The other day I was idly wondering what are the strangest combinations of items people buy at grocery stores. The kind of shopping cart that makes the cashier snicker and later tell his friends, “Dude, can you believe this guy came in and only bought condoms and apples?”
So I fired up Claude and started looking for any receipt data I could find.
For the final results, scroll to the bottom. Otherwise, read on to follow the journey.
Grocery stores keep this kind of data very close to the chest. There are consumer apps that collect receipt data (like
ReceiptHog
and
Fetch
) but they presumably just sell it to hedge funds or something. Years ago, however, Instacart
open-sourced
data from 3 million orders as part of a machine learning competition to optimize their recommendation algorithm. The data is still available on
Kaggle
and it’s very rich. What we want is the exact
opposite
of a recommendation algorithm but this data should work fine.
The Instacart data set includes the following:
3,214,874
orders
~10
products per order on average
49,688
unique products
134
unique “aisles” (product categories)
So all we have to do is look at every cart and see which combinations of items are least likely to appear, right? Let’s try that. For the sake of testing a few groupings, I included pairs, triples, and quads.
Product co-occurrence count: pairs
count
product 1
product 2
1
Banana
Extra Fancy Unsalted Mixed Nuts
1
Organic Hass Avocado
Baby Cucumbers
1
Organic Hass Avocado
Trail Mix
1
Bag of Organic Bananas
Bananas
1
Limes
Trail Mix
1
Organic Garnet Sweet Potato (Yam)
Clementines
1
Organic Avocado
Zero Calorie Cola
1
Organic Fuji Apple
Baby Cucumbers
Note: maximum pair co-occurrence count is 62,341.
Product co-occurrence count: triples
count
product 1
product 2
product 3
1
Banana
Bag of Organic Bananas
Soda
1
Banana
Bag of Organic Bananas
Clementines
1
Banana
Bag of Organic Bananas
Sparkling Mineral Water
1
Banana
Bag of Organic Bananas
Flat Parsley, Bunch
1
Banana
Bag of Organic Bananas
Sparkling Water
1
Banana
Bag of Organic Bananas
Fresh CA Grown Eggs
1
Banana
Bag of Organic Bananas
Red Seedless Grapes
Note: maximum triple co-occurrence count is 15,066.
Product co-occurrence count: quads
count
product 1
product 2
product 3
product 4
1
Banana
Bag of Organic Bananas
Organic Baby Spinach
Half & Half
1
Banana
Bag of Organic Bananas
Organic Whole Milk
Organic Zucchini
1
Banana
Bag of Organic Bananas
Organic Avocado
Organic Half & Half
1
Banana
Bag of Organic Bananas
Strawberries
Organic Lemon
1
Banana
Bag of Organic Bananas
Organic Hass Avocado
Spring Water
1
Banana
Bag of Organic Bananas
Organic Yellow Onion
Organic Fuji Apple
Note: maximum quad co-occurrence count is 3,828.
Hmm. These aren’t very interesting. Maybe that’s because there are
a
ton
of combinations that only occur once, and we’re only looking at the top 5-10, so we need a better way to sort them.
There are roughly 1.2 billion possible unique pairs of products (unordered pairs from a set of 50,000). About
97% of them never occur in our data
, and of the ones that do, around
22 million pairs
occur exactly once. When almost every pair is tied at zero or one, “least common” becomes meaningless.
So how should we sort? Claude had a good idea to rank each combination by “lift”: how frequently it
actually
appears divided by how frequently we’d
expect
it to appear. If you have two very common items like apples and oranges, you’d expect them to co-occur a lot. If they don’t, that’s notable, and the pair should rank higher. Now let’s try that.
Product co-occurrence by lift: pairs
lift
product 1
product 2
0.0007
Banana
Extra Fancy Unsalted Mixed Nuts
0.0010
Organic Hass Avocado
Baby Cucumbers
0.0012
Organic Hass Avocado
Trail Mix
0.0017
Bag of Organic Bananas
Bananas
0.0019
Limes
Trail Mix
0.0020
Organic Garnet Sweet Potato (Yam)
Clementines
0.0021
Organic Avocado
Zero Calorie Cola
0.0025
Organic Fuji Apple
Baby Cucumbers
Product co-occurrence by lift: triples
lift
product 1
product 2
product 3
0.0016
Banana
Bag of Organic Bananas
Soda
0.0019
Banana
Bag of Organic Bananas
Clementines
0.0032
Banana
Bag of Organic Bananas
Sparkling Mineral Water
0.0043
Banana
Bag of Organic Bananas
Flat Parsley, Bunch
0.0047
Banana
Bag of Organic Bananas
Sparkling Water
0.0056
Banana
Bag of Organic Bananas
Fresh CA Grown Eggs
Product co-occurrence by lift: quads
lift
product 1
product 2
product 3
product 4
0.0111
Banana
Bag of Organic Bananas
Organic Baby Spinach
Half & Half
0.0128
Banana
Bag of Organic Bananas
Organic Whole Milk
Organic Zucchini
0.0137
Banana
Bag of Organic Bananas
Organic Avocado
Organic Half & Half
0.0148
Banana
Bag of Organic Bananas
Strawberries
Organic Lemon
0.0155
Banana
Bag of Organic Bananas
Organic Hass Avocado
Spring Water
I think I see a pattern… Of course it’s unlikely that someone would buy both organic and non-organic bananas which points to the bigger issue that
50,000 products is too granular.
Every product has a dozen+ variants. There are
28 different condom products
and
110 different apple products in the produce department alone
, so our silly “condoms + apples” example will struggle to rise above the noise.
We have 49,688 products (too narrow) sorted into 134 aisles (too broad). We need something in between. Time to build a classifier! But how?
Enter GPC
It turns out nearly 100% of grocery retailers use a product classification system called
GS1 Global Product Classification (GPC)
.
GPC
is an ontology of everything you might find at a big-box or a grocer. It divides the universe of products into
segments
(camping, footwear, lighting, tools, food…), which contain
families
(bread, beverages, fruits, meat, vegetables…), which contain
classes
(beans, melons, peppers, fungi…), which contain
bricks
(chanterelles, enokitake, truffles, morels…).
Some bricks go even deeper into subtypes or attributes, e.g. Morels can have a growing method of
CONVENTIONAL
or
ORGANIC
, but for our purposes we can stop at the brick level. There are
5,318 bricks
in the full GPC system. Many segments (plumbing, live animals, industrial systems, etc.) aren’t sold at grocery stores, so we can drop them. That leaves
9 grocery-relevant segments with 1,697 total bricks
, and our products only populate about
1,000
of them. This feels like the right level of granularity and yields a
50x
reduction in complexity!
Time to fire up the ol’ classifier. Since we’re squeezing ~50,000 products into ~1,000 bricks, I used a standard two-stage approach:
Shortlist the brick options for each product.
First, get embeddings for every product and every brick. I did this locally using
qwen3-embedding:8b
. Then, for each product, find its 10 nearest bricks by cosine similarity.
Choose the best brick.
Use an LLM to select the best-fit brick for each product from the 10 options. This cost me ~$5 and took a few minutes, running 60 requests in parallel using
gpt-4.1-mini
. The full product-to-brick mapping is
here
in case it’s useful to anyone else.
Now let’s run our co-occurrence algorithm again and see if using bricks rather than products yields better results.
Brick co-occurrence by lift: pairs
lift
brick 1
brick 2
0.004
Wine – Still
Grains/Cereal – Not Ready to Eat (Frozen)
0.007
Spirits
Grains/Cereal – Not Ready to Eat (Frozen)
0.012
Beer
Baby Leaves
0.013
Spirits
Dates
0.019
Wine – Still
Stem Vegetables (Fresh Cut)
0.020
Spirits
Fish – Unprepared/Unprocessed
0.022
Beer
Fish – Unprepared/Unprocessed (Shelf Stable)
Brick co-occurrence by lift: triples
lift
brick 1
brick 2
brick 3
0.014
Cheese
Wine – Still
Frozen Grains/Cereal
0.019
Milk
Savoury Grain Meals (Shelf Stable)
Spirits
0.019
Cucumbers
Baby/Infant Specialised Foods
Spirits
0.020
Kale
Clementines
Wine – Still
0.021
Broccoli
Baby/Infant Specialised Foods
Spirits
Brick co-occurrence by lift: quads
lift
brick 1
brick 2
brick 3
brick 4
0.027
Banana
Strawberries
Garlic
Spirits
0.028
Milk Substitutes
Strawberries
Carrots
Spirits
0.032
Banana
Milk Substitutes
Baby Food
Spirits
0.035
Banana
Milk
Wine – Still
Savoury Grain Meals
Far less repetition! But there’s still something missing. Just because these are the rarest combinations doesn’t make them
interesting
. For some reason people don’t buy wine and frozen rice together, or beer and fish. So what? How do we make these combos spicier? Funnier?
We need a
humor index
for products.
What’s funny?
Claude dutifully scored all ~1,000 grocery bricks with a 0–1 humor score, based roughly on how taboo they are or how likely they’d be to come up in a stand-up comedy routine. The general ranges:
score
# of bricks
examples
0.0 – 0.1
726
Milk, Armenian Cucumber, Herbs, Prepared Fish
0.1 – 0.3
271
Confectionery, Cooking Wines, Breath Fresheners, Ear Care
0.3 – 0.5
50
Energy Drinks, Toilet Paper, Denture Care, Diet Aids
0.5 – 0.7
12
Wart/Corn Treatments, Foot Care, Medical Lubricants, Appetite Control
0.7 – 1.0
15
Condoms, Intimate Lubricants, Contraception, Enemas/Douches
(Humor scores for all bricks are
here
)
Now we can multiply each combo’s aggregate humor into the ranking alongside its rarity, and we should get much better results. Let’s try that.
Funniest rare combos: pairs
brick 1
brick 2
Garlic
Diarrhoea Remedies
Kale
Enemas/Douches
Garlic
Enemas/Douches
Flat Parsley
Condoms
Baby Food
Enemas/Douches
Coriander
Enemas/Douches
Baby Food
Condoms
Baby Food
Adult Diapers
Funniest rare combos: triples
brick 1
brick 2
brick 3
Cheese
Almond Milk
Intimate Lubricants
Milk
Almond Milk
Intimate Lubricants
Apples
Avocados
Intimate Lubricants
Onions
Raspberries
Antacids/Flatulence Remedies
Apples
Tinned Vegetables
Condoms
Milk
Almond Milk
Enemas/Douches
Funniest rare combos: quads
brick 1
brick 2
brick 3
brick 4
Milk
Almond Milk
Tomatoes
Antacids/Flatulence Remedies
Apples
Avocados
Ice Cream
Antacids/Flatulence Remedies
Cheese
Milk
Apples
Condoms
Bananas
Apples
Kale
Antacids/Flatulence Remedies
Bananas
Yogurt
Milk
Enemas/Douches
Much better! Some of these even made me laugh out loud. Kale and an enema? Parsley and condoms? Adult diapers and baby food? There’s gotta be a joke in there somewhere.
That said, in a large enough shopping cart, funny combos become more likely just by chance. If you’re ordering all your groceries for the week you’re going to pick up everything, from the boring to the spicier. So what happens if we look
only at small carts
, where the entire order is just 2 or 3 items?
Funniest small carts: 2 items
item 1
item 2
Vitamin D Milk
Ultra Thin Condoms
Cola
Ultra Thin Condoms
Bag of Organic Bananas
Anti-Diarrheal
Italian Bread
Ultra Thin Condoms
Oreo Chocolate Sandwich Cookies
Personal Lubricant
Bag of Organic Bananas
Incontinence Underwear
Funniest small carts: 3 items
item 1
item 2
item 3
String Cheese
Potato Chips
Omeprazole Acid Reducer
Black Cherry Yogurt
Hass Avocado
Omeprazole Acid Reducer
String Cheese
Black Cherry Yogurt
Personal Lubricant
String Cheese
Italian Bread
Personal Lubricant
Vitamin D Milk
Italian Bread
Anti-Diarrheal
Bag of Organic Bananas
Cola
Personal Lubricant
Small orders are less common but we still got some fun ones. Oreos and lube? Sounds like a good time!
