From a question to a study.

Use the hosted web app to draft a survey, explore possible answers, and ask real people. Start with an example, then make the study your own.

Priya is a fictional researcher exploring how customers think about a pricing change. Her example opens without login and includes all features, all six question formats, and a manually filled example. Its answers are illustrative, not customer evidence.

Start in the web app

Work with a shared workspace

Open Workspaces in the web app to create a workspace or choose one you belong to. Check the selected workspace before saving a study. Switching workspaces can reload the studio; resolve any unsaved-draft warning first.

To invite someone, ask them to copy their current account ID from Workspaces. Use that exact ID and choose a role, then privately share the one-use invitation link. There is no invitation email service. The recipient signs into that same account, reviews the invitation, and explicitly accepts it; opening a link alone does not join a workspace. An invitation code can also be pasted into the acceptance form.

Viewers can read. Editors can create and edit studies; opening collection and spending on AI require the owner’s member-spending permission. Owners can spend and manage settings. Administrators can manage editor and viewer membership; workspace-name changes are owner-only in this interface. Membership does not mean real-time co-editing.

If company sign-in is required, choose the workspace’s verification action before accessing its contents. Your workspace operator arranges company sign-in setup; it cannot be configured by entering an arbitrary organization ID here. The owner can require it only after configuration and verification. This is not a promise of production SSO certification.

What belongs in a study?

• Survey: your research purpose, question bank, and the context needed to understand it.
• Question: one thing you want to learn, with an answer format and a required or optional setting.
• Persona: a hypothetical perspective for a simulation. Human respondents answer as themselves.
• Model: an AI option available in your workspace for simulated responses or connected followups.
• Interview: a person’s session answering your survey, one question at a time.

A simulation combines authored questions, personas, and selected models. Adding more of any of them increases the work and potential AI cost.

Choose how people answer

The designer has six question formats. Choose the one that fits what you need to learn:

• Open text: let someone explain in their own words.
• Single choice: ask for one option from your list.
• Multiple selections: let someone choose several relevant options.
• Ranking: ask someone to order the options. Move controls support keyboard use.
• Scale: ask for a rating between the bounds you set.
• Dynamic question: begin with an authored question, then explore the answer through a bounded followup thread.

For choice questions, start with a preset, edit options individually, or bulk-paste one option per line and apply the list. Review the choices and their order. Mark a question optional when someone should be able to skip it.

Explore an answer with followups

Choose Dynamic question, write an opening question, and describe a research goal in 1–1,000 characters. For example: “Understand what makes a monthly bill feel unpredictable.” Set the followup cap from 1 to 5. This limits extra questions; it does not promise that every interview will use them all. Respondents can end the thread.

Simulations answer the authored opening question. Human followups appear in Discoveries with their wording, separately from the fixed-question comparison. A useful discovery can inform a later study where everyone receives the same question. Local mode uses labelled templates; live mode needs configured Not Organic access.

Build in conversation

Describe the decision you are trying to make and ask for question proposals. Refine them, then explicitly append the ones you want. Proposals do not silently replace your existing question bank. Review wording, options, and required settings before saving.

Preview without collecting answers

Use Try respondent preview to test the questions and controls. Preview does not create a saved interview or add answers to results. Dynamic preview followups are illustrative templates, not live AI.

Add an image, video, or audio clip

Add an already hosted HTTPS media URL and provide a description or transcript. File uploads are not available yet. Use media you have permission to share and make sure respondents can access it.

Respondents choose Load before external media is requested. Loading connects their browser to that media host. Test the URL in preview and provide a useful text equivalent for anyone who cannot load, see, or hear it.

Simulations require the text equivalent. Do not assume a simulated answer means the model saw the image, watched the video, or heard the audio.

Offer an AI voice interview

In Deploy → AI voice interviews, a member with both editing and inference-spending permission can enable paid voice for that study. If server setup is unavailable, the control stays disabled; ask your workspace operator to complete setup. The owner may need to reconnect their Not Organic account to authorize voice and accept Not Organic’s realtime logging consent.

Review the budget ceiling per connection and the connection limit for a rolling 24-hour window before enabling voice. The displayed budget bound multiplies those two limits; it is not a bill or a midnight reset. Each interview allows at most 3 voice connections, including reconnects. Respondent use can incur real charges against the owner’s Not Organic wallet. Disabling voice stops active voice connections.

Respondents explicitly opt in to AI voice and microphone access. Voice is processed remotely through Not Organic and OpenAI. Typed answers and on-screen choices remain available. Speech never submits an answer automatically: review or edit it, then submit. Optional browser dictation may use the browser’s own speech service and varies by device.

Confirmed voice seconds may exclude connections whose usage has not been finalized. Use the owner’s Not Organic wallet for actual settlement; the study’s counters are not exact billing. A visible control does not establish a verified live account or paid connection.

Record only with separate consent

Recording is a separate, explicit local opt-in, not a consequence of enabling voice. When supported, respondents can download audio and a timeline to their own device. Voxq does not upload or store that recording or persist the voice transcript on its server. Intentionally submitted survey answers are still saved normally.

Local recording does not make an AI voice conversation local-only: Not Organic and OpenAI still process the live voice remotely. Browser microphone, playback, and download support can vary, especially inside a website embed.

Choose perspectives to investigate

Give each persona a clear label, an age range where relevant, a description, and decision context. Focus on facts that matter: what they use, what they have tried, and what tradeoffs they face.

For Priya’s study, someone paying for their own tools and a team lead responsible for several subscriptions offer different hypotheses to explore. Neither stands in for actual customers.

Reuse an audience

Open Audience & models → Reusable audiences. Give the current audience a name and save it to the selected workspace. Saving does not run models. To reuse one, select it and confirm Replace draft audience. This copies personas and population targets into the unsaved draft; questions, product context, and models stay unchanged.

A human screener belongs to a particular study and is not saved in the library. Configure it again after loading an audience. The copy is independent of the library, so later changes do not silently rewrite the study.

Declare your population assumptions

In Population targets, enable weighting and set a percentage for each uniquely labelled persona. Use mutually exclusive segments. Shares must total 100%; edits are not automatically normalized. Add a source with population and date, or explicitly state that the mix is an assumption. Resetting to equal shares clears the old source so you can document the new assumption.

To weight human answers, choose an existing required single-choice question as a screener and map every option to a persona segment. Voxq does not infer someone’s demographic identity from their writing. Renaming personas, removing choices, or changing the screener can require repairing the targets before saving.

In Results → Population weighting, switch between Human and Predicted to inspect raw and weighted results, coverage, and excluded answers. If a positive-target segment has no usable answers for a question, its weighted result is withheld, not treated as zero. Human effective sample size is a weighting diagnostic, not confidence or a margin of error. These declared weights are not learned population calibration or proof of representativeness.

Choose available AI options

Select model tiers offered in your workspace. Options can produce different answers and costs. Start with a small run, read the responses, and add another model when the comparison helps your research.

Local mode uses deterministic demo responses and labelled followup templates, with no live model requests. Connected mode requires a configured Not Organic account and available model access. A model name in the interface does not by itself confirm that live requests can run.

Explore possible answers

In Simulate, check the questions, personas, and models, then review the estimate before running. Read generated responses for missing options, confusing wording, and assumptions to investigate with people.

Review failures or unavailable responses rather than treating them as answers. Demo responses help you learn the workflow, not evaluate prediction accuracy. If you change the study, keep track of which version produced the responses you compare.

Read what people actually said

Open Results to inspect individual answers and summaries for each format. Compare human and simulated answers to authored questions, checking each group’s sample size and reading explanations alongside totals.

Saved answers from incomplete interviews may contribute to a question’s results. Check how many people answered that question. Review dynamic threads in Discoveries; people may have received different followups.

Open Export results in Results, Discoveries, or Costs. Choose Prepare export, then Download CSV or Download JSON. Both contain the same prepared records and fields: study details, answers, skips, discoveries, comparisons, population details, and recorded inference costs. Use Refresh export to include newer data. A manually filled example demonstrates the views; it is not evidence from recruited respondents.

In a spreadsheet, filter record_type to separate answers from metadata. Structured details are JSON text in the details column in both formats. CSV protects formula-like text; JSON preserves original text. Null costs are unknown, not free. Answers may identify people even though access tokens and external respondent IDs are excluded. Review before sharing. Voice recordings and voice wallet charges are not included.

Share the study with people

1. Save your study and review it in respondent preview.
2. Open Deploy and enable collection when you are ready for real answers.
3. Copy the interview link, or copy the snippet to embed the survey on your website.
4. Read incoming answers in Results. Close collection to stop further answers; previously saved answers remain.

Respondents do not need a researcher account. Explain why you are asking, how you will use answers, and what people should avoid sharing. Do not describe a study as anonymous merely because login is unnecessary: its questions can still collect identifying details.

A submission through a published link is real collection, including your own test answers. Use preview to try the controls without saving data.

Know when AI is being used

Review the estimate before a simulation and recorded usage in Costs. An estimate is a forecast, not a hard spending cap. If a cost is unavailable, do not read it as zero.

Basic human answers do not themselves require AI inference. Simulations, drafting, connected adaptive followups, and opted-in AI voice can use paid AI. Local demo responses and templates make no live model requests; live voice is unavailable in local demo mode.

Research, Team, and deployment pricing describes proposed plans and allowances. This guide and the app’s estimates do not establish live subscription collection, monthly credit resets, top-ups, or limit enforcement.

Use simulations to improve the research

Model agreement is not proof of accuracy, demand, willingness to pay, or demographic representativeness. Recruit people who can inform the decision, inspect disagreement, and keep the limits of your sample visible.

Media descriptions are not model perception. Template followups are not live inference. An example study is not customer evidence. Keep these distinctions clear when explaining findings to a team.

Building an integration or managing your own deployment? Use the developer setup guide.