Karah and Julian are every day New Yorkers investigating the hidden uses of artificial intelligence affecting our everyday lives. They challenged us to use our platform to craft a new flavor of sparkling water unlike anything they’ve ever tasted and that they’d prefer to any of the other product on the market. To do so, we had to train the AI on their perceptions, learn their preferences, and run a rapid development cycle – before testing our results live on their show. Read on to see how we developed a flavored sparkling water better than LaCroix for two unique New Yorkers
Gastrograph AI is an ever-learning platform for predicting human sensory perception of flavor, aroma, and texture to predict consumer preference of food and beverage products. Consumer and professional sensory “reviews” are collected on Gastrograph Review, a sensory system for collecting high resolution data across any food and beverage product. The AI is trained on data collected through large-scale R&D undertaken by Analytical Flavor Systems (AFS) and proprietary data from each client on the platform to build targeted models for new product development and existing product optimization.
Consumer and professional sensory “reviews” are collected on Gastrograph Review, a sensory system for collecting high resolution data across any food and beverage product. The AI is trained on data collected through large-scale R&D undertaken by Analytical Flavor Systems (AFS) and proprietary data from each client on the platform to build targeted models for new product development and existing product optimization.
Meeting consumer preference of specific target cohorts is a challenge faced by companies locally and globally - as consumers across the world have different perceptions and preferences based on experiential and demographic factors.
Not every individual drawn from the same demographic (age, sex, and race) will share the same perception, but together, each demographic forms a unique distribution in their perception of flavor that can be learned and modeled. Individuals from the same demographic may perceive flavor drastically differently as each demographic is composed of sub-populations (a segment of the population) and cohorts (clusters of consumers that share similar preferences).
Gastrograph AI can predict perceptions of flavor, aroma, and texture across demographics. Just like Google Translate can turn French into Chinese and Chinese into English, Gastrograph AI can take sensory data from any demographic and translate it to the target demographic.
To train the Gastrograph platform on each demographic, we conduct Demographic Surveys of consumers around the world. Collecting data directly from target consumers allows the AI to learn the distributions of perceptions for the target consumers and create a baseline to predict for that market.
The ability to translate perception across demographics allows us to collect data from anywhere and predict how a product would be perceived by a new demographic without ever needing to ship product or recruit consumer panelist from the target group – increasing speed and accuracy of new product development.
As a company, Analytical Flavor Systems helps food and beverage brands develop or optimize products for specific target demographic segments– not personalized products for individuals. However, Sleepwalkers challenged us to push the limits of our technology and develop a personalized seltzer optimized to fit their preferences.
We knew our platform could model at the level of the individual. To accomplish this, we’d need to take our work on population dynamics and use it to predict a maximum likelihood estimator for both Karah and Julian. We had Julian and Karah taste a number of on-market products to parameterize their distribution of perception. This allowed Gastrograph AI to learn their individual sensitivities, anosmia, and biases. We already have the majority of the sub-populations of the United States surveyed and knew Julian and Karah were drawn from this population, so little to no new work on the demographic profile of Julian and Karah was necessary.
Consumer perception of flavor, aroma, and texture vary across demographics and cohorts - and consumer preferences varies in response to these differences in perception. Thus, it is necessary to model what a consumer perceives before you can model their preferences. Analytical Flavor Systems undertakes Market Surveys in both general (all-market) and focused (specific category) forms to reverse engineer the preferences that are driving the market in aggregate and in specific categories. To train the Gastrograph Platform on each demographic and sub-population, our Market Surveys include consumers from target demographics tasting local products in their home markets. Our R&D work focuses on the general case, where each consumer will taste products across a huge range of categories in the ready-to-eat and ready-to-drink space.
Analytical Flavor Systems maintains a team, our Forward Deployed Tactical Global Panel Team, that travels around the world to a different country or region to recruit local consumers to taste 80 - 100 products over 5 days. This never-ending leading-edge data collection grows our proprietary database and allows us to predict for 15+ target countries around the world.
The ability to predict preferences for a product by any demographic or cohort allows CPG companies to decouple the development of a new product from its marketing and in-country testing. Products can be developed anywhere and sent to their highest value market in which they are competitive. In addition – the ability to predict preferences for each target demographic or cohort allows Gastrograph AI to directly develop new high-preference products and optimize existing products.
Gastrograph AI was developed to predict perception and preferences for a target consumer population. To develop a personalized product for Julian and Karah, we modeled them as their own sub-population, allowing the AI to translate our proprietary trunk data, hundreds of thousands of sensory reviews collected over 10 years, into Julian and Karah’s perception. While we could only have them physically taste and review a few products during the interview, we could use the translated data to predict what they would perceive in most of the sparkling water products available in the United States, greatly expanding our data set for optimizing a new product for them.
Gastrograph AI can predict perception and preference across 6 demographic factors: age, sex, race, socio-economic status, past tasting experience, and smoking habits. These factors can be used in any combination to define demographic tasting population segments, allowing companies to develop high-preference competitive products targeted at specific cohorts of consumers.
Our Forward-Deployed Tactical Global Panel Team has developed a rapid grab-and-go methodology allowing us to bring new demographics online in 5 days per sub-market. Our team can be (almost) anywhere in the world with only 2 weeks advance notice on average.
Gastrograph AI must always be up-to-date on the latest products in each market – as consumer preference is always changing. To keep our models on the cutting edge of new preference acquisition, Analytical Flavor Systems maintains standing panels in New York City with more opening in the future. These standing panels review products identified or collected by our R&D teams in each market and our Forward-Deployed Tactical Global Panel Teams in the field. We monitor the press, social media, and the shelves for new product launches, new direct to consumer products, and new white-label products entering the market. Our standing panels focus on both the cutting edge and the mass market – because what’s new today could be mainstream tomorrow. In addition to monitoring the world for new preferences, our standing panels allow us to rapidly train the AI on new categories. Our standing panel produces a minimum of 1,000 reviews per week – allowing us to pick a market on Monday and deploy the entire system on a new category on Friday.
When we took on this challenge, we had our standing panel profile a new set of sparkling water products. Even though we already have hundreds of sparkling waters in the data set, we wanted to profile any new and innovative products on the market to capture the most up to date preference states and new preference acquisitions in the NYC market. The AFS Standing panel reviewed five new flavored sparkling waters from local and national brands. As formulation started, we used the AFS Standing Panel in New York City to profile each iteration and translate their reviews to Julian and Karah’s perception. While individuals in our standing panel represent a diverse range of consumers, our translation model allows us to collect data from any set of individuals and translate it into Julian and Karah’s perceptions.
In any competitive market, flavor, aroma, and texture are primary drivers of consumer choice. Functional benefits, uncompetitive markets, and price sensitivity can drive consumers to less preferred flavor-profiles – but every category is one R&D cycle away from becoming competitive. Consumers, not companies, define the competition. If a consumer can pick something instead of your product, it’s competing with you. It could be a narrow range of unsweetened tea or a cross section of ready-to-drink beverages; anything the consumer prefers as much or more than your product is in the competitive landscape.
Gastrograph AI gives CPG companies a competitive advantage in their ability to develop competitive products versus the other products that are currently available on the market. In order to develop the most competitive product for a target consumer demographic – you collaborate with the AI in scoping the competition. To do so, companies select their products' key competitors. The AI will then select further products from our Market Surveys and Standing Panel reviews that are preferred and available to the target consumer demographic or cohort. This competitive set will be used to populate the Market Topography Map and Preference Topology Map. By defining the competition in collaboration with the AI, CPG companies can choose to target new high-preference white space or to enter the densest preference space with a market-winning flavor profile.
We used Karah’s and Julian’s favorite products – all LaCroix flavors – in addition to products they were unfamiliar with as the competitive set. Mixing in familiar and unfamiliar products allows the Gastrograph platform to map areas of preference and dispreference, which is just as valuable as it allowed the AI to narrow the search space and focus on new competitive white space that they would love. Competitive sets can be as large as the entire market or as narrowly scoped as a single competitor. Different competitive sets could result in different optimizations and make sense for different goals. Because sparkling water has no functional sweetener systems and no functional ingredients, we had few concerns about time intensity effects and no concerns about masking. We thus selected all of the products in the unsweetened-flavored sparkling water category.
The development of feature based learning for flavor, aroma, and texture, allows for a directed product development approach, where every iteration of the products’ formulation is more preferred than the last. Gastrograph AI can guarantee a locally or globally optimal formulation in three iterations, allowing for rapid development and iteration cycles across products, bases, and demographic targets. The Gastrograph AI platform includes a set of product formulator tools that help a product formulator explore and achieve the optimal flavor profile for their brand. The tools include predictions and calibrations for flavor, aroma, and texture. The first round of formulations is exploratory. The AI will predict new flavor profiles into white space that may or may not be high preference and competitive. In this way, the AI can fill in the blank spaces on the map with potential local optimizations and quickly back-propagate the results into an updated and improved local model for higher resolution and accuracy predictions in each subsequent round. A standard product development cycle would start with 1 – 3 predicted Exploratory Flavor Profiles per target demographic and concept.
While Julian and Karah are both wonderful and unique individuals with their own perceptions and preferences, the AFS team had limited time to formulate and test multiple new potential formulations. We asked the AI to find any convergent set of preferences for Julian and Karah. This had the effect of making it more difficult for the AI as it had to find a new set of flavors that both Karah and Julian had never tasted and would both like more than any other product on the market. It also gave our team fewer chances of success, as we would only be able to have them taste a single product of our formulation at the final tasting.
Not every preferred flavor profile is right for every CPG company, brand, or product portfolio. Many companies are turning towards clean labels, low sugar, or unsweetened products. Gastrograph AI supports any set of constraints on flavor, aroma, and texture and can predict its optimization around those constraints. For example – we could lock sugar and other sweeteners at 0 and predict the best possible unsweetened flavor profile. This ability allows a diverse set of CPG companies - such as confections, noodle soups, salty snacks, alcoholic beverages, and baked goods - with different targets to utilize the Gastrograph AI Platform to develop and optimize products for their target consumer cohorts with ease and accuracy.
Every iteration, concept test, and formulation can be tasted on the Gastrograph Sensory system benchtop (by the formulators) or by your internal sensory panel. 3 – 5 reviews are all the Gastrograph System needs to make accurate predictions about the global perception and preference of the product underdevelopment. Each formulation tasted on the system trains and updates the model and increases predictive accuracy and resolution of the platform – the ability for the AI to pick out distinct and unique flavors (such as lemon VS Meyer lemon). Targets can be formulated in-house by your team or by a flavor house partner. AFS will extend the Gastrograph AI platform so it can be used anywhere that it’s needed.
When developing a new product, the initial exploratory targets are formulated and tasted on the system. Those targets will update the AI and help it to predict optimal targets. Some or all of the exploratory targets may move forward into the optimization cycles. At the end of the three iteration cycle, at least one product will be in an optimal state.
The AI was able to find a few optimal targets that would meet the challenge for both Karah and Julian. Our team started with three exploratory optimizations. One was dropped after the first round for technical reasons. The two remaining targets went through two rounds of optimization, with major changes made in round 2 and very minor changes made in the final round. Both had a high chance of success, and we chose the one the AI predicted a marginally better distribution for.
Flavor Houses have become an integral part of the CPG industry and reliable partners for the development or reformulation of successful products. Analytical Flavor Systems has worked extensively with flavor houses to achieve optimal flavor, aroma, and texture profiles – and many Flavor Houses are happy to use our services. Partnering with a Flavor House and AFS gives your products and developments the greatest chance of success.
A group of consumers who perceive the same flavor profile in a product may all like the product – that is not surprising. A group of consumers who perceive the same flavor-profile but only some like it while others dislike it also exist – that is called a divergent preference. A group of consumers across two or more demographics, each of which perceive differences in the flavor profile, who all like the same product is a convergent preference. Traditional product development techniques look for flavor profiles with the highest number of convergent preferences – in order to create a single mass market product that everyone likes. The Gastrograph AI platform allows you to target the entire population, to optimize a mass market product, or any subset of the population to create a targeted product that those consumers love. With Gastrograph AI, the choice is yours.
Analytical Flavor Systems is at the forefront of bringing artificial intelligence to the food and beverage industry. Like any new technology it must be tested and validated before it can be trusted. We suggest that every company run the first few products they develop on the Gastrograph platform though their standard pre-release validation procedure. For many companies, that’s a 60/40 test or a stack-rank preference test, where our formulation needs to be preferred more than the existing formulation or top competitor.
This was a risky challenge for AFS to take on. We are confident in the Gastrograph AI platform's ability to predict optimal flavor profiles for a population – but would it work at the individual level? A live recording of just two people tasting and sharing their opinion of a product we would formulate in under two weeks left a lot of room for error. There are always bias, mood, and other unpredictable effects that wash out in a crowd but can affect the perception and preference of a person on any given day. We tuned and tested our predictions and formulation capabilities and finally had to validate if we created something amazing.
As they say: the proof is in the pudding – or in our case, the seltzer! To the surprise and delight of Karah and Julian, the AI was able to predict and the AFS team was able to formulate a new flavor of seltzer that they’ve never tried before and liked more than any of the available products on the market.
Looking for the same success?
Data secrecy and security is a top level concern. Analytical Flavor Systems never shares data, formulations, predictions, or learnings between companies on the platform. Every company on the Gastrograph AI platform has its own branch on the model tree and learning only flows from the trunk to the branch. The data collected from the AFS Standing Panels, Demographic Surveys, and Market Surveys make up our trunk data. The trunk data is used by the AFS Data science team to build, train, and validate the AI. All data collected by clients on the platform, including competitive sets, R&D Formulations, Flavor House example applications, experimental grids, and current products, are client-specific and are stored in a proprietary branch.
Trunk Data + Trunk Models + Your Data = Your Models
Your models are updated and cross trained from the trunk models on your data; each branch on the system has its own set of models updated on the proprietary data stored in the branch. This allows each company on the platform to gain a competitive advantage by collecting more data from target consumers, demographics, and formulations. Gastrograph AI is used by some of the largest companies ever to be under NDA. While we’ve developed successful products across industries and regions – we keep all client partnerships and products that we work on trade secret.