The Dish Dragon is a tool to improve home cooking using the scientific method and large amounts of data. We’ve indexed 500,000 different recipes from sites like Serious Eats, Allrecipes and Food.com, and use that information to advice on improving your dishes.
We’re not a recipe site. We are not answering the question «How do I make Spaghetti Bolognese?». We are answering the question «How can I make my Spaghetti Bolognese better?».
Applying the scientific method to cooking is, of course, a tried and tested formula for success. We love the books of Harold McGee and the systematic changing of variables of Kenji Lopez-Alt’s Food Lab. However, there is one weakness with most approaches: they use their own taste as the measuring stick. Or at best the taste of their colleagues (and neighbours, we hear).
That wouldn’t really pass muster in science. And food is also notoriously cultural and individual, so a few people may not be a good indicator. They're not a representative sample. So to solve that, we use online reviews to investigate how large amounts of people like food combinations across different recipes.
Typically, any combination of two ingredients we mention will have at least 900 different reviews behind it!
Oh boy, do we know all about that! Yes, there’s a lot of issues and most of our work has gone into finding the signal in the noise.
The largest problem we have is that a lot of recipe sites inflate their reviews. Some recipe sites have perfect 5 star review scores for every single one of their thousands of recipes! Some of those might be fraudulent, and some might just have a small audience with their mum and dad as the main reviewers.
The other large problem is the more familiar one, that we don’t really know what the reviewer actually cooked, and whether or not they substituted every second ingredient. In fact, we don't know if they even cooked the dish at all.
Yes! We do a transformation of the data that goes a long way to solve this. We normalise the review scores on a site-to-site basis before we use it in our formula, and we use the deviation from the average site score to estimate the impact.
For example, a recipe with an average revew score of 5 stars on a site where every recipe has 5 stars will just be counted as «average» and have marginal impact on the final score. A recipe with a 4.7 star review from a site with average 4.3 star reviews and a lot of variation will have a more positive impact on the final score.
It solves two things: It defuses the inflated review scores and put them on an equal footing with the most honest sites. A small drawback is that they don’t add as much information as they could have, but at least they don’t distort the numbers.
The other major thing is that since we’re looking at how much better or worse a review score is than the normal, hopefully we reduce the impact of unreliable reviews. Yes, people might substitute olive oil for egg whites, but if we assume people behave similarly irrational across all recipes on the same site, we should at least be able to see what tastes relatively better.
In addition, we remove recipes that either trigger some indications that they are not reliable or have too low numbers to be useful. We actually end up just keeping the 20% most reliable recipes for the final review score calculations!
You can currently look up either single ingredients to see how other ingredients combine, or you can look up dishes and see which ingredients are common and gets the dish good or bad review scores. For each ingredient we show the baseline (the overall review score), the review score with the other ingredients in the mix, and for those interested, the 95% confidence interval around the extra ingredient.
For all ingredients where this confidence interval does not overlap with the 95% confidence interval of the baseline ingredient, we call the ingredients out as «Stand out combinations». These are the ones you should really look out for and consider if you should add to your dish!
But also have a look at the ones that are not so clear cut (the ingredients that improve another and so on), and definitively the poor combination when there are some. «Worse all around» means the same as for «stand out combinations», but the opposite!
Then, if you see some combinations that are interesting, if you click through you’ll get links to the recipes they were used in.
That happens! And it wouldn’t even be on here if people didn’t use it in recipes. So clearly there must be something?
In those cases, and if you want to use a particular ingredient combination, consider *why* this might get a bad score? Some things to consider is if it’s an ingredient that differs a lot between brands? Does the amounts matter a lot and are easy to get wrong? Does the preparation method make a big difference? This should be a moment to take pause and pay extra attention when you use a particular ingredient.
Sure, but we decided to show you the numbers rather than package it up. If it’s not useful, just look at the headlines and ingredient names.
Use shallots for onions, cook it with a parmesan rind and do not use sage. Also, be very careful if you’re adding chicken livers.
Well… you don’t really need to sign up. But if you want to keep up with us and communicate with us, just follow us on Twitter!
The same, just let us know on Twitter, or send us an email on mail@dishdragon.ai