Skip to content 
The challenge
The client understood its customers in broad terms, but not at the level of the product.
Then there was the size of it. Twenty thousand products, organised into eighty product areas, across seven countries, all needing to be looked at in enough depth to defend a portfolio decision. Done by hand, the work would have taken years, as the client's previous programmes of similar ambition had.
They suspected around a tenth of the range could be rationalised, but were not willing to do it by cutting the lowest sellers. They needed a basis for the decision grounded in customer need, and something they could keep using as the range changed rather than a report that would be out of date within the year.
What we did
The brief came from strategy and product rather than from the insight team: they wanted something they could use directly to make range and investment decisions. Three steps got them there, with human sign-off on everything the AI produced.
1. We built the questionnaire once and adapted it eighty times.
You cannot ask the same questions about coffee tables as about office chairs. We wrote the template for a handful of product areas, and Nucleus generated the choice drivers for the rest, with every one checked by the team. That part took days instead of weeks.
Sales data alone
Need at product level
One questionnaire for everything
Eighty, one per product area
Years of manual work
Five months, and repeatable in-house
What went into the model
200,000 survey responses
Seven markets, with a questionnaire adapted to every product area
Eighty segmentations
One for each product area, rather than a single view of the whole business
20,000 product descriptions
Read by language models to place each product against the groups in its area
Product images
Computer vision picking up what the descriptions leave out
The client's sales data
Brought into the same model, at the level of the individual product
2. Then we tested it on modelled customers.
Before anything went to field, we used Nucleus to generate modelled customers and had them answer the draft questionnaire. That did two things: it showed us where the questionnaire was going wrong, and it gave the strategy team something concrete to react to months before real results arrived.
Two questions turned out to be phrased in a way that led the model into overstating affinity, so we adjusted them and ran it again. None of this went into the final results: it was a way of checking the questionnaire before we spent money putting it in field.
3. Then we ran a segmentation for every product area and scored the products against it.
Eighty product areas meant eighty segmentations rather than one. Working from the survey data, Nucleus produced more than ten thousand first-draft segment profiles for analysts to work through.
We then took the finished profiles and combined them with the client’s twenty thousand product descriptions and product images, using language models and computer vision to give each product a score between 0 and 1 against each customer group in its area.
01
Choice drivers
Generated for the product areas researchers had not written by hand, then checked one by one.
02
Modelled customers
A draft questionnaire answered before fieldwork, to find what was phrased wrong.
03
Profiles and scores
First-draft segment profiles, then a 0 to 1 score for every product against every group in its area.
What the client walked away with
Each was sized for addressable market and for how widespread the unmet need was, so the shortlist came with a commercial case as well as consumer evidence.
01
Scores in their own BI platform
Teams in every market interrogate them directly, rather than waiting for a central team to cut the data.
02
A repeatable method
As the range changes they run the scoring themselves, instead of commissioning new fieldwork.
03
Better-informed range decisions
Rationalisation grounded in customer need, rather than in cutting the lowest sellers.
The outcome
The scores went into the client's own BI platform, so teams in every market can interrogate them directly rather than waiting for a central team to cut the data for them.
Products
20,000
scored 0 to 1 against every group in their area
Low against every group in its area marks a rationalisation candidate. High scores with strong sales mark where the money goes.
Customers
200,000
surveyed across seven markets
Eighty segmentations, one for every product area, rather than a single view of the business.
Delivery
5 months
not the years a manual equivalent would take
The client's previous programmes of similar ambition had run for years.
Handover
In-house
scoring the client runs itself
They have a repeatable method, so as the range changes there is no new fieldwork to commission.
The conclusion
Customer need varied between markets far more than anyone expected, enough that the same product can be a rationalisation candidate in one country and a core line in another. That is not something sales data shows you. The programme also brought product data, sales data and survey data together at the level of the individual product, making this depth of analysis possible across 20,000 products.
We needed decision-making metrics at a scale only your Nucleus powered approach could deliver.
Insight and Product Manager
Global home furnishings retailer
