Synthetic Research: AI’s Big Promise for Marketers Has a Catch

The synthetic data market is on track to grow from $267 million in 2023 to over $4.6 billion by 2032, and MarTech reports that 95% of insight leaders plan to adopt synthetic data within the next year. The speed advantage is undeniable: AI-generated personas compress research timelines from months to days at a fraction of traditional costs. But according to Greg Kihlstrom, writing for MarTech on April 17, 2026, synthetic research carries a structural problem that most marketing teams are not yet equipped to manage — and deploying it without a validation framework means your strategy is only as reliable as the biases baked into your AI model.

What Happened

On April 17, 2026, MarTech published a deep-dive from Greg Kihlstrom, Principal at The Agile Brand, laying out the current state of AI-driven synthetic research and the structural risks most brands are walking straight into. Kihlstrom frames the core tension plainly: competitive pressure to generate faster, cheaper consumer insights is colliding with the scientific rigor that makes those insights trustworthy. One side is winning right now — and it is not rigor.

Synthetic research, for teams not yet running it, is the practice of using generative AI and large language models to produce simulated consumer data: fabricated personas, synthetic survey responses, AI-generated focus group outputs, and modeled behavioral profiles. Instead of recruiting real respondents, running live focus groups, or fielding quantitative surveys across hundreds or thousands of actual humans, teams feed a large language model a description of the target audience and collect AI outputs as proxies for real consumer responses. The pitch is compelling: generate 5,000 synthetic consumer interviews in an afternoon, A/B test 40 different messaging concepts simultaneously, and iterate on product hypotheses at a speed that traditional research could never match.

The market is betting heavily on this approach. The synthetic data industry is projected to grow from approximately $267 million in 2023 to over $4.6 billion by 2032, according to data cited in Kihlstrom’s MarTech analysis. Adoption intent is even more striking than the market size numbers: 95% of insight leaders report plans to incorporate synthetic data into their research programs within the next twelve months. That is near-universal adoption of a methodology that most organizations have not yet formally stress-tested.

Here is where the catch reveals itself. The most intuitive approach to synthetic research — prompting ChatGPT, Claude, or Gemini with detailed consumer backstories to generate representative personas — actually produces outputs that are less accurate, not more. Kihlstrom cites research showing that more elaborate and detailed persona prompts increase homogeneity in LLM outputs rather than diversity. The model converges on a central, AI-friendly archetype rather than distributing responses across the messy, varied range of real human opinion. This is not a prompting problem you can engineer your way out of with better instructions; it is a structural limitation of how large language models generate text when given constrained role-play parameters.

The failure modes multiply from there. What Kihlstrom describes as “bias laundering” is one of the subtler and more dangerous problems in the synthetic research pipeline. Large language models are trained predominantly on internet-sourced text that skews heavily toward Western, educated, industrialized, rich, and democratically-governed perspectives — what researchers call WEIRD bias. When you generate synthetic consumer data from these models, the outputs inherit those skews but present them with the appearance of objective, neutral research data. Underrepresentation of non-Western consumers, lower-income segments, and non-English-speaking populations is not flagged as a limitation in the AI’s output. It is invisible, masked as representativeness. Your synthetic panel looks balanced on the surface while silently excluding the exact populations your strategy may depend on reaching.

Then there is what Kihlstrom calls the “Pollyanna Principle.” Synthetic AI respondents are trained to be helpful and agreeable. When they simulate consumer behavior, they reflect what they predict the researcher wants to hear — not what real humans would actually do or say. In one usability test cited in Kihlstrom’s analysis, synthetic participants reported completing every online course presented to them. Real human participants in the same test reported dropping out of most. The difference between those two outcomes — if you are building or validating an online learning product — is the difference between a viable business and a market failure. Kihlstrom also cites a GPT model that predicted positive consumer reception for pancake-flavored toothpaste, and a separate case where a base model overestimated willingness-to-pay for laptop projectors by 300%, until each model was fine-tuned against historical real-world survey data.

The underlying problem Kihlstrom names is the “synthetic persona fallacy”: the mistaken belief that prompting an LLM with demographic and psychographic attributes produces a functional equivalent of human psychology. It does not. LLMs generate statistically probable text given their training data. They do not simulate human decision-making, preference formation, or behavioral response under real purchasing conditions.

Why This Matters

The failure modes Kihlstrom describes are not abstract statistical concerns — they map directly to common, high-stakes marketing decisions: pricing thresholds, messaging hierarchy testing, product feature validation, audience segmentation, and persona development. If your team is using synthetic research to power any of these decisions without a validation layer, you are operating on data engineered to confirm your assumptions rather than challenge them.

The exposure differs by role and organizational structure, but no marketing function is immune.

In-house brand and consumer insights teams carry the greatest structural risk. The pressure to cut research costs and compress timelines is a constant reality inside most brand organizations, especially in CPG, retail, and consumer tech. Synthetic research looks like the obvious cost-saving solution, and its outputs look credible enough to present to stakeholders without obvious red flags. But without a researcher with quantitative or qualitative expertise auditing synthetic outputs against real-world benchmarks, AI-generated data travels directly into product roadmap decisions, campaign briefs, and audience strategies. A product concept that “tests well” with synthetic personas but carries a 300% inflated willingness-to-pay estimate is a launch disaster in development. By the time real-world numbers come in post-launch, the production spend, the media budget, and the retail distribution commitments have already been made.

Agencies and research consultancies face a different kind of exposure. If you are packaging synthetic research as a fast-turnaround service and selling it to clients as a credible substitute for primary research without validating the outputs, you are building a methodology liability into client engagements. When a campaign built on synthetic insight underperforms and the research methodology comes under scrutiny, “we ran AI personas against your brief” without a documented validation process is not a defensible position. The speed premium you charged for synthetic research becomes a liability when it cannot be backed by documented accuracy.

Marketing technology teams building internal AI research stacks need to recognize that the tool layer is not the governance layer. Wiring a base LLM into a persona prompt template and calling it a consumer research system generates outputs, not insights. Outputs become insights when they are validated, documented, and stress-tested for the specific failure modes Kihlstrom describes: representational bias, sycophantic response patterns, and factually incorrect predictions on novel products or underrepresented segments.

Solopreneurs and growth teams operating lean on research budgets face a subtler trap. Using ChatGPT to simulate customer discovery interviews feels like a reasonable workaround when full-scale qualitative research is out of budget. In narrow, well-understood product categories with abundant training data, a fine-tuned model can produce workable directional signals. But in novel product categories or with niche consumer segments, the Pollyanna Principle means the AI will confirm your assumptions rather than challenge them — which is precisely the opposite of what research is designed to do. The entrepreneur who validates their new product concept with synthetic AI interviews and receives uniformly positive feedback has not validated their idea. They have confirmed their own wishful thinking at scale.

What this development challenges most fundamentally is the premise that AI democratizes reliable market research. AI democratizes access to research at scale. It does not automatically democratize the accuracy or validity of the outputs. Getting trustworthy synthetic insights requires the same experimental design literacy, statistical thinking, and validation discipline as traditional research — applied to an entirely new set of tools and failure modes that most marketing teams have not yet built protocols to detect or prevent.

The Data

The scale of the synthetic research adoption curve makes the governance gap urgent. The table below maps the methodology across key performance and risk dimensions, drawing on data from Kihlstrom’s MarTech analysis and the Stanford HAI 2026 AI Index Report:

This comparison illustrates the core trade-off Kihlstrom articulates. Unvalidated synthetic research is not a faster version of traditional research — it is a different methodology with fundamentally different failure modes. Validated synthetic research using the Train Synthetic, Test Real (TSTR) methodology closes much of that gap, but only with the additional overhead of calibrating outputs against real human samples and fine-tuning models on domain-specific data.

Two figures from the Stanford HAI 2026 AI Index underscore why getting this right is urgent rather than theoretical: enterprise AI adoption reached 88% in 2025, and the estimated value of generative AI tools to U.S. consumers hit $172 billion annually by early 2026. AI is no longer an experiment inside most organizations — it is operational infrastructure. That means governance failures are not isolated to individual projects; they compound across every decision that relies on AI-generated inputs.

The TSTR methodology, developed in research spearheaded by Stanford and Google DeepMind and cited by Kihlstrom, produced results worth noting: models trained on synthetic data and validated against real-world samples achieved 85% accuracy replication and 98% correlation on social dynamics. That level of fidelity is meaningful and operationally useful — but it is an outcome of the TSTR process applied correctly, not a property of synthetic data inherently. The 85% figure requires investing in the validation infrastructure required to reach it.

Real-World Use Cases

The governance framework Kihlstrom describes is not theoretical — brands are already deploying validated synthetic research pipelines with measurable results. Here are five concrete implementations grounded in approaches covered across MarTech’s synthetic research analysis and Susan Ferrari’s coverage of AI shifts in market research.

Use Case 1: CPG Brand Accelerated Segment Validation

Scenario: A mid-market consumer packaged goods brand is planning a product extension targeting a demographic segment they have limited primary research on. A traditional segmentation study would require three months and a six-figure research budget — neither of which fits the launch timeline.

Implementation: Following the approach Dollar Shave Club used (cited by Kihlstrom), the brand builds a synthetic panel grounded in category-specific purchase behavior data pulled from their CRM and third-party behavioral panel data. Rather than generating personas from a bare LLM prompt with demographic attributes, the team fine-tunes a model on actual category survey responses before running the synthetic population. They then validate the synthetic panel’s output against a live survey fielded to 250 real consumers matched to their target demographic. The measured gap between synthetic and real responses on key preference metrics is used to apply calibration adjustments to the full synthetic output set.

Expected Outcome: Validated synthetic segment profiles delivered in days rather than months, with human-behavior-equivalent accuracy on preference and purchase intention metrics — consistent with Dollar Shave Club’s reported outcomes at a substantially lower research cost. The validation step adds two to three days but provides the calibration required to treat the outputs as decision-grade rather than directional.

Use Case 2: Agency Concept Screening at Scale with Control Calibration

Scenario: A creative agency needs to screen 25 campaign concepts for a retail client before presenting a prioritized shortlist. Fielding real consumer research across 25 concepts is cost-prohibitive; presenting a shortlist without any research backing is not credible to the client.

Implementation: The agency uses synthetic research to screen all 25 concepts, generating 500 synthetic consumer responses per concept using a fine-tuned model. The critical addition is a control calibration layer: three benchmark concepts with known real-world performance history are included in the synthetic screening. If the synthetic panel ranks those control concepts in the same order as their documented historical performance, the panel output is treated as calibrated. If the ranking diverges materially, the model is re-tuned before the remaining concept scores are used. The full process is documented in a methodology brief that accompanies the shortlist presentation.

Expected Outcome: A client-ready shortlist of five concepts backed by directional synthetic data calibrated against real performance benchmarks, delivered in 48 hours rather than the three to four weeks a traditional screening study requires. The agency maintains a documented validation layer that protects both parties if the research methodology is later scrutinized.

Use Case 3: SaaS Pricing Research Using TSTR

Scenario: A B2B SaaS company needs willingness-to-pay data for a new feature tier before launch. Running a Van Westendorp or Gabor-Granger study on their own customer base risks tipping off key accounts before the pricing announcement is ready.

Implementation: The team applies the TSTR methodology directly: they train a synthetic model on historical pricing survey data from comparable feature launches in their category, generate synthetic WTP distributions for the new tier, and validate by running a blind survey of 150 external panel respondents matched to their ideal customer profile against the same pricing questions. Given the documented 300% WTP overestimation risk for novel products cited by Kihlstrom, the team applies a calibration adjustment based on the measured delta between synthetic predictions and real responses in the validation sample before using the data in their pricing model.

Expected Outcome: Validated WTP estimates delivered in two weeks rather than six, with a documented methodology the product and finance teams can rely on for the commercial model. The validation step is treated as non-negotiable — pricing decisions built on uncalibrated synthetic data represent one of the highest-risk applications of the methodology, given the systematic overestimation patterns the research has documented.

Use Case 4: On-Premise AI for Sensitive Customer Feedback Synthesis

Scenario: A financial services firm wants to synthesize three years of customer service conversation transcripts into behavioral personas for their digital product team. The transcripts contain personally identifiable information and cannot be transmitted to an external API under their data security policy.

Implementation: Following the approach described in Ferrari’s MarTech analysis of AI shifts in market research, the team deploys Google’s Gemma models in an on-premise configuration running entirely within their corporate infrastructure. No customer data reaches an external API endpoint. The model synthesizes behavioral patterns and segments from the historical transcripts, generating personas that are validated against a sample of hand-coded transcripts reviewed and confirmed by an internal analyst before the full synthesis is accepted into the insight repository.

Expected Outcome: Rich, behaviorally-grounded customer personas derived from three years of actual service interactions, generated in weeks rather than the quarters a fully manual analysis would require, with full data security compliance and a human expert validation layer anchored in spot-review of the highest-stakes outputs.

Use Case 5: Multi-AI Continuous Sentiment Synthesis

Scenario: A direct-to-consumer e-commerce brand wants to run ongoing consumer sentiment synthesis from social listening and customer feedback data at scale, but cannot staff a research team to manually quality-check every AI output cycle.

Implementation: Building on the multi-agent validation architecture described in Ferrari’s MarTech analysis, the brand deploys a three-model pipeline: Model A generates synthetic consumer sentiment summaries from incoming social and feedback data. Model B independently analyzes the same inputs for sentiment signals and key themes without seeing Model A’s output. Model C compares both outputs, identifies material inconsistencies, and generates a confidence score for each synthesis cycle. Any output where Model C’s confidence score falls below the team’s threshold is automatically routed to a human review queue rather than flowing directly into the insight repository. Only flagged exceptions — typically 10-15% of cycles — require human attention.

Expected Outcome: Continuous consumer sentiment synthesis at the scale the brand needs, with a built-in peer-review validation layer that catches hallucinations and inconsistencies before they reach decision-makers. This mirrors the self-validation architecture Ferrari identifies as one of the defining AI shifts reshaping market research in 2026 — moving AI from isolated output generators to integrated systems with quality control embedded in the workflow rather than bolted on afterward.

The Bigger Picture

The synthetic research governance problem does not exist in isolation. It is one node in a much larger reckoning the industry is navigating between AI capability and AI readiness. The Stanford HAI 2026 AI Index Report frames the macro condition precisely: the report documents “a widening gap between what AI can do and how prepared we are to manage it,” with technical capabilities improving, investment accelerating, and adoption spreading faster than governance infrastructure can follow. Enterprise AI adoption hit 88% in 2025. Generative AI reached 53% population adoption within three years of launch — a faster curve than either PCs or the internet. The tools are everywhere. The operational discipline to deploy them reliably is not.

In market research specifically, this gap has compounding consequences. Research is the input layer for every major marketing decision — brand strategy, product development, pricing, media planning, campaign creative, audience segmentation, competitive positioning. If the research layer produces plausible-looking but systematically biased outputs, the errors multiply across every function consuming those insights. An agreeableness bias in synthetic usability testing does not just corrupt one product decision. It can shape the roadmap, the positioning, the launch investment, and the channel strategy built on top of that research — all of which will then be evaluated against market reality that the synthetic data was never designed to accurately represent.

Susan Ferrari’s April 15, 2026 MarTech analysis of three AI shifts reshaping market research signals that the infrastructure layer is catching up with the governance problem in real time. Anthropic’s Projects feature — which gives AI systems persistent memory across sessions — directly addresses one dimension of synthetic research reliability: context continuity. Instead of generating insights cold from a single prompt with no institutional memory, researchers can now build AI systems that accumulate knowledge across sessions, synthesizing patterns across years of brand tracking data without starting fresh each time. That does not solve the Pollyanna Principle or WEIRD bias on its own, but it substantially reduces the context deficiency problem that makes base LLMs unreliable for longitudinal and complex research programs.

Google’s Gemma on-premise models address the data security and access dimension — unlocking previously off-limits sources like customer service transcripts and PII-containing survey archives for synthetic analysis within corporate security perimeters. Multi-AI validation systems address the hallucination and inconsistency problem through peer-review-style architecture. These are not incremental improvements to existing synthetic research workflows. They are infrastructure changes that make governed, validated synthetic research practically deployable for teams that could not build those guardrails six months ago.

The direction the industry is moving is clear: AI-driven research is becoming core infrastructure, not experimental capability. The teams building governed, validated synthetic research pipelines now will accumulate a durable methodology advantage as the market matures around standards and best practices. The teams treating synthetic research as a cost-cutting shortcut without validation are building a liability that will eventually surface in a product launch, a campaign, or a strategy decision that traces back to data engineered to confirm what the team already believed.

What Smart Marketers Should Do Now

The governance gap in synthetic research is a solvable problem, and the solutions do not require a large team or sophisticated technical infrastructure. Here is how to get ahead of it.

1. Implement TSTR as the non-negotiable default for any decision-grade synthetic research.

Train Synthetic, Test Real is not an advanced enterprise methodology — it is the minimum viable validation framework for synthetic research being used to drive actual decisions. The Stanford and Google DeepMind research cited by Kihlstrom demonstrated 85% accuracy replication and 98% correlation on social dynamics when TSTR is properly executed. The mechanics are straightforward: generate your synthetic data set, then validate its predictions against a real human sample before treating any outputs as decision-grade. The real sample does not need to be large — 100 to 200 respondents matched to your target population is typically sufficient for calibration. Make this a non-negotiable project requirement. If there is no budget or timeline for a validation sample, there is no budget for decision-grade synthetic research — only for directional ideation.

2. Build and enforce a persona transparency checklist before any synthetic data leaves the research function.

Kihlstrom’s governance framework specifies four dimensions that must be documented for every synthetic research output: the application domain (exactly what decisions this data will inform), the target population specification (which consumer segment the synthetic panel was designed to represent and how it was constructed), data provenance (what training or fine-tuning data the model was calibrated against and where that data originated), and ecological validity (whether the synthetic research environment sufficiently mirrors the real-world context being studied). This does not require elaborate documentation infrastructure. A single-page brief per study, completed before the output is distributed, forces the team generating synthetic data to surface the assumptions built into the methodology before those assumptions become invisible inputs to strategy. If any of the four dimensions cannot be clearly documented, the research should not be treated as decision-grade.

3. Stop using base LLMs for synthetic research on novel products or underrepresented consumer segments.

Base models generate outputs that reflect their training distribution — which is dominated by existing, well-represented categories and demographics. For established product categories with abundant training data, a calibrated base model may produce workable directional outputs. For novel products, the documented failure mode is systematic overestimation of willingness-to-pay (300% in the case cited by Kihlstrom) and a default-to-positive reception pattern that does not reflect actual consumer behavior. For underrepresented segments, WEIRD bias means the model’s outputs are approximations built on data that does not represent those populations. In both cases, primary real-world research is not optional — it is the baseline against which synthetic data must be calibrated before it carries any evidentiary weight for strategy decisions.

4. Invest in fine-tuned models built on your own historical data rather than generic persona prompts.

The Dollar Shave Club case study makes the practical gap concrete: synthetic panels grounded in category-specific behavioral data produced human-behavior-equivalent results. Generic demographic persona prompts produced homogeneous, Pollyanna-contaminated outputs. The reliability of your synthetic research is directly proportional to the quality and specificity of the calibration data you feed it. Teams that invest in structuring their own historical research data — past surveys, customer interviews, behavioral analytics, CRM records, transaction histories — and using it to fine-tune or calibrate their synthetic models will get materially more accurate outputs than teams running bare LLM prompts against audience briefs. This is a data infrastructure investment that compounds in value over time, not just a prompt engineering exercise.

5. Deploy multi-AI validation architectures for any continuous or high-volume synthetic research pipeline.

For teams running synthetic research at sustained scale — continuous brand tracking, ongoing concept screening, high-frequency sentiment synthesis — a single-model approach creates a single point of failure for hallucinations and systematic bias. The multi-AI peer-review architecture Ferrari describes in her MarTech analysis — one model generates, a second independently analyzes, a third flags inconsistencies and scores confidence — substantially reduces the probability that any single model’s errors propagate unchecked into the insight layer. Most marketing technology teams can implement a three-model validation pipeline within their existing AI infrastructure in a matter of weeks. For any research output feeding strategy-level decisions, the setup investment is justified by the risk it eliminates — the alternative being a single model running unsupervised and generating outputs with no internal accountability mechanism, precisely the “methodological black box” Kihlstrom warns against.

What to Watch Next

The synthetic research methodology landscape is moving quickly. These are the specific developments worth tracking over the next six to twelve months.

Validated synthetic research platforms. Several market research technology companies are building synthetic data capabilities with TSTR-style validation designed into the product by default rather than requiring teams to build their own validation frameworks. In Q2 and Q3 2026, watch for platform launches or major feature releases that offer built-in real-sample validation as a standard workflow. First movers here will define the market standard for what “validated synthetic research” means as a commercial product category, ahead of any industry standard-setting body.

Industry standards frameworks from MRS and ESOMAR. The Market Research Society and ESOMAR have both signaled intent to develop guidance on AI-generated consumer insights. Watch for formal standards documents or proposed frameworks in Q3 2026 that establish minimum validation requirements for AI-driven research to qualify as primary research data. When those standards arrive, synthetic research governance will shift from a best-practice discussion to a compliance requirement — which will accelerate adoption of validation frameworks faster than any vendor movement could.

Fine-tuned vertical models for specific research domains. As the limitations of base LLMs become more widely understood, expect purpose-built models optimized for specific research contexts: B2B purchase intent simulation, retail consumer behavior modeling, healthcare patient persona generation. These models will be calibrated on domain-specific training data and will reduce some of the WEIRD bias problems that make base models unreliable for niche or non-Western consumer segments. Watch which market research technology firms partner with vertical data providers to build the first category-specific research models — these will set the performance benchmarks the rest of the market will be measured against.

Multi-agent research system benchmarks. The multi-AI validation architecture is still early in its market research deployment. Over the next six months, expect the first publicly available performance benchmarks from brands running multi-agent systems for continuous research synthesis. The critical metric is not output speed but validation catch rate — how frequently does the multi-model system flag errors that a single-model pipeline would have passed through to decision-makers unchecked? That data will determine how broadly the architecture gets adopted.

Persistent AI memory in enterprise research platforms. Anthropic’s Projects feature and comparable persistent memory capabilities from other providers are still being adopted at the enterprise level. As these tools become integrated into established research platform workflows — rather than requiring custom implementation — longitudinal brand tracking and multi-year insight synthesis will become substantially more accessible. The first enterprise research platform to ship persistent AI memory as a native feature will have a meaningful adoption advantage in the insights market.

Bottom Line

Synthetic research is not a shortcut — it is a methodology that delivers real speed and cost advantages when properly governed, and produces dangerously overconfident, bias-contaminated outputs when it is not. Kihlstrom’s analysis makes the structural failure modes explicit: bias laundering, the Pollyanna Principle, and the synthetic persona fallacy are predictable outputs of base LLMs deployed without fine-tuning and validation, not edge cases. The TSTR methodology and persona transparency checklist are practical, implementable solutions available to any team today, not theoretical constructs waiting for future tooling to support them. With 95% of insight leaders planning synthetic data adoption within the next year and enterprise AI already embedded in 88% of organizations according to the Stanford HAI 2026 AI Index, the governance gap is an operational problem that cannot wait for industry standards to arrive. The brands that build validation-first synthetic research pipelines now will carry a durable accuracy and speed advantage into the next phase of the market — and the brands that treat synthetic research as a cost-cutting shortcut will eventually make a strategy-level decision off data engineered to confirm what they already believed.