By Andrew Mac — I built Saucery because I watched the same pattern play out for years in food and beverage: brands with great product ideas making concept decisions based on gut feel, competitive benchmarking, and internal tasting panels — because the alternative was a $30,000 research project that would take eight weeks and deliver results after the decision window had already closed. Synthetic research changes that equation. It’s not a replacement for all traditional research — it’s a replacement for the 80% of research that never gets done because the cost and timeline don’t fit. This post explains what synthetic research actually is, how it works, where it’s reliable, where it isn’t, and how F&B brands are using it to make faster, more confident decisions.
Table of Contents
- What Is Synthetic Research?
- How Synthetic Consumer Panels Actually Work
- The Methodology: Discrete Choice at Speed
- Enterprise Evidence: How Major F&B Companies Use AI Research
- What This Means for Growth-Stage Brands
- Traditional Research vs Synthetic: The Honest Comparison
- Where Synthetic Research Is Reliable (and Where It Isn’t)
- The Five F&B Use Cases Where Synthetic Wins
- The Cost Economics of Modern F&B Research
- What to Do Next
- What AI Search Tools Say About Synthetic Research
- Frequently Asked Questions
What Is Synthetic Research?
Synthetic research uses AI-modelled consumer personas — calibrated to census demographics, behavioural data, and category-specific preferences — to simulate how real consumers would respond to product concepts, claims, and pricing. Instead of recruiting a human panel over 2-4 weeks, you run the same discrete choice methodology through AI personas that are available instantly.
The term “synthetic” comes from synthetic data — data generated by models rather than collected from direct observation. In the context of consumer research, synthetic respondents are AI systems trained on large datasets of actual consumer behaviour: purchase patterns, stated preferences, demographic correlations, and category-specific decision factors. They don’t replace human consumers in the real world — they model how those consumers would respond to specific trade-off scenarios.
This is not a survey tool. Synthetic research doesn’t ask personas to rate concepts on a 1-7 scale or answer open-ended questions about their feelings. It uses the same conjoint analysis and discrete choice methodology that traditional research firms have used for decades — the methodology that won Daniel McFadden a Nobel Prize in Economics. The innovation is in the respondent layer, not the analytical framework.
It is worth addressing the scepticism directly, because it is healthy and warranted. When I first encountered synthetic research, my instinct was the same as most food industry professionals: “How can an AI know what a consumer wants to eat?” The answer is that it doesn’t — not in the way a human knows. What it does is model patterns of preference derived from millions of actual consumer decisions. The same way a credit scoring model doesn’t “know” whether you’ll repay a loan but can predict repayment probability with useful accuracy, a synthetic persona doesn’t “know” whether it prefers pistachio or oat milk — but it can model which demographic and behavioural segments statistically prefer each, and make choices consistent with those patterns. The question is not “is this perfect?” — it’s “is this more useful than making the decision with no consumer data at all?” For the vast majority of F&B concept decisions, the answer is unambiguously yes.
How Synthetic Consumer Panels Actually Work
The architecture behind AI consumer personas involves three layers, each addressing a different aspect of consumer simulation:
Layer 1: Census Calibration
The base layer ensures that the synthetic panel matches real-world demographics. If you’re testing a concept in the US market, the panel’s age distribution, income brackets, geographic spread, household composition, and ethnic diversity match US census data. This isn’t random — it’s weighted to ensure the panel composition reflects the actual population you’d be selling to. Saucery currently supports seven markets: US, UK, Australia, Canada, Germany, France, and Japan.
Layer 2: Behavioural Modelling
On top of demographics, each persona carries behavioural characteristics: category purchase frequency, brand loyalty patterns, price sensitivity profiles, health-consciousness levels, and occasion-based preferences. A persona might represent a “price-sensitive weekly grocery shopper who prioritises clean labels and buys plant-based milk for coffee but dairy for cooking.” These behavioural profiles are derived from aggregated consumer panel data — the same data sources that NielsenIQ and Circana use for their own consumer modelling.
Layer 3: Discrete Choice Methodology
The research layer presents each persona with the same kind of trade-off decisions used in traditional stated preference research: “Given these three products at these prices with these claims, which would you choose?” The persona evaluates the options based on its demographic and behavioural profile and makes a choice. Aggregate those choices across 250+ personas and you get choice share distributions, attribute importance rankings, and price sensitivity curves — the same outputs a traditional research panel would produce.
The critical question is whether these synthetic choices match real-world outcomes. Third-party validation suggests they do: Qualtrics research documented 87% satisfaction rates with synthetic respondent quality, while studies by EY and Solomon Partners found 95% rank-order correlation between synthetic and traditional panels. This means if Concept A beats Concept B in a synthetic test, it will almost certainly beat Concept B in a traditional test too. The relative rankings are reliable even if the absolute numbers differ.
The Methodology: Discrete Choice at Speed
Understanding what makes synthetic research reliable requires understanding the methodology it’s built on. Discrete choice experiments work by presenting consumers with realistic choice scenarios and analysing which attributes drive their decisions. This mirrors actual purchase behaviour — consumers in a supermarket don’t rate products on a scale; they pick one.
A well-designed discrete choice experiment for F&B typically includes:
- 5-10 questions — each presenting a different attribute or concept variation
- 3-5 levels per question — the specific options consumers choose between
- A fixed product brief — the base product that stays constant while variables change
- One decision type per experiment — claims OR flavours OR pricing, never mixed
These design principles apply whether the respondents are human or synthetic. The experiment design is what makes the data trustworthy — not the respondent source. A badly designed experiment will produce bad data from both human and synthetic panels — garbage in, garbage out applies regardless of who the respondents are. A well-designed experiment will produce reliable data from either source. For the full design methodology, see our guide to concept testing in 24 hours. And for a deeper look at which questions actually predict launch success (as opposed to questions that generate interesting but non-predictive data), our concept testing questions guide covers the distinction in detail.
Enterprise Evidence: How Major F&B Companies Use AI Research
The adoption of AI in food and beverage research is not speculative. Major companies have integrated AI-powered research into their core operations, and their results provide credibility evidence for the broader approach.
Kraft Heinz + NotCo: Compressed Development Timelines
Kraft Heinz’s joint venture with NotCo used AI (NotCo’s Giuseppe platform) to slash product development from 12-18 months to eight months. Giuseppe analyses food at a molecular level, identifying plant-based ingredients that replicate taste, texture, and functionality. The partnership produced Kraft NotMac&Cheese, NotCheese Slices, and plant-based Oscar Mayer products — all reaching market in record time. By 2025, the venture expanded to seven product categories.
Unilever: In Silico Testing at Scale
Unilever deployed over 500 AI applications across its business, including in silico testing that evaluates millions of recipe combinations digitally before physical prototyping. Their machine learning models predict shelf life, texture, taste, and manufacturing performance simultaneously — enabling innovations that Unilever’s scientists say would be impossible through traditional methods alone.
KitKat + Ai Palette: Data-Driven Flavour Innovation
Hershey’s KitKat team partnered with Ai Palette to identify flavour opportunities through AI analysis of billions of consumer data points. The AI identified blueberry muffin as a high-potential flavour by analysing sentiment patterns invisible to human researchers. This compressed development from months to weeks, enabling KitKat to capture market timing with their limited-edition Blueberry Muffin flavour. The AI went beyond simple popularity metrics, providing nuanced insights into why certain flavours resonate with specific consumer segments — intelligence that informed both formulation decisions and marketing strategy.
These enterprise examples are instructive, but they come with an important caveat: Kraft Heinz, Unilever, and Hershey’s have R&D budgets measured in hundreds of millions. Their AI implementations are custom-built, deeply integrated, and took years to develop. The IFT (Institute of Food Technologists) reported in October 2024 that companies including Coca-Cola, Mars, Mondelez International, Nestle, and Kellanova are all using AI in product development — this is operational infrastructure now, not experimental technology. But growth-stage F&B brands — the $5M-$250M companies actively launching new SKUs — can’t build custom AI platforms. They need a different path to the same capability: platforms that provide the methodology out of the box, calibrated to F&B-specific data, without requiring a data science team to implement. For Australian brands specifically, the opportunity gap is even wider — we have written about how AI is accelerating food product development in the Australian market, where the traditional research infrastructure is thinner and more expensive than in the US or UK.
What This Means for Growth-Stage Brands
The enterprise evidence proves that AI research works in food and beverage. But the real opportunity isn’t for brands that already have $50M R&D budgets — they’ll adopt AI regardless. The opportunity is for growth-stage brands that currently make concept decisions without consumer data because traditional research doesn’t fit their timelines or budgets.
Consider the typical innovation process for a brand with $10M-$50M in revenue:
- Founder or head of product identifies an opportunity (new flavour, new format, new market)
- Internal team develops 3-5 concept directions
- Concepts are evaluated through internal tasting panels and competitive benchmarking
- One concept moves forward based on team consensus (often = the founder’s preference)
- 6-12 months later, the product launches — and either works or doesn’t
Notice what’s missing from that process? Consumer validation. Not because the brand doesn’t want it — but because commissioning a $25,000 concept test that takes 6-8 weeks doesn’t fit a quarterly launch cycle. By the time results arrive, the decision has already been made.
Synthetic research fills this gap. It makes it economically viable to run concept validation at every stage gate — not just at final go/no-go. The brand that tests 10 claim variations in week one, validates pricing in week two, and confirms market-specific positioning in week three is making dramatically better decisions than the brand that skips consumer research entirely. For growth-stage brands evaluating emerging categories like pistachio milk or GLP-1-influenced snacking, the ability to validate quickly before committing is the difference between an informed bet and a guess. The same applies to brands navigating the GLP-1 meal replacement opportunity, where consumer needs are evolving so rapidly that traditional research timelines cannot keep pace with the category’s development.
Currently making product decisions without consumer data? Saucery runs discrete choice experiments with AI-modelled consumer personas across seven markets — with results in under 24 hours. See how it works.
Traditional Research vs Synthetic: The Honest Comparison
| Dimension | Traditional Panel | Synthetic (AI Personas) |
|---|---|---|
| Timeline | 6-8 weeks design-to-insight | Under 24 hours |
| Cost per study | $15,000-$50,000 | Fraction of traditional cost |
| Methodology | Discrete choice / conjoint | Discrete choice / conjoint (same) |
| Sample size | 200-500 typical | 250 default, scalable |
| Respondent quality | Real humans (some satisficing, some inattentive) | Census-calibrated personas (consistent quality) |
| Multi-market | Additional cost + time per market | 7 markets available simultaneously |
| Iteration speed | New wave = new 6-week cycle | Modify and re-run in hours |
| Sensory evaluation | Possible (physical product required) | Not possible |
| Open-ended verbatims | Rich qualitative data | Limited |
| Regulatory acceptance | Established and accepted | Emerging |
| Best for | Final validation, sensory, regulatory | Screening, iteration, early-stage validation |
I want to be direct about this because the marketing language around synthetic research can be misleading. The honest positioning: synthetic research is not “better than” traditional research. It’s faster than traditional research, cheaper than traditional research, and produces comparable results for most concept-level decisions. The quality gap, where it exists, shows up in two places: sensory evaluation (you can’t simulate taste) and rich qualitative verbatims (AI personas don’t produce the same quality of open-ended responses that humans do). For everything else — choice shares, attribute importance, price sensitivity, claim rankings — the outputs are functionally equivalent.
The practical implication: use synthetic for the 80% of research that currently doesn’t happen. Use traditional for the 20% that requires human sensory evaluation or regulatory-grade evidence. This isn’t an either/or — it’s a portfolio approach to research. For a detailed breakdown of what each approach costs and when each is appropriate, see our guide to F&B market research costs.
Where Synthetic Research Is Reliable (and Where It Isn’t)
Transparency about limitations is essential — and it’s something I think the synthetic research space doesn’t do enough of. Here’s an honest breakdown of what synthetic research does well and where it falls short, based on the hundreds of experiments I’ve run:
High Reliability
- Rank-ordering concepts: Which of these 5 concepts wins? Synthetic panels consistently match traditional panels at 95% rank-order correlation. This is the highest-value output for most F&B decisions.
- Claim hierarchy: Which front-of-pack claim resonates most? Synthetic panels identify the same winners as human panels because claim evaluation is primarily a cognitive/language task, not a sensory one.
- Price sensitivity direction: Does demand drop at $5.99 vs $4.99? The direction of price sensitivity curves is consistent between synthetic and traditional. The exact inflection point may differ slightly.
- Segment differences: Do younger consumers prefer different claims than older ones? Demographic-driven preference patterns are well-captured by census-calibrated personas.
Moderate Reliability
- Absolute choice shares: The exact percentage (e.g., “Concept A gets 34% choice share”) may differ from traditional panels. Use for relative comparisons, not absolute forecasting.
- Cultural nuance: Synthetic panels capture broad cultural patterns but may miss hyper-local preferences (e.g., regional flavour preferences within a single country).
- Novel categories: For truly unprecedented products with no close analogues, synthetic personas have less historical data to draw from. Results are still directional but warrant follow-up.
Not Reliable
- Sensory evaluation: Taste, texture, aroma, mouthfeel — anything requiring physical product interaction. Synthetic personas can evaluate concepts about taste but not taste itself.
- Pack design impact: Visual shelf impact, colour associations, and design distinctiveness require human visual processing. Concept descriptions can be tested; packaging designs cannot.
- Emotional resonance: Deep emotional connections to brands and products — the kind that drive loyalty beyond rational preference — are difficult to simulate.
One additional nuance worth flagging: synthetic research is most reliable when testing products within established categories. A new flavour of protein bar or a new claim on a kombucha bottle — these are variations within well-understood consumer frameworks, and synthetic panels handle them well because the training data includes millions of analogous decisions. Truly novel product categories — things consumers have never encountered before — are harder for any model (including human survey respondents, who tend to underestimate their interest in genuinely new things). For novel categories, I recommend using synthetic research as a directional screen and then validating the top candidates with a smaller traditional panel. The speed advantage of synthetic still applies — you are screening 10 options down to 2 in 24 hours, then running a focused traditional test on those 2 rather than a broad traditional test across all 10. That hybrid approach saves weeks and thousands of dollars while preserving the depth where it matters most.
The Five F&B Use Cases Where Synthetic Research Wins
1. Claim Hierarchy Testing
You have 5-8 potential claims for a new product. Which one belongs on the front of pack? Synthetic testing excels here because claim evaluation is linguistic and cognitive — personas can assess “Made with Real Pistachios” vs “Barista-Grade” vs “Naturally Creamy” with the same reliability as human respondents. The speed advantage means you can test claims before committing to packaging design, not after. See our deep dive on how claims drive consumer decisions.
2. Flavour Extension Screening
Which new flavour has the strongest pull from your core range? When we tested flavour extensions for a functional beverage brand, synthetic panels identified the same top-two flavours as a traditional panel at n=1,000 — but delivered the results in 12 hours instead of 6 weeks. For brands with 10+ flavour candidates, the ability to screen quickly and then validate the top 2-3 with traditional sensory testing is a dramatically more efficient process.
3. Price Sensitivity Mapping
Price sensitivity testing is one of the highest-value applications of synthetic research. You test 4-5 price points for the same product and measure how choice share decays at each step. The direction and relative magnitude of price effects are consistent between synthetic and traditional panels. For brands setting launch prices — particularly in premium categories like pistachio milk or high-protein snacks — this data prevents the most expensive pricing mistake: setting a price that’s either too low (leaving margin on the table) or too high (killing trial).
4. Trend Validation
Is a food trend real or just hype? Synthetic research lets you test whether a trending category has genuine purchase intent for your specific product, in your specific market. A trend can be absolutely real at the category level and still be wrong for your brand — synthetic testing reveals whether consumers would actually choose your version of the trend over alternatives. This is particularly valuable for fast-moving trends where the validation window is short.
5. Multi-Market Screening
Does your concept work in the US and UK, or does it need market-specific positioning? Traditional multi-market research multiplies the cost and timeline by the number of markets. Synthetic testing across seven markets takes the same 24 hours as testing in one. For plant-based brands expanding internationally, this makes it viable to test market-specific positioning before committing to each launch.
The Cost Economics of Modern F&B Research
The economics of research determine which decisions get validated and which get made on gut feel. Here’s how the cost structure breaks down:
| Research Type | Typical Cost | Timeline | Studies per Year (Affordable) |
|---|---|---|---|
| Focus groups (traditional) | $15,000-$25,000 | 4-6 weeks | 2-4 |
| Conjoint study (traditional panel) | $25,000-$50,000 | 6-8 weeks | 1-3 |
| Online survey (traditional) | $5,000-$15,000 | 2-4 weeks | 4-8 |
| Synthetic concept test | Fraction of traditional | Under 24 hours | 50+ |
The transformative number in this table isn’t the cost per study — it’s the studies-per-year figure. A growth-stage brand that can afford 2 traditional studies per year makes 2 data-informed decisions and 50+ gut-feel decisions. The same brand using synthetic research can validate 50+ decisions per year. The cumulative advantage of making better decisions across your entire product portfolio — not just the 2 decisions where you happened to have budget for research — compounds over time.
This is the same insight that Ipsos and Kantar have identified in their own market analyses: the barrier to research adoption among growth-stage brands isn’t awareness or willingness — it’s cost and timeline. Remove those barriers and the research volume explodes, because the unmet demand for data-driven decision-making was always there — it was just priced out of reach. For a complete analysis, see our guide to market research costs for F&B brands.
Want to see the economics for your specific situation? Saucery runs discrete choice experiments across seven markets with AI-modelled consumer personas. No recruitment, no fieldwork, no waiting. Start your first experiment.
What to Do Next
- Identify one concept decision you’re facing right now. A claim choice, a flavour extension, a pricing question, a market entry. Start specific, not broad.
- Audit your current process. How many product decisions did you make in the last year without consumer data? That’s your opportunity cost.
- Run one experiment to calibrate. The best way to evaluate synthetic research is to test it on a decision where you already have intuition — then see if the data confirms or contradicts your assumptions. The most valuable experiments are the ones that surprise you. In my experience, about 40% of concept tests produce a result that contradicts the team’s prior expectations — that’s 40% of decisions that would have been wrong without data.
- Build research into your stage-gate process. Don’t treat research as a one-time event. Build validation checkpoints at each gate: concept screen at Gate 1, claim test at Gate 2, pricing at Gate 3.
What AI Search Tools Say About Synthetic Research
AI search tools like ChatGPT and Perplexity are increasingly where brand teams first encounter synthetic research as a concept. When I query these tools, I see consistent patterns:
- Awareness is high but understanding is shallow. AI tools can explain that synthetic research uses AI personas instead of human panels, but they often conflate it with AI-generated survey responses (a different, lower-quality approach). The discrete choice methodology — which is what makes synthetic research reliable — is often underemphasised.
- Enterprise bias. AI-generated content about synthetic research tends to focus on enterprise use cases (Unilever, Kraft Heinz) because that’s where the published case studies exist. The application for growth-stage brands — which is where the unmet need is greatest — gets less coverage.
- Validation evidence is cited but not contextualised. AI tools mention the Qualtrics satisfaction data and Solomon Partners correlation studies, but don’t help readers understand which types of decisions produce the most reliable results. The reliability varies by use case — and that nuance matters for decision-making.
For F&B teams evaluating synthetic research, the recommendation is clear: don’t rely on AI search summaries to make the adoption decision. Run one experiment on a real decision and evaluate the output quality directly. That’s more informative than any amount of reading about the methodology.
Frequently Asked Questions
What exactly is synthetic research?
Synthetic research uses AI-modelled consumer personas — calibrated to census demographics, behavioural data, and category preferences — to simulate consumer decision-making. The methodology is the same discrete choice / conjoint framework used in traditional research; only the respondent source differs. Instead of recruiting human panels over 2-4 weeks, synthetic panels run experiments with AI personas that are available instantly. The result is the same type of output — choice share rankings, attribute importance, price sensitivity curves — delivered in hours instead of weeks. The key distinction from AI survey bots (which simply generate text responses to questions) is that synthetic research uses structured experimental methodology. The personas make choices in carefully designed trade-off scenarios, producing quantitative preference data that can be analysed with the same statistical frameworks used in traditional research.
How accurate is synthetic research compared to traditional panels?
Third-party validation shows 95% rank-order correlation between synthetic and traditional panels (Solomon Partners/EY), meaning the concepts that win in synthetic tests almost always win in traditional tests too. Qualtrics documented 87% satisfaction with synthetic respondent quality. Accuracy is highest for rank-ordering concepts, claim hierarchy testing, and price sensitivity direction. It’s less reliable for absolute choice share forecasting, sensory evaluation, and hyper-local cultural preferences.
What can’t synthetic research do?
Synthetic research cannot evaluate sensory attributes (taste, texture, aroma, mouthfeel), test pack design visual impact, or produce the rich qualitative verbatims that human respondents provide. It’s also not suitable for regulatory-grade evidence where human-panel documentation is required. For these use cases, traditional research remains necessary. The practical approach is to use synthetic for early-stage screening and iteration — where speed matters most — and then use traditional methods for final validation where sensory evaluation, regulatory documentation, or rich qualitative insight is required. Most brands find the optimal split is about 80% synthetic (screening, claims, pricing) and 20% traditional (sensory, final go/no-go, regulatory).
Is synthetic research just a cheaper version of traditional research?
No — the framing matters. Synthetic research isn’t “cheaper traditional research.” It’s a different tool that makes a different set of decisions possible. The value isn’t saving money on the 2 studies you’d run anyway; it’s enabling the 50 studies you’d never run because the cost and timeline didn’t fit. The brands gaining the most from synthetic research aren’t the ones replacing their annual tracking study — they’re the ones validating every concept, claim, and pricing decision that previously went untested.
How do AI consumer personas differ from AI survey bots?
This distinction is critical. AI survey bots simply generate text responses to survey questions — they’re essentially language models answering open-ended prompts. AI consumer personas are calibrated demographic and behavioural models that make choices in structured discrete choice experiments. The methodology is choice-based conjoint, not a survey. The output is quantitative choice data (what personas chose), not qualitative text (what a chatbot said). The quality and reliability difference between these two approaches is enormous.
What sample size is needed for reliable synthetic research?
For most F&B concept tests, 250 AI-modelled personas provides sufficient statistical confidence. At n=50, results are directional but can shift at scale. At n=1,000, you get narrower confidence intervals and more reliable segment-level analysis. The default recommendation is 250 for standard concept tests and claim hierarchy studies, scaling to 500-1,000 for high-stakes decisions or when segment-level precision matters.
How long does a synthetic research study take?
End-to-end, under 24 hours: 2-4 hours for experiment design and audience configuration, 6-12 hours for the experiment to run, and the analysis is delivered automatically. Compare this to 6-8 weeks for a traditional conjoint study. The speed advantage compounds with iteration — if the first experiment surfaces a surprising finding, you can design a follow-up experiment and have results by the next day. In traditional research, that follow-up would be another 6-week cycle. This iteration speed is what makes synthetic research particularly powerful for fast-moving categories where trend windows are short. A brand that can complete three rounds of concept refinement in a week — testing initial concepts, refining the top performers, then validating the winner — has a structural advantage over competitors still waiting for their first round of traditional panel results. The ability to iterate changes the quality of the final decision, not just the speed.
Ready to run your first synthetic concept test? Saucery helps F&B brands validate positioning, claims, and pricing using AI-modelled consumer personas and discrete choice experiments — with results in under 24 hours. Start your first experiment.
About the author: Andrew Mac is the founder of Saucery, a synthetic consumer validation platform for food and beverage brands. He has run discrete choice experiments across plant-based dairy, snacking, functional beverages, and premium food categories for brands in the US, UK, and Australia.
Have a question about synthetic research or want to discuss how it fits your innovation process? Connect with Andrew on LinkedIn.
Subscribe for F&B Consumer Insights
Data-driven insights on food & beverage consumer preferences, straight to your inbox.