The pour-igin
of species
The pour-igin
of species
You’re in a rush and you need to grab a bottle of wine for a special occasion. You’ve got $40 and no preference for red or white. Which wine are you buying?
Tap on a bottle or start scrolling to get a random wine
Step 2
We collected data for 9,300 wines that appeared on the popular app Vivino and used Chat-GPT Vision to help us categorize the animals on the labels with this prompt: This image is a wine label. Do you see any animals or humans on it? Read more in the methodology.
Sometimes it was spot on. It correctly identified all of these labels as birds even though the artistic styles were different.
But other times, we had to manually review and edit.
GPT was particularly bad with elephants. It also struggled when animals weren’t the focal point and with what counted as an “animal.” Sometimes anthropomorphic or mythical creatures were included, sometimes not.
After cleaning up the data we grouped the animals into broad categories based more on common perception than binomial nomenclature (genus and species) — sorry science purists — and omitted categories with fewer than 20 animals, leaving us with these 16 groupings. Here’s the animal on the wine you picked.
To get a general understanding of the wines, we found the median price of each animal group.
Then we looked at the rating.
But that only gets us so far. What we really care about are the best deals, wines whose price is below the median and whose rating is above the median. This is…
The median price and ratings show us something important: what we all assume expense and quality look like.
As we’ll see, there are patterns and exceptions. But what became most clear over the course of our study was that the question was not simply about predicting price: it was about predicting value.
If your budget is $10, your budget is $10 — it doesn’t matter if you know which animals are expensive. A much more useful thing to know is which animals are undervalued, the ones that tend to stay cheap, even as quality increases.
For that, let’s dive deeper into the data.
More expensive →
Instead of looking at price and rating independently, let’s look at them together using a scatterplot. Price is on the y-axis here, so the more expensive the wine, the closer to the top. Rating is on the x-axis and the higher quality wines are toward the right.
The animal group on the wine you selected is highlighted in yellow.
Overall, the median price for wines with animals on the label ($26.99) was $3 cheaper than wines without animals ($29.99), but both groups had a median rating of 4 stars.
This means that — above all else — you’re better off buying any animal than no animal at all.
Among animal wines, farm animals like cattle, pigs, and sheep have both lower ratings and lower prices.
While bottles featuring animals commonly found on heraldic crests or coats of arms — like cats, bears, and mythical creatures (oh my!) — have higher prices and higher ratings.
Let’s zoom out, with $150 as the maximum price. We figured that if you’re spending more than $150 on a bottle of wine, you’re probably not solely choosing based on the label anyway.
Now let’s explode the chart so we can see all 1,409 animal wines we identified under $150. Here’s the bottle you selected.

Sauvignon Blanc
Tomtit
bird
White
The wines follow a logarithmic trend, where rating increases with price, but with diminishing returns.
To find the best deals — the wines at or below the median price and at or above the median rating — we’re going to divide the scatterplot into 4 quadrants.
What we really want to look out for are wines in the bottom right corner — these are the best deals.
The bottle you picked with a bird on it is in this quadrant. Nice intuition!
So, can the animal on the label help us buy a good cheap wine?
Here are all the amphibians and reptiles. 5.3% of them are good deals. You can’t even really see the good deals on the chart because no amphibian and reptile wines below the median price appear beyond the 4 star median rating.
Here are the pigs, with a score of 10.7%. Good deal pig wines, like amphibians and reptiles, are a little hard to see on the chart because most of them fall on the median rating line. There is one outlier though: Fat Pig Cape Vintage, priced at $23.45 with a 4.3 star rating.
And here are the cats. 9.6% are good deals. Cat wines are much easier to see on the chart because there are a lot more of them, but there’s still a relatively small percentage of good deals. That’s because their median price ($38.43) is nearly $9 more than the overall median ($29.99).
Last up, the birds. 16.5% are good deals. Like cat wines, there are just more of them, but because their median price ($25.99) is $4 less than the overall median ($29.99), we see a higher percentage of good deals.
Birds had the highest percentage of good deals that we’ve seen so far, but let’s go back to our bottle lineup to see which animal group comes out on top.
Bottom line: bet on fish. 24.2% of them are in the magical good deals quadrant.
But, this is just our definition of a deal. What’s yours?
and filters to set your own limits. And keep scrolling to see an animal-by-animal breakdown.
Now you know how to spot the best deals, what if you don’t care about price and quality and just want to ball out with your favorite animal?
Use the left and right arrows to navigate through all 16 animal groups.
In the end, one bottle had to emerge victorious. When looking at the entire scatterplot, and focussing on the bottom-right corner, a few bottles stand out as great deals. Below $20, the highest-ranked wines scored 4.3 stars, but these were all dessert or sparkling wines.
For the best bet on a standard, dinner-ready, red or white wine, the crown goes to Mount Fishtail’s Sauvignon Blanc from New Zealand’s Marlborough region.
New Zealand wines are delicious and, ultimately, a better predictor of quality than any of the animals we looked at. Across all categories and variables, the only dataset that maintained its price as the quality increased was New Zealand wines. When in doubt, buy Kiwi.
After over 9,000 bottles and years of searching for the perfect animal wine, it turned out it was being grown on a vineyard at the seat of Mount Fishtail, tucked away in the Marlborough Ranges, a short jaunt across the Cook Strait from Wellington, which is where I wrote this article.
Cheers to that coincidence, and cheers to Mount Fishtail.
Methodology
Initial test data was manually collected across several New Zealand supermarkets. With hypothesis in hand, we expanded to programmatic collection.
To get our initial list of wines, we used the vivino-api package, querying with different types of filters similar to what a user sees on Vivino’s website or app and excluding any duplicates. The filters included things like wine types (Red, White, Rosé, etc.), price range ($0–$10000+), and average rating (0–5 stars). For each wine, we collected the following: Vivino ID, name, year, winery, country, region, type, rating, number of ratings, price, currency, and a url with the image of the label. All details were collected in March 2024 and may not reflect current price or rating, especially if those have changed dramatically.
To avoid over-prioritizing wines whose brands included animals, we limited each brand to one wine per type (Dessert, Fortified, Red, Rosé, Sparkling, White), choosing the wine with the most ratings as a proxy for which wine would be more well known. This gave us a list of 9,314 wines to download images for.
We then fed the label images into the OpenAI API (gpt-4-vision-preview model). We tested several prompts including open-end questions like What’s on this image?, but ultimately the prompt that worked the best was This image is a wine label. Do you see any animals or humans on it? On a scale of 0-1, with 0 being 'not certain at all' and 1 being 'very certain', how sure are you?
This provided us with responses that looked something like this: I do not see any animals or humans on the wine label. The label contains text and a small emblem at the top, but there are no discernible figures of humans or animals. I am very certain about this assessment; so on the scale from 0 to 1, my confidence level is 1. Occasionally, GPT would respond with something like I'm sorry, I can't assist with these requests. In those cases we reran the prompt or manually reviewed the label. We ended up abandoning GPT’s certainty prediction because it wildly swung toward the poles: only 0.6% of the wines had confidence scores of something other than 0 or 1.
We then manually reviewed a subset of wines to check GPT’s accuracy. GPT was particularly bad with pachyderms. It also struggled when animals weren’t the focal point and with what counted as an “animal.” Sometimes anthropomorphic or mythical creatures and insects were included, sometimes not. Knowing these shortcomings, we paid more attention to manually reviewing these types of labels. There still may be a few labels that slipped through the cracks, but overall this should be a solid sample.
We ended up identifying 1,488 animal wines, or about one sixth of all wines in our initial gathering. We then manually grouped them into larger categories. Some animals rolled up, others did not. For example:
Were rolled up
- lion big cat cat
- hawk raptor bird
- ant walking insect bug
Stayed the same
- horse horse horse
- bear bear bear
- fish fish fish
This opened up a lot of existential questions about what constitutes an animal, like “Is a duck a bird?” or “Is a zebra a horse?” We tried to strike a balance between common perception and scientific naming with binomial nomenclature (genus and species), often going back to the question “What would a kid call this animal?”
We limited our analysis to animal groupings with at least 20 wines, knocking out things like bats, monkeys, rodents, and marsupials. Our apologies to Yellow Tail, the lone kangaroo label in our dataset.
To calculate the percentage of wines in each animal group that were good deals, we found the percentage of wines at or below the overall median price ($29.99) and at or above the median rating (4 stars) for each animal group.
For the trend line, we calculated the best fit regression (logarithmic) and used Harry Stevens’ d3-regression package to draw the line.
We also bucketed the wine by price, rating, type and country to compare each animal groups’ distribution to all wines. We then calculated a Z-score for each bucket to find animal wines with statistically more or statistically less wines per bucket than all wines.
To draw the 3D wine bottles we used Adobe Illustrator and Adobe Dimension, roughly following this tutorial from Silver Moon Design School. We exported 8 different views of each bottle from Dimension and combined them in an image sprite for the 360° rotation.