Why we don't ship synthetic respondents

The category we sat out

Simile raised $100M in February 2026 on the promise of consumer simulation. Quantilope, Listenlabs, Consumr.ai, H-in-Q and Perspective AI are funded and shipping. Toluna calls it "AI-native research." The pitch deck story is good: faster, cheaper, no recruitment friction, 90% alignment with real survey data on structured tasks.

Theia is in the AI market research category. We didn't build a synthetic-respondents product. This piece is about why.

The critique others have already published

We are not the first to raise concerns. The published methodology critique is strong and growing:

• NIQ documented the structural bias in synthetic-respondent training data
• Bellomy published "Synthetics Challenges Market Researchers Should Consider"
• VerianGroup detailed the limitations of AI-generated responses in social research
• Perspective AI wrote "Why Fake Respondents Can't Replace Real Customer Research"
• Phys.org covered academic findings that "GenAI could push consumer research toward generic, biased results"

Five publications. Five different critiques. One missing piece: what's the alternative?

The five failure modes, briefly

For completeness:

01 — Derivative intelligence. A synthetic respondent is exactly as good as the training data behind it. Outside the training distribution — new products, cultural shifts, B2B specialist categories — the model fabricates plausible-sounding answers without knowing what it doesn't know.

02 — Sycophancy. RLHF-trained models are structurally biased toward agreement with the researcher. Multi-turn synthetic interviews drift in the direction the prompt suggests. Sycophancy is a property of how the models are built, not a bug to be patched.

03 — Mode collapse on minorities. Synthetic respondents over-represent the average of training data. Minority preferences, edge-case attitudes, and niche segments collapse into the dominant pattern. For brands trying to find growth at the edges, this is exactly the wrong bias.

04 — Western-context bias. Pricing preferences, willingness to pay and brand-loyalty patterns are deeply cultural. Training data is overweight in English-language Western content. Cross-market consumer behaviour outside the West gets distorted.

05 — Recursive model collapse. As synthetic respondent outputs increasingly enter training data, the bias compounds. Synthetic respondents in 2028 will be partly trained on synthetic respondents from 2026. The signal-to-noise ratio degrades over time without anyone noticing.

The published validation failure

One example for proof of concept. In 2025 academic work testing synthetic respondents against known outcomes, the models predicted 83% electoral participation against an actual 49%. The errors are not symmetric. Better prompts don't reliably reduce them.

For a board-grade brand decision, a CMO presenting "consumers say they'll convert at 35%" derived from synthetic respondents is presenting evidence that is reasonably likely to be wrong by a multiple. That's not a tool problem; that's an architecture problem.

What we built instead

Theia's approach is revealed preference at scale, structured into a graph.

Every claim in a Theia deliverable comes from real consumer behaviour:

• Search behaviour — what consumers actually type into Google and Amazon, clustered into demand pockets
• Purchase signal — sales movements from 1P (Vendor Central) and 3P (Stackline, GfK)
• Voice — reviews, YouTube transcripts, web articles, Reddit, BazaarVoice, AI Overview citations
• Citation — which brands LLMs cite for which queries

This is what consumers do. Not what synthetic personas predict they would do.

The data is processed through native-language extraction (German stays in German until canonical mapping), Fixed Entity Architecture (expert-curated ontology, not LLM-discovered), and math for connections (cosine, Leiden, HHI for graph operations — not LLM judgement).

The result is an intelligence layer with five properties synthetic respondents structurally cannot have:

The honest case where synthetic respondents fit

We are not saying synthetic respondents have no place. Three legitimate use cases:

• Early-stage hypothesis exploration — "what should we even be testing?"
• Well-anchored attitudes in mature, well-documented product categories
• Scenario simulation — counterfactual exploration that no real-data approach can answer

If your decision falls cleanly into one of these three buckets, a synthetic-respondent platform may be the right tool. For everything else — competitive strategy, brand repositioning, PE diligence, regulated services, B2B specialist categories — synthetic respondents are not the answer.

What this means for the buyer

If you are commissioning AI-driven market research in 2026, three questions worth asking any vendor:

1. "What signal is your output derived from — real consumer behaviour or AI-generated personas?"
2. "For each claim in this deck, can you show me the original source URL?"
3. "What's your EU AI Act readiness for the August 2026 deadline?"

Vendors that synthesise outputs cannot answer (1) honestly. Vendors that ship synthetic-respondent outputs cannot answer (2) at all. Vendors that ship live-only outputs (no snapshots, no run-id replay) cannot answer (3).

The shape of the AI market research category over 2026-2027 is going to be defined by how procurement teams answer those three questions. Vendors that answer all three honestly will keep enterprise contracts. Vendors that can't will lose them.

What we'll publish next

Next post: The deterministic harness in AI research — why reproducibility, audit trail and budget assertion are the engineering properties enterprise procurement is about to demand, and what they look like in production.

---

Theia is structured market intelligence — revealed preference, native-language extraction, integrated across four pillars. For consumer brands, industrial B2B, research firms and PE. See the platform or read the methodology.