Artixol
Methodology

What this can and cannot tell you.

Synthetic research is useful when it is used for what it is. This page is how we describe it.

In short

Artixol results are experimental model output. They are best at ranking options and finding objections, and they are not a census, a forecast or a substitute for field research.

How personas are built

You write a brief. A language model generates personas from it, with a role, personality, background, languages and attitudes. If the brief names a country or city, names and languages follow that culture; otherwise the default is the United States. First and last names come from the same community. Places and relationships are only added if you turn them on.

How studies run

Each persona answers individually, from its own card and memories. Instruments use fixed answer scales (for example yes to no, or use / share / ignore / skeptical) with a reason in the persona's voice. Message tests can fetch the stimulus URL, and Ask can use web sources when the question needs live facts.

How memory works

When a study or shock runs, each persona records episodic and semantic memories and updates topic attitudes. Later studies see them, unless you clone with fresh minds. We do not retrain a model per persona.

Known limits

  • Model bias. Personas inherit the biases and blind spots of the underlying language model.
  • Not a representative sample. A staffed audience is not drawn from a population frame. Percentages describe the staffed audience, not a market.
  • Stated, not observed. Answers are generated, so they say nothing about real behavior such as willingness to pay.
  • Brief sensitivity. Different wording of a brief produces different audiences. Rerun with variations before trusting a result.
  • Sample size. Audiences are capped at 200 personas today.

A checklist before you act on a result

  1. Is the decision reversible? If not, field a human study.
  2. Does the result hold when you reword the brief and rerun?
  3. Do the reasons make sense when you read individual personas?
  4. Would five real conversations confirm it?

How we suggest presenting results

Say you ran a living sample, not a chatbot. Share the split and the export. Share the reverse from Lab. And say which people you would still recruit in the field, because the model only tells you where to look. Do not present the percentages as the market.

Common questions

How accurate are the results?

Artixol results are experimental model output, not a census or a forecast. They are best used to find where to look, which messages to test with real people and which segments to recruit. See the methodology page for limits.

Can I replace my survey panel with Artixol?

No. Use Artixol to sharpen the questionnaire, narrow options and decide where to spend a field study. Keep human research for decisions that need real evidence.

Do personas just agree with everything?

Personas are prompted to answer from their own job, daily life, languages and memories, and replies must vary by person. We also record structured choices, which makes disagreement visible in the splits.

Where do the personas come from?

They are generated from your brief by a language model. If you name a country or city, names and languages follow that culture. If you do not, the default is the United States, including coherent migrant communities.

Questions about the method?

We are happy to walk through how a study is built and where it is weak.