the model does produce a number are named above, and neither is a survey metric.
"What about bias in the model?"
Real. Every research method carries sampling bias: panels skew toward people who take surveys for money. Ours skews differently. That's why VETT is built for direction, comparison, and screening, where relative signal matters more than an absolute point estimate.
"Isn't $9 too cheap to be serious?"
The cost of traditional research is fieldwork: recruiting, incentives, quality control, project management. Remove fieldwork and the instrument and the analysis remain. That's what you're paying for, and it's the part that carries the rigour.
"Where would you still use a panel?"
Regulatory or advertising claims needing human substantiation. Final go/no-go on large irreversible capital. Very small specialist B2B populations. Longitudinal tracking of the same individuals.
Use VETT to explore, screen, iterate, and narrow, and to answer the hundred smaller questions nobody was ever going to fund research for. Where a decision is large and irreversible, confirm with fieldwork.
VETT gives you established research frameworks, and the statistics that go with them, run in minutes instead of weeks.
That's a different product from a human panel. For most of the decisions you make in a year, it's the one that actually gets used.
Pick a decision, and the setup advisor maps it to the right instrument.