Estimating liking before you make the product
A predictive liking model is the relationship between the sensory profile of a product and the liking consumers give it. Once that relationship has been measured over enough products in a category, you can estimate the liking of one that does not exist yet.
That turns development into a directed search: instead of making ten prototypes and seeing which one goes down well, you know which way to move and which ones are not worth the trouble.
The first prototypes were measured with consumers; the last two are estimated by the model
Model trained on 24 products · Expected liking, scale 1 to 9 · 90% confidence band
The product and the figures in the chart on this page are fictitious. We never share our clients’ data.
An expected liking, and its margin
The model is not fitted on a product: it is fitted on the map of which attributes move liking in your category, and by how much. That map comes out of an earlier study, preference mapping, and that study is what decides whether there is a model at all.
With the map in place, you only have to describe the profile you want a product to have (creaminess, acidity, aroma intensity, whatever the category happens to have) and three things come back.
- Expected liking
- Not a grade: the estimated liking that profile would get, on the same scale the category was measured on.
- The margin
- It always sits next to the figure and is never tucked away, because it is what tells you whether two prototypes really are different or whether the model cannot tell them apart.
- The valid range
- How far you can move each attribute before you leave what has been measured. It is written into the report, attribute by attribute.
From the map to the model, in four steps
The first three produce a model. The fourth is what turns it into something you can decide on.
The whole category gets measured
Descriptive profiling with a trained panel and liking with real consumers, over the same samples. That cross is the preference map, and it is the starting point: without it there is nothing to model.
Preference mapping →The shape, not the direction
Hardly any attribute behaves in a straight line: many have an optimum in the middle. The model has to reproduce the curve the map found, not a straight line that flattens it.
It is tested against itself
Some of the products are held back, the model is fitted on the rest, and it is then asked to estimate those. A model that only explains the data it was trained on is of no use to anyone.
It is checked against a real product
The prototype the model points to is made and measured with consumers. That is where you see whether the estimate holds, and by how much it is off. Without that check a model is a hypothesis with decimal places, and the decimals give an impression of precision the hypothesis has not yet earned. The check is part of the project, not an extra.
Three decisions taken ahead of the plant
All three are taken before a production run is spent, and none of them can be taken without an estimate in front of you.
Ranking ideas before making them
A list of formulation ideas gets ranked by estimated liking, and only the ones worth the cost of making go to the plant.
What it costs to go down each step
How much liking is left behind at each step of a reduction, before the formula is committed.
Reformulation →The profile a segment would ask for
Which sensory profile it would take to win in a given segment, and whether the formula can reach it.
What comes out of the fit is options. No model signs off a recommendation.
Someone who knows the category reads it and decides what makes sense in your market and what cannot be touched in the formula. The model narrows the ground; choosing is still a matter of judgement.
Two conditions we say up front
A liking model is a tool with conditions of use. These are the two that have wrecked the most projects when nobody raised them in time.
It is only valid inside the range measured. The model knows the category it was shown. Ask it for the liking of a flavour nobody on that panel has ever tasted and it does not estimate it: it invents it. Outside the range measured, the figure means nothing.
It needs enough different products. Four similar references give it no variation to learn from. You need products that separate from one another on the attributes that matter, and that sometimes means measuring the competition before modelling anything.
If your category does not meet the conditions, we tell you and we do not fit the model. What is usually missing is measurement, not mathematics: first the set of measured products is widened, then the modelling happens.
Do you already have enough data for a model?
Tell us what you have measured so far and on how many products. We will tell you whether it is enough to fit a liking model, what is missing if it is not, and what you will be able to decide with it in front of you.