Park et al. [1] introduced generative agents as computational agents that use LLMs, memory, reflection, and planning to produce believable individual and social behavior. Argyle et al. [2] argued that LLMs can be conditioned on socio-demographic backstories to simulate human samples, coining the term algorithmic fidelity for the ability of a model to reproduce subgroup response patterns. Aher et al. [3] proposed Turing Experiments for evaluating how LLMs replicate findings from human-subject studies.

These works share a core assumption relevant to the present study: conditioning matters. LLM outputs are not merely one undifferentiated model voice. They can vary meaningfully with context, prompt, agent memory, or demographic framing. The present work extends this assumption into commercial discovery. We ask not whether synthetic personas reproduce known survey results or known interview themes, but whether their open-ended pain statements map to a changing external market.

The closest published prior art is Teutloff’s study of synthetic founders [4]. That work docks human-subject founder interviews against synthetic founder and investor personas using the same interview protocol. It reports convergent, partial, human-only, and synthetic-only themes. The method is important because it treats synthetic output as something to be compared against an external reference rather than accepted directly.

The present work is inverse to that design. We do not dock synthetic personas to a known human response dataset. We elicit open-ended pains first, classify them into locked taxonomies, and then validate the resulting themes against venture-market evidence. Where Teutloff measures fidelity to known responses, we measure discovery against an external commercial environment.

The skeptical position is also central to this paper. NN/g defines synthetic users as AI-generated profiles that attempt to mimic user groups and warns that user research needs real users for most evaluative decisions [6]. ACM Interactions critiques synthetic personas as a potential fallacy in which LLM completions are treated as evidence without human-grounded validation [7]. MeasuringU reviews experiments with synthetic users and points to broader validity concerns [8]. A systematic review of 182 papers on synthetic participants similarly raises concerns about cognitive misalignment, stereotyping, and limits of behavioral simulation [9].

Our design accepts much of this critique. We do not claim that synthetic personas are human subjects, have agency, or can validate demand. We use them as hypothesis generators and require external validation. The paper’s key methodological choice is therefore symmetric validation: every theme that crosses the high-volume threshold is checked against external market data and assigned a status.

Each study used culturally stratified synthetic personas generated inside the TwinSim platform. Each persona was generated from a 16-section cultural seed profile and constrained by a voice-discipline layer intended to prevent register inflation and generic LLM voice. The full seed-generation method and the voice-discipline mechanism are not disclosed in this paper. This omission is deliberate: the paper discloses the experiment, prompts, classification method, validation status definitions, and aggregate results, while reserving the persona construction mechanism for a separate technical report.

The India persona set had previously been externally calibrated against Indian public data sources with a Spearman correlation of 0.839 across selected behavioral and demographic variables. Equivalent calibration work was conducted for the other markets, but the detailed calibration schemas are outside the scope of this paper. This paper should therefore be read as an empirical report on a discovery instrument, not as a full technical disclosure of the persona generator.

The underlying model was DeepSeek V3 via OpenRouter, with Claude Haiku used as fallback for failed generations. DeepSeek V3 is a mixture-of-experts LLM described in the DeepSeek-V3 technical report [5]. The internal study logs treat July 2024 as the model knowledge cutoff used for the post-cutoff validation argument. The studies themselves ran between March and May 2026.

The B2C prompt was issued to 1,258 personas across India, UAE, Australia, and Southeast Asia:

Given your life, which problem have you faced where the problem is not just obvious, it is so common that it is not even considered a problem to be solved? Avoid extremely common problems like better air conditioner. Things like cracked heels or women in India are true unique problems, take that as example and answer.

The B2B prompt was issued to 175 German finance and compliance personas:

Given your professional role, name a problem so common in your daily work that nobody considers it solvable. Avoid generic complaints about software or meetings. Think about the friction you have stopped noticing because it has always been there.

The prompt deliberately avoided product stimuli, examples of possible startups, or solution suggestions. The B2C prompt contained a culturally concrete seed example to push respondents away from generic consumer complaints. The Germany B2B prompt removed the cracked-heels example and instead specified professional friction.

Five studies were run between March and May 2026. Table 1 summarizes the sample composition.

Persona responses were classified using a three-stage hybrid pipeline.

1. 1.

   Regex pass. Theme-specific keyword patterns were defined in English and relevant local languages or romanizations: Hinglish for India, Bahasa Malay and Bahasa Indonesia where relevant, Tagalog, Thai, Arabic romanization for UAE, and German for Germany.

2. 2.

   TF-IDF similarity pass. Responses not captured cleanly by regex were compared to multilingual theme-anchor descriptions using TF-IDF cosine similarity.

3. 3.

   Manual unmatched pass. Unmatched responses were reviewed manually to identify net-new themes. When a new theme was accepted, it was added to the taxonomy and the pipeline was rerun.

After iterative development, each geography’s taxonomy was frozen as Locked Taxonomies v1.0. Future studies can therefore measure new samples against fixed theme definitions. Coverage ranged from 66.0% to 94.2%, meaning that 66.0% to 94.2% of responses matched at least one locked theme. The unmatched tail was not forced into categories.

TF-IDF was used instead of multilingual sentence-transformer embeddings because of local environment constraints during analysis. This is a limitation. It likely reduces recall for semantically equivalent responses expressed with different local phrasing. A replication should rerun the classification layer with modern multilingual embeddings while preserving the locked taxonomy for comparability.

For each high-volume theme, defined as five or more personas except in UAE where the smaller sample used a threshold of four, the author searched for currently funded local startups or commercial actors addressing that exact problem in the relevant market. Validation searches used public funding news, startup databases, company directories, and direct company-site verification. The validation window was March to May 2026.

Each high-volume theme was assigned one of four statuses:

• •

  Validated: multiple funded local competitors or established commercial actors exist for the theme.

• •

  Category-forming: funded local competitors raised meaningful capital or launched after the model cutoff, making direct training-data retrieval less plausible.

• •

  Partial gap: adjacent products exist, but no funded actor clearly owns the specific persona-named framing.

• •

  Unowned: the pain appears commercially interpretable, but no funded local company or clear commercial owner was found.

Validation status assignments were author-judged. The detailed company-to-theme validation log is not reproduced here, because it is a commercial annex and may contain productized opportunity mappings. Aggregate rates, status definitions, prompts, sample summaries, and locked taxonomies are disclosed.

For market $m$, let $H_{m}$ be the set of high-volume themes in that market, $V_{m}$ be the subset classified as Validated, and $C_{m}$ be the subset classified as Category-Forming. We define:

The Discovery Index is not a measure of truth, product-market fit, or market size. It measures the share of persona-surfaced high-volume pain themes that correspond to funded or category-forming venture activity in the same market. A higher value suggests more venture-saturated coverage of the persona-surfaced opportunity space. A lower value suggests more partial-gap or unowned space, assuming the validation process is complete.

Across five studies, 1,433 personas surfaced 212 distinct pain themes. High-volume theme counts ranged from 14 to 30 per study. The Discovery Index ranged from 40% to 79%. Table 2 is the centerpiece comparison.

The values do not move randomly. India B2C returned the highest Discovery Index at 79%. UAE, Australia, and Southeast Asia clustered tightly at 40–43%. Germany B2B sat between those values at 58%. This ordering is consistent with a market-saturation interpretation: a high Discovery Index means that more persona-surfaced themes are already visibly funded, while a lower index means that more high-volume themes remain partially owned or unowned.

This interpretation is suggestive rather than conclusive. The study has one sample per geography and validation was author-judged. Still, the pattern is directionally coherent across culturally distinct markets and across one B2B population type.

India B2C. The India panel contained 385 personas across 15 cultural cohorts. It produced 47 distinct themes, 24 of which crossed the high-volume threshold. Nineteen of those 24 were validated or category-forming, giving a Discovery Index of 79%. Three high-volume themes were classified as category-forming because relevant funded activity occurred after the model cutoff. Coverage was 72.5%.

UAE B2C. The UAE panel contained 103 personas, with 88% expat and 12% Emirati native composition. It produced 36 themes, 14 high-volume themes, and a Discovery Index of 43%. Coverage was 94.2%, the highest among the five studies, likely reflecting both smaller sample size and a more demographically concentrated pool.

Australia B2C. The Australia panel contained 385 personas across Anglo-Australian and minority cultural strata, with metro, regional, and rural-remote representation. It produced 50 themes, 21 high-volume themes, and a Discovery Index of 43%. Coverage was 66.0%.

Southeast Asia B2C. The Southeast Asia panel contained 385 personas from the Philippines, Malaysia, and Thailand. It produced 60 themes, 30 high-volume themes, and a Discovery Index of 40%. Coverage was 66.0%. This study produced the strongest cultural self-stratification signal, discussed below.

Germany B2B. The Germany panel contained 175 finance and compliance professionals. It produced 60 or more raw themes compressed into 19 high-volume themes. Eleven of the 19 were validated, giving a Discovery Index of 58%. Coverage was 84.0%. The result suggests that the method generalizes beyond B2C consumer panels into role-specific professional discovery.

Table 3 reports validation statuses. India has the largest validated/category-forming share. Southeast Asia has the largest unowned count. Germany has a relatively high validated share but also multiple partial gaps in Mittelstand-oriented finance and compliance workflows.

The Southeast Asia study pooled personas from the Philippines, Malaysia, and Thailand. The prompt did not ask for country-specific themes, yet high-volume themes self-stratified by country in culturally coherent ways. Table 4 shows the strongest examples.

This matters because a generic regional prompt such as “name Southeast Asian consumer pain points” would be expected to produce broad tropes unless explicitly instructed to stratify by country. In this study, country-specific themes emerged from persona conditioning and open-ended elicitation. We do not report a controlled direct-prompt baseline in this draft; that comparison should be included in a preregistered replication.

The most commercially interesting repeated pattern was not merely that many themes mapped to funded categories. It was that partial-gap themes often sat inside already-funded categories. Personas named a layer adjacent to the legible layer solved by incumbents.

We describe this as the wrong-layer pattern. In market after market, the funded category solved the visible layer: placement, recruitment, discovery, digitization, workflow software, or compliance tooling. Personas named the lived operational layer: capability, coordination, accountability, physical operations, or cross-system glue. To preserve the commercial value of the detailed hypotheses, this paper reports the pattern at the layer level rather than reproducing the full company-to-theme mapping.

This pattern is compatible with, but not proof of, the claim that culturally grounded persona construction elicits context-specific frictions that direct prompting tends to flatten. It is also compatible with a more conservative interpretation: well-conditioned LLM completions can recombine latent local knowledge into plausible market hypotheses, and external validation is required to separate signal from hallucination.

The Discovery Index should be read as a market-context measurement, not as a model leaderboard. A high score does not mean that the personas are better in that market. It means that a higher share of high-volume persona-surfaced themes can already be matched to funded or category-forming local activity. A lower score does not mean failure. It may mean the market has more unowned or partially owned commercial whitespace.

This interpretation makes the India result especially informative. India produced the highest Discovery Index because many high-volume consumer pains mapped to funded or category-forming activity. UAE, Australia, and Southeast Asia clustered lower, suggesting less saturated coverage of persona-surfaced pain themes. Germany B2B landed in the middle, consistent with a mature but uneven B2B software landscape.

Validation asks whether a known stimulus produces a favorable response. Discovery asks what problem space a population names when no solution is supplied. This difference matters because synthetic users are most vulnerable when they are treated as substitutes for real respondents in decisions that require behavioral evidence. A discovery instrument has a different burden: it must produce candidate hypotheses worth investigating, not final truth.

The method in this paper therefore has a lower but cleaner claim. The personas surfaced themes. Those themes were classified. High-volume themes were checked against external venture-market evidence. The resulting Discovery Index is measurable. The detailed hypotheses still require real-world research, founder interviews, customer discovery, or market testing.

The results do not refute skepticism about synthetic users. They narrow the debate. Critics are right that LLM completions are not people and should not be treated as human evidence. However, this study suggests that synthetic personas may still be useful when three conditions hold: the personas are deeply conditioned, the prompt elicits open-ended discovery rather than evaluative approval, and every output is validated against an external reference.

This shifts synthetic personas from “fake respondents” toward “structured hypothesis generators.” That shift is important. It makes the method less grandiose, but more defensible.

The experiment is reproducible in principle because the prompts, validation status definitions, aggregate results, and locked taxonomies are disclosed. However, three components are withheld or only described at high level: the 16-section seed schema, the voice-discipline mechanism, and the detailed company-to-theme validation log. This is a tradeoff between academic transparency and commercial protection.

The most important next step is independent replication. An ideal replication would use the same prompts, fresh persona samples, the locked taxonomies, multilingual embedding classification, and independent dual-reviewer validation. It would also test a direct-prompt baseline and at least one alternate base model.