Understand the method

Fictional people, individually evaluated

Each dot represents a distinct fictional profile with interests, professional background, experience, brand familiarity, purchase intent and priorities. Every completed dot has its own Groq request and validated response. Profiles are designed assumptions, not a sample recruited from a population.

Comparisons with matched conditions

Two ideas use the same saved profiles, prompt version, model and settings. Each request sees only one creative. Dispatch order alternates across profiles to reduce systematic timing differences. Two communities use one creative with different audience configurations. Model outputs remain correlated and can reproduce model biases; these are not independent human observations.

What the colours mean

Green (+) is a favourable reaction. Red (−) is an unfavourable reaction. Orange (?) is mixed, uncertain or needs more information. Grey is waiting; a cross indicates a failed or cancelled response. Counts and a text table provide an alternative to colour.

Reading an image

Static JPEG, PNG and WebP images are safely decoded, stripped of metadata and resized to fit 1600 × 1600 pixels before the actual image is supplied to Groq's Qwen vision model for each response. This may reduce very small text. When either idea contains an image, both arms use the same vision-capable model. Uploading alone does not mean analysis happened.

AI Research Setup

Auto uses a separate Groq planning request to recommend a method, flag question bias, suggest an exploratory usage tier and identify missing audience details. It produces advice, not research results. Approve the recommendation or choose Independent Audience explicitly. Setup analyses the question and supplied text; images are evaluated during the simulation. The required trial credits are shown before starting.

Automatic recovery

Recoverable provider failures retry automatically with bounded backoff while retaining each original profile, model and completed response. Recovery uses no additional trial credits. Invalid requests stop safely; if the response set cannot finish, the report remains incomplete.

Question Bias Check

A separate model request can suggest neutral evaluation instructions. You approve or reject its suggestion. A single-test draft can compare original and neutral instructions using the same creative. This is a test of question wording, not a second advertisement.

What these results cannot tell you

Approval percentages are proportions of generated responses. They are not estimates of human purchase probability, conversion rate, market size or statistical significance. More simulated profiles do not correct a shared model bias. Use the findings to improve questions and identify objections for subsequent interviews, surveys or real experiments.

Interactive Community

Interactive Community is a restricted text-only OASIS beta. Each participant gives an initial opinion, takes part in two private discussion rounds and then gives a final opinion. One social action is allowed per turn. Participants see their initial opinion and the latest up to eight comments within a 6 KB window; the report records exactly what was visible. They may abstain or keep their opinion unchanged. Two-community comparisons use separate social databases, matched profile seeds and speaking order. No conversation crosses between communities.

Only the administrator and explicitly authorised beta accounts can start interactive jobs after operational verification. Ordinary beta accounts begin with five participants. Larger sizes are administrator-only and remain limited to sizes actually verified. Images stay with the independent engine. A larger single-community test does not establish that a two-community comparison of the same size is supported. Availability follows the configurations actually verified. No demonstrated prediction of human behaviour is claimed.