Structural discovery of combinations
JEM finds your best kits, bundles and baskets while testing far less. Not a guess: it's proven mathematics — the computation is exact and checked on every run.
In plain words
Imagine you run a store and want to sell bundles — say "grab banana + strawberry + granola". The problem: there are hundreds of possible combinations, and finding out which one sells best means testing with real customers, which costs money.
We analyzed 3 million real purchases from a supermarket and found two things:
Some items go together (banana and avocado sell together all the time) and some never do (regular banana and organic banana — the customer picks one or the other). The tools companies use today score each product on its own and see none of this.
Every test of one bundle teaches you something about all the similar ones. The result: you can find the winning bundles with a fraction of the testing — in our simulation on this real data, about 8× less.
Important: this supermarket analysis is just one example of the application — the same method works for any catalog (delivery combos, menus, ad sets). And it's not guesswork: the math behind it is grounded and proven. A theorem guarantees it extracts the most signal possible from the data — for the kind of reconstruction it does, it's the best possible; there's no better version.
Less testing = less money burned on bad bundles and an answer fast enough to reach Black Friday with the right bundle. That's what we're selling to stores.
The problem
A bundle's value is in the combination, not the loose items. But the number of combinations grows far too fast to test one by one on real traffic.
Every session spent on a losing bundle is money you don't get back. And picking "by gut" usually leaves the champion out.
How it works
Instead of treating each bundle as an isolated test, JEM uses the structure of the combinations to get the most out of every observation.
Overlapping bundles are connected. One observation on a bundle already informs every other bundle that shares products — no starting from scratch each test.
How much comes from each item, each pair (the synergy) and each trio. The champions almost always live in the pair.
The computation is provably exact — checked to machine precision on every run. No training, no GPU, no manual tuning: it runs in seconds over a spreadsheet.
What sets it apart
Models that score product by product stall before the best bundles — they're blind to exactly the synergy between items. JEM sees precisely the part they miss.
In the market baskets we measured, the biggest share of what makes a good bundle sits in the synergy layer — the band ordinary models can't represent.
Results
Simulation and diagnostic figures, measured on real market data.
Where it fits
Rule of thumb: it pays off where the decision is about a set — and testing burns traffic or budget.
Holidays, sales, Black Friday.
Build the combo customers actually want together.
Which trio of ads works best together.
Assortment for pages and digital shelves.
Pricing bundles and combined plans.
Replenishment and repurchase together.
How it starts
We measure, on your historical data, how much synergy exists and how much your current process leaves on the table. It ends in a go/no-go.
If there's no meaningful synergy, we'll tell you — and you don't go further. No integration and no personal data: a simple history export is enough.
Fill this in and we compose a ready-to-send email — a 30-minute chat, no commitment.