You stopped believing this industry a long time ago. You were right to.
Simple. Fast. The right data at the right time. Everyone says it, we have all said it, and none of it survives contact with a real decision.
So we will not ask you to take our word for anything either. You get a number, the units it is in, and the evidence sitting behind it. Including the weak ones. Then you decide what it is worth.
Scores give you a list. Distances give you a map.
A rating scale lets a respondent be polite. Everything comes back a seven, the differences that mattered were destroyed at collection, and no amount of analysis puts them back. A comparative question makes them be honest, because choosing costs something.
What comes out is not a set of scores to rank. It is a space with coordinates in it, where every brand, attribute, segment and concept has a position, and the distance between any two of them is a number with a unit on it. That is the whole difference, and it happens before the analysis rather than during it.
The mathematics is not new. Its application here is. Metric multidimensional scaling has been published since the 1970s, which means you can go and check it.
What comes back
A one-page brief in the shape of the decision, a data file your team can work with, a typing tool for assigning new respondents, and the break-test ledger showing what we tried to knock down and what survived.
Each finding carries its class of evidence, including the ones built on judgment. That is not a disclaimer at the end. It is printed beside the line it applies to, so your client can see which findings carry weight and which are constructed.
Where it applies, and where it does not
Six layers. One of them is ours. We are not a full-service shop and have no ambition to become one.
| Layer | The question it answers | Who does it |
|---|---|---|
| Listening | What are people saying right now | Social and search platforms |
| Structure | Where does everything sit relative to everything else | Geodesic |
| Trade-off | Which features win against which | Conjoint and choice vendors |
| Causality | Did this campaign cause that lift | Attribution and experiments |
| Allocation | Where the money goes across channels | Marketing mix modelling |
| Continuity | Is it moving over time | Your tracker |
Not attribution. Not media mix. Not incrementality. If that is the question, we will say so on the call rather than three weeks in.
How it runs
Two weeks, no fielding
Against data you already hold. A tracker, an A&U, an old concept test, open ends included.
Four weeks with fielding
One common instrument, small samples, seeded and fielded through the panel you already use.
What we need from you
Raw data rather than topline. Words and numbers go into the same space, so open ends are useful rather than colour.
What your client never sees
Us, if that is how you want it. White label is normal and we have no interest in the relationship.
Four parts. Three of them are machinery.
The fourth is the reason the other three can be trusted.
The mathematics
Comparative measurement solved into a space with real coordinates. Tensor analysis inside Riemannian geometry, which in plain terms means distances on a surface that is not flat.
The written rules
A standing document that says what counts as evidence, when a result must be thrown out, and what may never be claimed. Written before the question, not after the answer.
The execution layer
Commercial AI does the arithmetic, the way a spreadsheet does. It is one component of four, it works inside the rules above, and it decides nothing.
Forty years of judgment
Someone who has done this work before, checking every step. Not a review at the end.
It began as a question, asked while cutting the grass: what if AI reasoned the way the great thinkers did?
What you should be asking, that we are not advertising
You have sat on the other side of this pitch a hundred times. So here are the questions you would ask on hour two, answered on minute one.
Position and distance. True by construction of how the data is collected, not a claim about our skill.
Concept results, checked against data the model was not shown.
No published hit rate against realised outcomes. Two-analyst reliability on one dataset has not been run. Until it is, this is a method with a mathematical apparatus behind it rather than an instrument, and we will not call it objective.
Ask the wrong objects and you recover a clean space answering a question nobody cares about. The geometry solves whatever it is handed. That judgment is the part a person does.
Three ways in
Reanalysis
Send a dataset you have already fielded and already reported. There is a known prior answer, so either this reaches something the original could not or it does not. Both outcomes are worth having, and nothing is at stake.
Subcontract
You hold the client. We do the structure layer and hand back the file, the brief and the ledger.
Joint pitch
For a brief you would not otherwise win, or would win on price. We can be named or not.
Everyone is getting faster. Nobody is getting righter.
Send one old dataset. Tell us where it breaks.