Drunk on Data: Finding the Optimal Recipe for a Manhattan Cocktail

To show you Gastrograph’s optimization system in action, we challenged ourselves to improve the average Manhattan cocktail and come up with the perfect recipe. Cheers!

As someone involved in product development, you’re under more pressure than ever to innovate so you can compete with the new CPG brands that are appearing on shelves all the time. But it’s hard to take risks or make any kind of decision without some backup.

Should you focus on launching entirely new products? Altering existing recipes? If so, how do you know what you need to change to positively impact liking?

With thousands of dollars at stake, you want data on your side to justify your product choices. But conventional product development methods don’t give you that data. They’re leaving you in the dark.

Enter Gastrograph AI. Our system gives you a reliable, fast, and cost-effective way of adjusting food and beverage products to increase likeability. We give you the data you need to make smart product decisions.

To show you Gastrograph’s optimization system in action, we challenged ourselves to improve the average Manhattan cocktail and come up with the perfect recipe. Cheers!

How We Did It

The process we went through with the Manhattan cocktail matches how we typically approach product optimization. Our system is built on primary sensory data from in-person tastings — we’ve got the world’s largest sensory database, and we’re adding to it all the time. We start by running an analysis of the initial product (or in this case, cocktail) to get a baseline. Then we set the system to find the best formulation for a target audience.

Collected Sensory Data

We conducted tastings of several different formulations of the cocktail and used the data to train our AI system to understand how the balance of each ingredient impacts people’s preferences.

For our study, we had to define an ‘average’ Manhattan cocktail. We took the standard recipe of 2:1 whiskey to vermouth plus a dash of bitters and used brands that are popular for Manhattans.

The average Manhattan recipe, according to Gastrograph:

  • 60g Wild Turkey 101 Rye Whiskey (66.39% of the total drink)
  • 30g Martini & Rossi Rosso Vermouth (33.20% of the total drink)
  • 0.375g Angostura Aromatic Bitters (0.41% of the total drink)

To build the training models, we conducted 194 sensory reviews (with real human tasters) of 23 different formulations. The 23 formulations covered the most extreme ratios and meant we could train the system to understand the range of possible recipes. For example, 100% whiskey with 0% vermouth to 0% whiskey with 100% vermouth.


Mass A is the whiskey, Mass B is the vermouth, and Mass C is the bitters

We got tasters to fill out a Gastrograph-style flavor profile, including intensity scores for each flavor and an overall Perceived Quality score out of seven. The sensory reviews gave us all the data we needed to train the AI to understand the impact the different ingredients have on preference.

Analyzed the Initial Product

We ran the Gastrograph AI system to predict how consumers would respond to the average recipe — defining which flavor attributes they would perceive and at what intensity.


The flavor profile our system generated for our average Manhattan recipe

Gastrograph’s objective flavor profiles always show the same 24 flavor attributes that cover the gustatory flavor space, so we’re able to directly compare different products or versions of products. Green indicates the flavor had a positive impact on preference, red indicates a negative impact, and black indicates neutral.

The blue areas show which flavor attributes people will perceive when they taste the product — the further from the center the blue extends, the more intensely they’ll perceive the flavor.

For example, for the average Manhattan recipe, the profile shows that people will perceive 'wet' flavors with three out of five intensity and 'rich' flavors with two out of five intensity.

Next, we ran the system to forecast consumers’ preferences for the average recipe. The result was an average of 3.7 out of 7.

'Hedonic Score' is the preference score. 3.7 out of 7 ... definitely room for improvement!

Found the Optimal Recipe

We set the Gastrograph system to find the optimal formula for a Manhattan cocktail — the recipe it came up with would get higher PQ scores than the average Manhattan.

Our goal was to improve the cocktail for people who already enjoyed it. So we set the system to target those who showed high preference in our analysis of the average Manhattan.

If we went through the same process in real life, we'd have to get our target audience to taste hundreds of versions of the cocktail — all fractionally different. Though that might sound fun for Manhattan drinkers, it would be extremely time consuming, not to mention that our tasters would get tired (...and drunk).

Instead, because we had already trained our system to understand how consumers taste, it could run through different formulations of the cocktail and accurately predict how they would perceive them and how much they'd like them. Our system found the best recipe in a shorter time than it'd take you to hold in-person tastings, no hangovers necessary.

The optimized recipe:

  • 67.83% whiskey
  • 32.16% vermouth
  • 1.38% bitters

That's a slightly increased percentage of whiskey and bitters and less vermouth compared with our initial recipe.

The flavor profiles show the impact of the new recipe on how people will perceive the flavors. For example, people will perceive more 'sugar' flavors in our new version than in the average cocktail.


The perceived intensity of ‘dry’ flavors also increases in the optimized version

The Gastrograph system predicted preference for the new recipe: a PQ score of 4.5/7.


The red line shows the average recipe, and the blue shows the optimized recipes.


The analysis of the new recipe shows an increase in the percentage of the population who would give a high PQ score: 50% of people would give the new formulation 5/7.

It also predicts a lower percentage of people would give 4/7 compared with the average recipe — that means we’ve switched people who liked the cocktail “a bit” (4/7) to liking it “a lot” (5/7).

What Optimization Means For You

Our Manhattan cocktail case study is an example of the product recipe optimization that's possible with Gastrograph AI. Companies can apply this process in several ways to improve their product development and innovation.

1: Perfect Existing Products

Identify which flavor attributes make customers enjoy your product and then tweak your recipe to increase that preference. Get people who like your product to fall in love with it.

2: Target New Demographics

Our sensory database covers more than 30 countries around the world. We're able to analyze your product and then predict what the preference for it would be in different locations.

We've helped companies take products that were popular in one location and alter their recipe so they could relaunch on a whole new continent.

3: Create and Innovate

With Gastrograph, you’re able to learn which flavor combinations different demographics enjoy. You can create entirely new products that target their specific preferences.

Product Development Doesn’t Have to Be Stressful

Gastrograph AI can transform your product development. We’ll help you take your product decisions from risky, stressful nightmares to data-led creative adventures. Request a demo.

Similar posts

Stay up to date with Gastrograph AI

Be the first to know about company updates and industry-related news, straight from our internal subject matter experts.