The asset is the history, not the algorithm
The history is an archive of products measured on the same scale throughout, not a folder of unrelated studies. That is what makes a model trained on top of it worth anything.
Comparing two products measured years apart only works if the instrument has not changed in between. The panel and its scale →
It starts with a decision, not with data
No project here starts from whatever data happens to be available. It starts from a question that today is answered by eye and is expensive to get wrong: which prototype goes to the plant, how far an ingredient can be cut, which segment deserves a product of its own.
Once that question is written down, the next thing is to see how much of the answer is already measured in the category. That is the difference between designing a study and ordering one.
What changes when the category is already measured
A study that starts with the baseline in front of it can concentrate where the uncertainty actually is, instead of spending sessions confirming what the category already has on record. It shows in the size of the design and in how long it takes to answer.
Two pieces of work turn that baseline into a recommendation.
- Preference mappingOne market, or several? The cross between what the panel measures and what consumers answer, with the opposite-taste groups an average leaves hidden.
- Predictive modelsWhat if the product does not exist yet? Estimating the liking of a formula that has not been made yet, and knowing which of the ideas on the table deserve a real sample.
Who works with the history
The Advanced Analytics & AI team sits inside the consultancy, not in a department of its own, and works on the same data the trained panel and the consumer fieldwork produce. That closeness is what avoids the usual failure of data projects in research: elegant models built on variables nobody measured carefully.
Statistics was part of the method from the start, because crossing the two measurements was already a model even if nobody called it one. What has changed is the size of the history and the tools that read it. The cross, in Preference mapping →
Three limits of the history
These are the limits of the history itself. The limits of the model trained on top of it — how far it can extrapolate, and why it has to be checked against real product — are on Predictive models.
The input has to be real
The history is fed by measurements taken with a trained panel and with consumers of the category. Analytics cannot fix a badly taken measurement: it inherits it and passes it on.
There is no history of everything
Coverage is uneven. In some categories the baseline has been building for years and in others it has to be measured first, and we tell you which of the two you are in before proposing anything.
One client’s data is not another’s
The history is only ever used in aggregate and anonymised form, under the confidentiality agreement that comes with every project. What one study adds to its category baseline never resurfaces attached to the client that ran it.
The history is not an archive: it is a baseline. What has been measured is not measured again.
And whatever is missing gets measured before anything is trained on top of it. That order is what separates a number from a decision.
Do you have data sitting idle?
If you have run sensory or consumer studies in recent years, you probably already have half of what is needed. Tell us what you have stored and we will say whether it works as a baseline or whether measuring has to come first.