Human Flavour Perception, Simulated

TasteNET simulates how people perceive flavour, across 65 families of taste, mouthfeel and aroma and about 2,000 descriptors. On the sphere, the further a dot sits from the red core, the more it is liked.

Liked most: Hot, Roasted and Crisp.

Liked most: Citrus, Carbonation and Juicy.

Liked most: Umami, Hot and Sweet.

Liked most: Umami, Meaty and Hot.

Liked most: Hot, Warming and Kokumi.

Liked most: Citrus, Floral and Carbonation.

Liked most: Sweet, Fatty and Caramel.

Liked most: Hot, Umami and Roasted.

The Structure of Taste

Every bite passes through eight layers of one person's taste, ordered by how fast they change. The slowest sit at the centre.

Every human is born liking sweet, rejecting bitter and wary of foods never eaten.

Fixed. Shapes perception.

Receptor genes set how bitter or fragrant the same molecule tastes to you.

Fixed for life. Shapes perception.

The mother's diet reaches the infant through amniotic fluid and breast milk.

Set in early windows. Shapes perception.

A cuisine's flavour principles, identity, and how family and peers eat.

Changes over decades. Shapes expectation.

Fear of new foods is highly heritable, and sensation seekers like spicy food more.

Changes over years. Shapes expectation.

Every bite rewrites it: young children accept a food more after tasting it daily for 8 to 10 days.

Changes with each exposure. Shapes expectation.

The same food pleases when hungry and less when full, and a food just eaten loses its appeal.

Changes within hours. Tunes liking in the moment.

Venue, price, packaging, colour and sound set an expectation before the first taste.

Changes within seconds. Shapes expectation.

Sweet LikingBitter RejectionWariness of the New1Innate Bitter Receptor TAS2R38Sweet and Umami ReceptorsHighly Palatable DimensionAcquired Dimension2Genes Flavour in Amniotic FluidFlavour in Breast MilkFormula FlavourChildhood Sweet PreferenceChildhood Bitter Sensitivity3Early Life Cuisine Flavour PrinciplesIdentityFamily TableSocial ModellingFood Environment4Culture Food NeophobiaSensation SeekingReward Sensitivity5Traits Repeated ExposureSpread of FamiliarityFlavour-Flavour LearningFlavour-Nutrient LearningHabit6Learning Hunger and FullnessSensory-Specific SatietyFullness and SmellGut Microbiota7Body State Price TagVenueLabels and PackagingColourSound8Setting and Cues One Bite

Methodology: How TasteNET Twins Are Built

TasteNET decodes the human-food relationship by applying first-principles methods across its system design. This section explains, for food and beverage research teams, how a digital twin is built in four steps: measure real people, build the twin, place the twin in its market, and simulate the decision.

Engraving of Santorio Santorio seated in a weighing chair hung from a steelyard, beside a table set with a meal: the frontispiece of Medicina Statica
Santorio Santorio in his weighing chair, weighing himself after a meal. Wellcome Collection.

What Makes TasteNET Different

Our key differentiator is first-party sensory and behavioural data, modelled through principles grounded in the real human decision journey.

Four Steps

We built our own consumer-facing data pipeline around the 16 Taste Types, the 8 Foodie Personalities and the Food Memory Interview.

Our taxonomy

  • 16 Taste Types: a three-part psychophysical assessment of aroma, taste and mouthfeel preferences, covering 53 aromas, 32 forced-choice taste and texture pairs and 7 intensity dimensions.
  • 8 Foodie Personalities: 98 questions on what drives a person's food choices, from price and health to social influence, brand loyalty, exploration and sharing.
  • Food Memory Interview: a self-paced session of six questions on food origins, peak sensory memories, a real trial purchase and emotional eating, capturing the lived experience that structured instruments cannot reach.
Try the three instruments yourself askyumi.com ↗

Before predicting acceptance, we measure intensity and preference curves for every key sensory and behavioural attribute.

Our design principles

  • Each region or cohort is profiled with more than 1,000 consumer profiles for statistical depth and fine-grained resolution.
  • A clear algorithmic matrix separates bio-mechanical thresholds, such as receptor-level sensitivity, from cognitive perception data, such as preference, memory and cultural association.
  • Biophysical parameters become mathematical Sensory Power Indices, so sensory perception can be compared across cohorts in the same analytical space.
  • Model outputs are supported and validated through controlled-environment focus groups.

Every simulation runs against a live World State, a continuously updated layer of cultural signals, trending flavours, health narratives and social dynamics specific to your target market.

For example

  • If Heritage Detox Brew is trending in China in February 2026, twins in that region factor this cultural momentum into their purchase and advocacy decisions.

Only bottom-up simulation at the individual level lets brands move beyond averaged metrics to evidence-backed decisions: what to launch, how to position it and who will buy it.

Our AI capacities

  • Interprets user intent with domain-specific awareness, combining product category, regional context and sensory variables.
  • Synthesises first-party data with secondary reports and public datasets to generate evidence-backed recommendations.
  • Operates on pre-modelled sensory baselines and returns results at the resolution of region, cohort or individual.
  • Accepts proprietary uploads such as product briefs and formulation notes, with no cross-client data sharing and no model retraining exposure.
Three instruments, one consented profileReal people on askyumi.comAromaTasteTextureSocialPriceLoyaltyHealthExploreShareOriginsSensoryDecisionFeeling01020392 items · 3 parts98 items6 questionsOwned by the userConsented, anonymisedStrongerWeakerReport back16 TasteTypes8 FoodiePersonalitiesFood MemoryInterviewFoodie profileReport back
Intensity and preference, for every attributeLikingIntensityReceptorthresholdSweetSourSaltyBitterUmamiCohort ACohort BPreference peakSensory Power Index1,000+1,000+
A live World State behind every twinCultural signalsTrending flavoursHealth narrativesSocial dynamicsSepOctNovDecJanFebweigh the trend inpurchase and advocacyDecision shiftedUnaffectedTrend signalHeritage Detox BrewTwins in China
Individual decisions show what the average hidesDislikeLikewhere few twins sitWhat to launchHow to position itWho will buy itWill buyWill notUndecidedAverage 5.2Recipe BMorning ritual45% of twins45%40%15%

What Twins Are Not

  • Real peopleOne twin eachAveraged persona

    Not Synthetic Personas

    No synthetic data. No averaged personas.

  • MeasuredNever collectedTaste and smellFood behaviourFood memoriesDNAHealth recordsBiometrics

    Not Genetic or Biometric

    No genetic data, no personal health records, no biometric data and no DNA.

The Science Behind It

References

Traditional consumer research asks what people think. TasteNET measures the drivers behind the answers.

Psychophysics captures each person's stable sensory fingerprint: detection thresholds, JAR windows and preference patterns. Thermodynamics explains why those preferences hold: every food decision seeks a lower-energy equilibrium, where sensory reward outweighs the cost of processing.

No two people taste the same food: receptor genetics, psychophysics and how perception is measured.

  1. Bartoshuk, Duffy & Miller (1994) PTC/PROP tasting: Anatomy, psychophysics, and sex effects. Physiology & Behavior
  2. Kim et al. (2003) Positional cloning of the human quantitative trait locus underlying taste sensitivity to phenylthiocarbamide. Science
  3. Bufe et al. (2005) The molecular basis of individual differences in phenylthiocarbamide and propylthiouracil bitterness perception. Current Biology
  4. Keller et al. (2007) Genetic variation in a human odorant receptor alters odour perception. Nature
  5. Bartoshuk et al. (2003) Labeled scales (e.g., category, Likert, VAS) and invalid across-group comparisons: What we have learned from genetic variation in taste. Food Quality and Preference
  6. Bartoshuk et al. (2004) Valid across-group comparisons with labeled scales: The gLMS versus magnitude matching. Physiology & Behavior
  7. Hayes & Keast (2011) Two decades of supertasting: Where do we stand? Physiology & Behavior
  8. Puputti et al. (2018) Multidimensional measurement for individual differences in taste perception. Food Quality and Preference
  9. Rozin (1982) “Taste-smell confusions” and the duality of the olfactory sense. Perception & Psychophysics
  10. Breslin & Beauchamp (1997) Salt enhances flavour by suppressing bitterness. Nature
  11. Zampini & Spence (2004) The role of auditory cues in modulating the perceived crispness and staleness of potato chips. Journal of Sensory Studies
  12. Spence & Levitan (2021) Explaining crossmodal correspondences between colours and tastes. i-Perception
  13. Yeomans et al. (2008) The role of expectancy in sensory and hedonic evaluation: The case of smoked salmon ice-cream. Food Quality and Preference
  14. Plassmann et al. (2008) Marketing actions can modulate neural representations of experienced pleasantness. PNAS
  15. King & Meiselman (2010) Development of a method to measure consumer emotions associated with foods. Food Quality and Preference
  16. Lee et al. (2023) A principal odor map unifies diverse tasks in olfactory perception. Science
  17. Rosenthal & Chen (2023) Dimensions of food texture: A conceptual discussion. Journal of Texture Studies
  18. Doty (2025) Odors as cognitive constructs: History of odor classification and attempts to map odor percepts to physical and chemical parameters. Chemical Senses

How food choices are made, measured and spread: preference learning, decision drivers and social contagion.

  1. Rozin & Fallon (1987) A perspective on disgust. Psychological Review
  2. Rozin (1990) Acquisition of stable food preferences. Nutrition Reviews
  3. Berridge (2009) ‘Liking’ and ‘wanting’ food rewards: Brain substrates and roles in eating disorders. Physiology & Behavior
  4. Shiv & Fedorikhin (1999) Heart and mind in conflict: The interplay of affect and cognition in consumer decision making. Journal of Consumer Research
  5. Raghunathan, Naylor & Hoyer (2006) The unhealthy = tasty intuition and its effects on taste inferences, enjoyment, and choice of food products. Journal of Marketing
  6. Haws, Reczek & Sample (2017) Healthy diets make empty wallets: The healthy = expensive intuition. Journal of Consumer Research
  7. Onwezen et al. (2019) The development of a single-item Food Choice Questionnaire. Food Quality and Preference
  8. Roininen, Lähteenmäki & Tuorila (1999) Quantification of consumer attitudes to health and hedonic characteristics of foods. Appetite
  9. Lichtenstein, Ridgway & Netemeyer (1993) Price perceptions and consumer shopping behavior: A field study. Journal of Marketing Research
  10. Bearden, Netemeyer & Teel (1989) Measurement of consumer susceptibility to interpersonal influence. Journal of Consumer Research
  11. Van Trijp & Steenkamp (1992) Consumers’ variety seeking tendency with respect to foods: Measurement and managerial implications. European Review of Agricultural Economics
  12. Tangney, Baumeister & Boone (2004) High self-control predicts good adjustment, less pathology, better grades, and interpersonal success. Journal of Personality
  13. Cai, Chen & Fang (2009) Observational learning: Evidence from a randomized natural field experiment. American Economic Review
  14. Salganik, Dodds & Watts (2006) Experimental study of inequality and unpredictability in an artificial cultural market. Science
  15. Granovetter (1978) Threshold models of collective behavior. American Journal of Sociology
  16. Watts (2002) A simple model of global cascades on random networks. PNAS
  17. Centola & Macy (2007) Complex contagions and the weakness of long ties. American Journal of Sociology
  18. Centola (2010) The spread of behavior in an online social network experiment. Science
  19. Christakis & Fowler (2007) The spread of obesity in a large social network over 32 years. New England Journal of Medicine
  20. Pachucki, Jacques & Christakis (2011) Social network concordance in food choice among spouses, friends, and siblings. American Journal of Public Health
  21. Bass (1969) A new product growth for model consumer durables. Management Science
  22. Rand & Rust (2011) Agent-based modeling in marketing: Guidelines for rigor. International Journal of Research in Marketing

Food as who we are: culture, class, family, early flavour learning and the memories that taste and smell bring back.

  1. Barthes (1961) Pour une psycho-sociologie de l’alimentation contemporaine. Annales: Économies, Sociétés, Civilisations
  2. Douglas (1972) Deciphering a meal. Daedalus
  3. Bourdieu (1984) Distinction: A social critique of the judgement of taste. Harvard University Press
  4. Fischler (1988) Food, self and identity. Social Science Information
  5. Nemeroff & Rozin (1989) “You are what you eat”: Applying the demand-free “impressions” technique to an unacknowledged belief. Ethos
  6. Rozin & Millman (1987) Family environment, not heredity, accounts for family resemblances in food preferences and attitudes: A twin study. Appetite
  7. Rozin et al. (1999) Attitudes to food and the role of food in life in the U.S.A., Japan, Flemish Belgium and France: Possible implications for the diet-health debate. Appetite
  8. DellaPosta, Shi & Macy (2015) Why do liberals drink lattes? American Journal of Sociology
  9. Majid (2021) Human olfaction at the intersection of language, culture, and biology. Trends in Cognitive Sciences
  10. Mennella, Jagnow & Beauchamp (2001) Prenatal and postnatal flavor learning by human infants. Pediatrics
  11. Tulving (1972) Episodic and semantic memory. Organization of Memory, Academic Press
  12. Tulving & Thomson (1973) Encoding specificity and retrieval processes in episodic memory. Psychological Review
  13. Herz & Schooler (2002) A naturalistic study of autobiographical memories evoked by olfactory and visual cues: Testing the Proustian hypothesis. American Journal of Psychology
  14. Willander & Larsson (2006) Smell your way back to childhood: Autobiographical odor memory. Psychonomic Bulletin & Review
  15. Mojet & Köster (2005) Sensory memory and food texture. Food Quality and Preference
  16. Köster & Mojet (2015) From mood to food and from food to mood: A psychological perspective on the measurement of food-related emotions in consumer research. Food Research International

Simulating people with language models: what works, where twins distort, and how fidelity is tested.

  1. Park et al. (2023) Generative agents: Interactive simulacra of human behavior. UIST ’23
  2. Park et al. (2024) Generative agent simulations of 1,000 people. arXiv:2411.10109
  3. Argyle et al. (2023) Out of one, many: Using language models to simulate human samples. Political Analysis
  4. Horton (2023) Large language models as simulated economic agents: What can we learn from Homo silicus? NBER Working Paper 31122
  5. Brand, Israeli & Ngwe (2023) Using LLMs for market research. Harvard Business School Working Paper 23-062
  6. Goli & Singh (2024) Frontiers: Can large language models capture human preferences? Marketing Science
  7. Arora, Chakraborty & Nishimura (2025) AI-human hybrids for marketing research: Leveraging large language models (LLMs) as collaborators. Journal of Marketing
  8. Toubia et al. (2025) Twin-2K-500: A data set for building digital twins of over 2,000 people based on their answers to over 500 questions. Marketing Science
  9. Peng et al. (2026) Digital twins are funhouse mirrors: Five systematic distortions. Science Advances
  10. Binz et al. (2025) A foundation model to predict and capture human cognition. Nature
  11. Bisbee et al. (2024) Synthetic replacements for human survey data? The perils of large language models. Political Analysis
  12. Wang, Morgenstern & Dickerson (2025) Large language models that replace human participants can harmfully misportray and flatten identity groups. Nature Machine Intelligence
  13. Kirk et al. (2024) Understanding the effects of RLHF on LLM generalisation and diversity. ICLR 2024
  14. Chuang et al. (2024) Beyond demographics: Aligning role-playing LLM-based agents using human belief networks. Findings of EMNLP 2024
  15. Zhang et al. (2025) Leveraging dual process theory in language agent framework for real-time simultaneous human-AI collaboration. arXiv:2502.11882
  16. Ziabari et al. (2025) Reasoning on a spectrum: Aligning LLMs to System 1 and System 2 thinking. arXiv:2502.12470
  17. Piao et al. (2025) AgentSociety: Large-scale simulation of LLM-driven generative agents advances understanding of human behaviors and society. arXiv:2502.08691
  18. Anthis et al. (2025) Position: LLM social simulations are a promising research method. ICML 2025

Validation: Where and When a Twin Can Be Trusted

How long, how far, how deep.

Evidenced Not Yet Verified Twin Profile

    The three questions clients ask most are three edges of one question. Each dot is one decision event: solid where there is evidence today, hollow where it still lies beyond the edge.

    Time Is a profile valid for months, or for the long term? One Quarter: The default horizon. All Phase I evidence comes from a single sitting, inside this quarter. Several Quarters: Adjacent evidence only: a segment-level simulator held across four months. Years: No data yet.

    Scenario What can it predict, and what can it not? Stable Preference: Phase I matched 90 to 95%, against 50% by chance. Emotion-Driven: The emotion task recovered 66 to 74% of picks, against 25% by chance. Context Reversals: Phase I matched 45 to 49%, close to the 50% of chance.

    Depth If it surfaces unconscious motives, how is that verified with the person? Stated Preference: What the survey measures directly. Recognised Motives: Evidence needed: picking one's own description out of a line-up. Unspoken Motives: Evidence needed: predicting behaviour across situations better than the person can.

    Validity Over Time

    How long a twin holds depends on how fast the market moves.

    Backtested No Backtest Yet Deeper Tone, Longer Shelf Life

      A twin is a portrait of one moment, and the market keeps moving. We take a finished twin back into the 3 to 12 months of real market that followed and check in which month its calls start to drift. The faster a market moves, the shorter the shelf life.

      Fast Market Momentum New products, channels and trends turn over quickly here, and choices move with them.

      Moderate Market Momentum The market moves, but mainstream preferences and occasions stay fairly stable.

      Slow Market Momentum Mature categories, with tastes and habits that change slowly.

      Scenario Coverage

      Endless occasions, eight variables.

      TrustedNeeds CalibrationUntestedPass Line

        Eating and drinking are set by a person, a product and a situation. The person is the twin profile, the product is the category, and what is left over is the occasion. Cut it into eight variables that do not overlap, four in the person and four around, each validated against real people on its own.

        Time Where this moment sits in the day, the week and the year.

        Ambience How the surroundings feel: heat, cold and weather, plus light, sound and smell.

        Place Where the person is when deciding, and what they are doing.

        Company Who is present and what role the person plays.

        Purpose What this bite is for, and for whom.

        Slack Whether time and money are enough right now.

        Mood The feeling of the moment.

        Body What the body needs right now.

        Depth

        What is said, and what goes unsaid.

        Evidenced Not Yet Verified Twin Profile

          An iceberg draws one waterline and fixes it in place. Here there are two lines and both move: the edge of words follows how well one asks, today's edge follows the evidence. System 1 and System 2 are shorthand, not two machines; the five layers shade into one another. One food or drink decision is cut from choice down to belief, and each dot is one question put to a twin.

          Choice Picked in a survey, not yet done in a real setting.

          Reason Slow, effortful, verbal.

          Motive Unsaid until asked, recognised once shown.

          Gut Feeling Fast, automatic, seldom said.

          Belief Priors, slowest to change, hardest to put into words.

          This section shows where and when a TasteNET digital twin can be trusted, for food and beverage teams deciding how far to rely on a simulation. Trust has three edges: how long a profile stays valid, what a twin can and cannot predict, and how deep into motive its answers can be verified. Across 20 benchmarks, TasteNET twins reach 86% average accuracy against real consumer behaviour.

          Cross-section of Mount Chimborazo with the names of plants written at the altitudes where they grow, flanked by columns of measurements
          Alexander von Humboldt and Aimé Bonpland, Geographie der Pflanzen in den Tropen-Ländern, 1807.

          The Current Fidelity Report

          Backtests agree with real consumers 86% of the time. Case validation looks underneath that number.

          Does the twin answer as the person does? Figures come from the backtest and Phase I.

          Does the twin speak from its own profile, and for the right reasons? Case validation, 86 items and 628 answers.

          Does the twin fabricate or contradict itself? Case validation, lower is better.

          Backtest Overall

          20 benchmarks
          86%

          Overall agreement across 20 benchmarks. Chance differs by benchmark, so this figure does not convert into a margin over chance.

          Stable Preference

          Phase I, chance 50%
          90 to 95%

          Two-way choices in ordinary situations, such as sparkling or still water on a weekday afternoon. Above 90%, the steadiest ground a twin stands on today.

          Emotion-Driven

          Phase I, chance 25%
          66 to 74%

          A dish-choice task under an emotional state, where chance is only 25%. The hit rate is around 70%, so its margin over chance is no smaller than that of stable preference.

          Context Reversals

          Phase I, chance 50%
          45 to 49%

          Items where the same person flips with the occasion. The hit rate is close to chance, and this is where the scenario edge sits today.

          Profile Read-In

          Pass mark 90%
          93 to 95%

          Whether a twin answers in line with preferences its profile states outright. This is the first gate of validation: above 90% the profile counts as read in correctly, and tension items follow.

          Generalisation and Tension

          Not stated in the profile
          78 to 92%

          Two traits in the profile are set against each other, or the twin is asked about foods the profile never mentions, to see whether it reaches the right direction. Eight independent samples per item, scored by item.

          Mechanism Hit

          The right reason
          50 to 88%

          A right answer is not enough: the reason has to trace back to the matching line in the profile. The spread comes from how the profile is written. The older style that lists instances gives 50%. Writing by category with conditional sentences, or the new itemised format with reaction times, both give above 80%, up to 88%. The top case excludes items affected by one prompt-template problem.

          Fabrication

          Share of answers
          2 to 8%

          Experiences or details a twin supplies with no basis in its profile. Thicker data means fewer: 2% with first-person data, 4 to 8% without.

          Outliers

          Share of answers
          0.4 to 3.7%

          Single answers that directly contradict the profile. This is why one sample is never used for a decision: production takes the majority of several samples and filters reasons for consistency.

          Figures against real people come from the backtest and Phase I; profile and cleanliness figures come from case validation, 86 items and 628 answers on one model. The three use different definitions and are not directly comparable.

          How Accuracy Is Verified

          Every verdict is checked against human answers the twin never saw. Three answer keys feed the same four tests, and each verdict covers one condition at one level of inference.

          Run for every twin. It is scored on items held out from its profile, and its hit rate is reported as a share of the person's own consistency: two weeks apart for noise, two months apart for the time edge.

          A twin population reruns a question whose real outcome is already known, from the market or a past study. The result is scored on two axes: accuracy against what consumers actually did, and completeness, the breadth of decision factors captured.

          Built with your team on studies you already hold. Your past results become the held-out answer key, so each verdict covers your own categories, markets and conditions.

          Answer Key

          Day 0Ringed items are held out
          TwinEight samples per held-out item
          2 WeeksSame items again: the ceiling
          2 MonthsCeiling for the span

          AnswerHeld OutChanged on RetestTwin Samples

          Many samples per item, so a distribution is compared, not one lucky draw.

          Twin PopulationReruns a question with a known outcome
          Real ConsumersAccuracy: what actually happened
          Decision FactorsCompleteness: filled ones the twins also named
          • Taste
          • Texture
          • Price
          • Pack
          • Claim
          • Occasion
          • Habit
          • Peers
          • Health
          • Brand

          ChoseDid NotNot Reached

          One run is set against one known outcome, on both axes at once.

          Your Studies
          Conditions you name: market, segment or occasion
          Concept TestPurchase intent
          Sensory PanelLiking and JAR
          Price TestWillingness to pay
          Usage and AttitudeHabits and occasions

          TrustedNeeds CalibrationWalk AwayUntested

          Past results are held out and rerun blind. A cell with no study stays untested.

          Precondition

          Four Tests

          Verdict

          • TrustedUsable for decisions at this level.
          • Needs CalibrationDirection is usable. Quantitative claims need a human sample for calibration.
          • Walk AwayThe signal itself is unreliable. Use human research for this cell.
          • UntestedExploration only, not for decisions.

          Flip the switches to try it. An anchor expires when the model, prompt, item type or population changes; thresholds are starting points, re-registered against decision risk before formal use.

          Pushing the Edges Outward

          Online, large-sample retests and comparisons. Offline, cross-modal validation in controlled settings.

          WorkstreamModeScenarioDepthTimeWhat It Does
          Retest PanelOnlineThe same users are tested again after two weeks and after two months, giving decay curves for the sensory, behaviour and memory layers.
          Transfer GradientOnlineTested step by step along transfer distance: new items, new categories, new kinds of behaviour.
          Three Grades of EvidenceOnlineSelf-recognition, witnesses who know the person, then behavioural prediction across situations.
          Ablation and Human BenchmarkOnlineInjected information is removed block by block and scored against the person's own survey on three layers.
          Academic ExperimentsOnlinePhase I completed with a university partner; follow-up studies run with further academic partners.
          Singapore Crossmodel Sensory LabOfflineA space where every sensory cue is controlled, opening in early 2027. Same person, same food, different light, sound, scent, tableware and company.
          Restaurant SandboxOfflineA computational social experiment with a university lab at a Beijing restaurant, run in parallel against actual ordering records.
          Eye Tracking and EEGOfflineLab studies calibrate with ratings and behaviour first, then add eye tracking and EEG as evidence for unspoken motives.

          Set the Scope, Then Trust the Answer

          Put your question in, see which side of the edge it falls on, and which route it takes.

          When
          Inside all three edges
          You Provide
          Demographic filters
          You Get
          Twin interviews and simulation right away
          First Step
          Today
          When
          Beyond the scenario edge
          You Provide
          Demographic requirements, main use cases and categories
          You Get
          Custom twins that carry a memory of the brand
          First Step
          50 twins
          When
          Beyond the depth or time edge
          You Provide
          A human sample to revisit and a retest cadence
          You Get
          A human anchor and a verdict for this condition
          First Step
          A validation design drawn up together
          TimeScenarioDepth
          • Twin
          • Human
          • Your Question
          • Edge Crossed
          • New Edge
          Detail of Nymph of the Luo River, a handscroll after Gu Kaizhi in the Palace Museum, Beijing: the nymph in a crescent boat, four figures under willows, a drummer and a spirit crossing the water

          Frequently Asked Questions

          Why do two people taste the same food differently?

          People taste the same food differently because taste is shaped by eight layers that vary from person to person. Receptor genes set how bitter or fragrant the same molecule tastes. Early-life exposure, culture, personal traits and learning shape what a person expects. Body state and setting then tune liking in the moment.

          Why does the same person prefer different flavours on different occasions?

          Because liking moves with body state and setting. The same food pleases more when hungry than when full, and venue, price, packaging, colour and sound set an expectation before the first taste. TasteNET models preference per occasion: a hungry morning and a summer afternoon give the same person two different liking patterns.

          Which flavour attributes does TasteNET model?

          TasteNET models flavour perception across 65 families of taste, mouthfeel and aroma and about 2,000 descriptors. The underlying assessment measures each person's liking of 53 aromas, 32 forced-choice taste and texture pairs and 7 intensity dimensions.

          How is a TasteNET digital twin built?

          A TasteNET digital twin is built in four steps. First, one real person is measured with three instruments: 16 Taste Types, 8 Foodie Personalities and the Food Memory Interview. Second, the twin is built from that person's intensity and preference curves. Third, the twin is placed in a live World State of its market. Fourth, the twin simulates the decision.

          What are the 16 Taste Types?

          16 Taste Types is Digitaste's psychophysical assessment of a person's aroma, taste and mouthfeel preferences. It covers 53 aromas, 32 forced-choice taste and texture pairs and 7 intensity dimensions, and sorts sensory preference on four axes: salt or acid, coating or texture, light or rich flavour, fresh or mature tone.

          What are the 8 Foodie Personalities?

          8 Foodie Personalities is Digitaste's framework for how a person chooses food. It uses 98 questions on what drives food choices, from price and health to social influence, brand loyalty, exploration and sharing. It sorts people on three axes: rational or impulsive, self-focused or social-focused, settler or explorer.

          What is the Food Memory Interview?

          The Food Memory Interview is a self-paced session of six questions on food origins, peak sensory memories, a real trial purchase and emotional eating. It captures the lived experience that structured instruments cannot reach, and it gives each TasteNET twin its food memories. All three instruments can be tried at askyumi.com.

          How many real consumers are behind each market in TasteNET?

          Each region or cohort in TasteNET is profiled with more than 1,000 consumer profiles, for statistical depth and fine-grained resolution. In total, TasteNET draws on 110K+ real-world sensory profiles across 118 regions in 38 countries.

          How is TasteNET different from food trend reports?

          Trend reports describe a market in general terms. TasteNET simulates how individual consumers will choose one specific product. Each twin is built bottom-up from one measured person, and every simulation runs against a live World State of cultural signals, trending flavours and health narratives, so a trend enters as context for the decision.

          What research is TasteNET based on?

          TasteNET draws on four research areas: sensory science and psychophysics, food choice behaviour and social contagion, food identity and memory, and the simulation of people with language models. The Methodology page lists more than 70 academic references across the four, from taste genetics to generative agent simulations of 1,000 people.

          What data are the digital twins built on?

          Every digital twin is grounded in first-party sensory and behavioural data collected through our self-reporting foodie community. Our database includes 110K+ real-world sensory profiles spanning 128 persona types across 118 regions. Each twin carries measured flavour preferences, food choice values, and a demographic profile. Unlike generic AI consumer insight platforms that rely on synthetic personas, TasteNET digital twins reflect how people actually perceive and choose food, based on who they are in the real world.

          Does TasteNET account for real-time market trends and cultural context?

          Yes. Every simulation runs against a live World State, a continuously updated layer of cultural signals, trending flavours, health narratives, and social dynamics specific to your target market. For example, if Heritage Detox Brew is trending in China in February 2026, digital twins in that region will factor this cultural momentum into their purchase and advocacy decisions, so simulations reflect the market as it is today.

          How is TasteNET's 86% accuracy measured?

          86% is the overall agreement between TasteNET twins and real consumers across 20 backtest benchmarks. Results vary by task: 90 to 95% on stable preferences (chance 50%), 66 to 74% on emotion-driven dish choices (chance 25%) and 45 to 49% on choices that flip with the occasion (chance 50%).

          What can TasteNET twins not predict yet?

          TasteNET twins are weakest where the same person flips with the occasion: on these context reversals the hit rate is 45 to 49%, close to chance. There is also no evidence yet for horizons of years, and motives a person cannot put into words are not yet verified. For these cases, use human research.

          How long does a TasteNET digital twin stay valid?

          The default horizon for a TasteNET twin is one quarter. Beyond that there is adjacent evidence only: a segment-level simulator held across four months. There is no data yet for years. Shelf life depends on how fast the market moves, and Digitaste backtests twins against the 3 to 12 months of real market that followed.

          How does Digitaste verify a twin's accuracy?

          Every verdict is checked against human answers the twin never saw. Each twin is scored on items held out from its profile, with eight samples per item, and its hit rate is reported as a share of the person's own consistency on retest. Four tests follow: direction, magnitude, subgroups and decision.

          Can TasteNET be validated against our own past studies?

          Yes. Digitaste builds custom benchmarks with your team from studies you already hold, such as concept tests, sensory panels, price tests and usage and attitude studies. Your past results are held out and rerun blind, so each verdict covers your own categories, markets and conditions.

          What do Trusted, Needs Calibration, Walk Away and Untested mean?

          They are the four verdicts TasteNET gives for one condition. Trusted means usable for decisions. Needs Calibration means the direction is usable, but quantitative claims need a human sample. Walk Away means the signal is unreliable and human research should be used. Untested means exploration only, not decisions.

          Do TasteNET twins make things up?

          Rarely, and it is measured. In case validation, 2 to 8% of answers contained details with no basis in the twin's profile, falling to 2% with first-person data. Separately, 0.4 to 3.7% of single answers contradicted the profile. For this reason TasteNET never uses one sample for a decision: production takes the majority of several samples.

          Can TasteNET surface motives consumers cannot put into words?

          Partly, and TasteNET marks which part is evidenced. Stated preferences are measured directly. Motives a person recognises once shown, and motives that stay unspoken, need further evidence: self-recognition, witnesses who know the person, and behavioural prediction across situations. Digitaste is adding eye tracking and EEG studies as evidence for unspoken motives.

          Is TasteNET validated with academic research?

          Yes. Phase I of the validation was completed with a university partner, and follow-up studies run with further academic partners. Offline, a restaurant sandbox with a university lab compares twins with actual ordering records, and the Singapore Crossmodel Sensory Lab opens in early 2027 for controlled cross-modal tests.

          Can I see the full TasteNET validation report?

          Yes. Digitaste shares the validation report on request. It contains the full benchmark list, the method and the results.