Gastrograph AI Frequently Asked Questions
How does Gastrograph work?
Gastrograph AI models people, products, and preferences. It learns how different cohorts of consumers around the world perceive flavor. It learns the flavors of individual products on the market in various countries, and it learns how to map those perceptions to those preferences. So the AI can predict for perception for different demographics, for other products anywhere in the world.
Why does it work?
AI as a technology has enabled a deeper interpretation of small, medium, and large-sized data; GG learns from all of its (big) data to create a mathematical space for products to be analyzed.
What problems does Gastrograph seek to solve for CPG companies?
Right now, all sensory and consumer insights are a cost center. To understand anything you're doing, you need to run more and more and more tests, and none of that has a compound ROI.
You're not building value. You're not making an asset. Using our technology and building up your branch, you're building a data asset that can be mined for future insights and reduces the amount of work you need to do in the future.
So not only is it faster and more accurate, but you're able to do that at a reducing cost basis to get a positive ROI out of all the consumer, product, and market research you're doing.
What does predictive mean?
To have a predictive framework means you can take all the past data and make predictions without running a new experiment. The insights from a predictive framework are statistical in nature. They can have predictive layers, meaning we can predict unidentified flavors preferences and run hundreds of scenarios of predictions on simulated products.
How is using Gastrograph different from traditional methods?
Sensory data is slow to collect and is very expensive. When using the traditional method, the study intends to answer specific questions about a particular product, so the data collected is only used once and becomes useless. With Gastrograph, the data collected gets used for various projects and demographics; this allows you to reduce, reuse, and recycle data collection.
Where does the data come from?
We have two standing panels: one panel in New York, where people come daily to taste various products available on the market, with a child panel running once a week. The second standing panel we have is in Shanghai.
We also have a team that travels worldwide to collect data on the markets. So far, we can model sensory perception and preference in over 30 countries, and we are continuously expanding.
Who owns the data?
Gastrograph AI owns all the data but operates on a trunk and branch model. Our trunk is built from our standing panels in New York and Shanghai and other collections and acts as the foundation of predictions. Any data collected from prototypes, requested products, or data collected by a client goes directly into their unique branch, and they have exclusive access. Predictions get made using our trunk data and their branch. Investing more into their unique branch allows a larger branch and better predictions.
How accurate are Gastrograph's predictions?
The only public double-blind validation study is with Ajinomoto. We ran nine coffee products in Japan, predicted for nine consumer demographics in China, validated it on an N=242 person CLT. We were able to show that we were as accurate, if not more accurate, for perception and preferences across all nine products in all nine demographics. We considered that to be the gold standard. For preference modeling, our minimum is an 85% accuracy rate.
How do we keep data up to date?
We update our models regularly. We update every five years for the countries defined as mature (such as the US or European countries). We update our models every 2 or 3 years for countries evolving faster, such as China.
How has Gastrograph's predictive quality been validated? Why is this important?
We believe that for you to see value in Gastrograph, you have to use it. For you to use it, you have to trust it. So we think the validation studies in one of the pillars of that type of trust, saying, "Here is indisputable evidence the system is working."
You can find our full validation study with Ajinomoto here.
How are insights delivered?
The system's goal is to make actionable insights that allow you to make decisions. We will work with you to find the method you need to apply those insights at your company's best, whether that's in deck format for presentations or in CSV values, for you to plug into your existing data analysis tools. Depending on the project's goal, those insights might be descriptive around the way that the world currently is, a prediction about the way that the world will be, or prescriptive, a prediction around what you should do.
Are our system and data proprietary? If so, can you speak to why this is important or a differentiator for Gastrograph?
Our system and data are proprietary. Gastrograph has been collecting sensory data for over 12 years, creating longitudinal data and experience working with that data to build increasingly more accurate models. Unlike other traditional methods where information gets used once, Gastrograph data could be reused and recycled indefinitely to make predictions.
Who tastes the products?
We use humans in the loop AI. If human preferences get built on individual and unique perceptions, and your tongue is perfectly predictive of what you're going to taste, then you need humans in the loop. We use AI on the data collection and generalization, and prediction sides. Being able to take data from any given random group of individuals, non-represented samples of consumers, and use that data to model and predict what any other consumer cohort will perceive in the product and predict the distribution preferences.
How can you make broad predictions from such low sample sizes?
Making predictions uses all of the data and leverages all of the data we've ever collected. We can train big data, so inference only needs small sample sizes.
To ensure the quality of our data, we have baseline measurements, internally, the outright reject data. We also have two internal metrics, experience score, and trust score. An experience score is built up over time and quantifies a user's ability to identify subtlety and nuance. So put another way, it's how descriptive they are, how good they are at identifying flavors. A trust score gets calculated on a peer-review basis, where we can see whether or not there were missing flavors that we would expect them to taste in the product based on what other people are tasting.
Is there an expert panel that evaluates the product? Please explain why or why not experts get included in our panels.
No, we use an average consumer. We would have less confidence in the data if we used expert tasters because any form of training distances professional panelists from the consumer's perception and experience.
There's no reason to assume that any two individuals on a panel should score things the same way or use the same limited tasting vocabulary. It leads to things like offloading. It leads to things like meta-analysis, where I'm guessing what the rest of the panel will say, so I get to the correct answer. So instead, we use one hundred percent trained but uncalibrated consumers to collect data that are predictive of consumer perception, the consumer preference, which at the end of the day is what's going to determine the success or the failure of that product.
How do you account for subjective differences in tasting?
The AI works by modeling at three different levels. It works at first, the individual level, the environmental level, and then the product level. So from last to first, when we're looking at a product, we're looking at the multitude of reviews we get around that product and the physiological response to underlying sensory stimuli.
Can we show an increase in sales after we work with a company to optimize a product?
The short answer is no; nobody will share their sales numbers. So instead, when we work with companies, we think of validation at three different levels.
Internal validation. Internal validation gives us an idea of what types of error rates we expect. We will never launch models with below high eighties, low nineties percent accuracy, using accuracy in the general sense.
The second level of validation that we do is head-to-head validation. We will make a prediction, and then you'll test it against prior information you have, and we can show that those things match or that we're more accurate.
Finally, the most extensive form of validation we do, the only thing that we would say is validating, is a double-blind validation study. We were able to show that our predictions were equivalent or more accurate than the validated panel.
How does Gastrograph work better than what CPGs are already doing?
Currently, major CPG companies are doing internal validations, which take up to three years to know whether a product will succeed or fail. We're able to help companies create products that consumers like more repeatedly. We believe that it is essential to optimize for repeat consumption. Current validation doesn't account for whether there will be a sustained preference for repeat consumption. In contrast, our tool, because of the way we collect the data and because of the way we build the models, is highly targeted and predicted for repeat consumption scenarios.
How can CPGs trust the data they receive from Gastrograph?
We are always learning, always up to date, actionable insights.
Create a better-tasting world
Get in touch to speak with our team so we can learn about the challenges you're facing.