Methodology

1. What AskHundred Is — and Isn't

AskHundred is an AI-powered exploration tool that simulates diverse audience reactions to your questions, ideas, and products. It generates a panel of 100 (or 1,000 for Panel Finder) synthetic personas, each with a unique demographic profile, personality, and background. Each persona independently evaluates your question and provides their perspective.

2. Panel Construction

2.1 Demographic Diversity

Each panel is generated with strict diversity requirements:

• Gender parity: ~50/50 male/female ratio (adjustable for targeted panels)
• Age distribution: 18 to 75 years, following a normal distribution centered on 35 (std dev 12), reflecting real population demographics
• Socioeconomic range: Realistic mix for the target population — executives, professionals, employees, laborers, students, retirees
• Geographic spread: When targeting a country, personas are distributed across major cities, mid-size towns, and rural areas — not concentrated in the capital
• Cultural diversity: Names, occupations, and backgrounds are culturally appropriate for the selected country or region

2.2 Personality Profiles ("Thick Personas")

Each persona goes beyond basic demographics. They have:

• 3 personality traits (specific, not generic — e.g., "detail-oriented perfectionist" not just "careful")
• 2 personal values (e.g., environmental sustainability, financial security, creative freedom)
• Communication style: direct, nuanced, analytical, or emotional — this shapes how they express their opinion
• Background: 1-2 sentences describing their life context, habits, and decision-making influences
• Occupation category: tech, healthcare, education, business, creative, trades, student, retired, service, or agriculture

2.3 Anti-Consensus Mechanisms

Real focus groups don't produce unanimous agreement. Neither do AskHundred panels:

• Temperature groups: 30% of personas are "analytical" (measured, logical), 40% "moderate" (balanced), 30% "emotional" (expressive, instinctive). Each group uses a different AI temperature parameter (0.3, 0.7, 0.9) to generate varied response styles.
• Disruptors: 5-10% of personas are deliberately contrarian, skeptical, or indifferent — mirroring real-world survey participants who disagree, don't care, or misunderstand the question.
• Independent generation: Responses are generated in small batches (5 personas per batch) with no cross-contamination. Persona A cannot influence Persona B's answer.
• Value spectrum: Personas span the full range of values: security-minded, environmentalist, liberal, conservative, progressive, pragmatic, apolitical — in realistic proportion for the target population.

3. Response Generation

3.1 How Personas "Think"

Before answering, each persona's AI model is instructed to consider:

• How their personality traits influence their reaction (a skeptic ≠ an enthusiast)
• What their age and life experience imply (a 68-year-old retiree ≠ a 22-year-old student)
• How the question concretely impacts their professional life
• Their communication style (direct ≠ nuanced ≠ analytical ≠ emotional)
• Their daily habits, neighborhood, family — what touches them in daily life

3.2 Web-Augmented Context

For questions about current events, trending topics, or products that exist in the real world, AskHundred automatically fetches recent web context via a search agent. This context is provided to all personas, but each persona weighs it differently based on their personality — some follow trends, others deliberately go against them.

3.3 Language Handling

Personas respond in the same language as the question. If you ask in French, all 100 personas answer in French with culturally appropriate vocabulary. Names, cities, and occupations are localized to the target country.

4. Quality Assurance

4.1 Content Moderation (Pre-test)

Every question is screened by a moderation agent before execution. Questions involving hate speech, violence, illegal content, or CSAM are rejected. Legitimate sensitive topics (politics, religion, health) are allowed. Uploaded images undergo a separate NSFW screening.

4.2 Quality Gate (Post-test)

After every test, an independent quality evaluation agent scores the results on four dimensions:

• Diversity score (0-100): Are opinions genuinely varied, or is there artificial consensus?
• Coherence score (0-100): Does each response match the persona's profile?
• Realism score (0-100): Do responses sound like real people, or like AI-generated text?
• Overall score (0-100): Combined quality assessment

Individual responses are also flagged if they are off-topic, duplicative, generic, or break character. Flagged responses are retained but annotated for transparency.

4.3 Cross-Model Validation

For A/B tests and multiple-choice questions, results are cross-validated against 4 independent AI models (GPT-4o, Gemini Flash, Mistral Large, Perplexity Sonar). This measures consensus across different AI architectures and flags results where models significantly disagree — an indicator that the outcome may be less reliable.

5. Known Limitations & Biases

We are transparent about the limitations of AI-simulated panels:

• Not real humans: AI personas simulate diversity but cannot replicate lived experience, emotional nuance, or cultural subtlety at the same depth as real people.
• Training data bias: The underlying AI models reflect biases present in their training data. We mitigate this through explicit diversity instructions and cross-model validation, but cannot eliminate all bias.
• Western-centric defaults: When no geographic target is specified, personas may skew toward Western cultural norms. We recommend always specifying a target country or region for more relevant results.
• Novelty sensitivity: AI personas may over- or under-react to truly novel concepts that have no precedent in training data.
• Professional domains: Results for highly specialized domains (medical, legal, financial) should be treated with extra caution, as AI personas cannot replicate the deep expertise and ethical constraints of real professionals.

6. Best Practices for Reliable Results

1. Be specific in your question. "Which logo do you prefer?" is better than "What do you think?"
2. Target your panel. A French-targeted panel about a French product gives more relevant results than a global panel.
3. Use as hypothesis generator. Treat results as "here's what to explore further" not "here's the final answer."
4. Cross-reference with real data. Use AskHundred to narrow down options, then validate winners with real users.
5. Check the quality score. Results with quality scores below 60 should be interpreted with caution.
6. Read the verbatims. Individual persona quotes often contain more insight than aggregate scores.

7. Technical Specifications