# Writing a research study

Write a study your twins answer honestly: start from the decision, pick a type for each question, and test your twins with a backtest.

Source: https://point11.ai/docs/campaigns/research-studies  
Updated: 2026-10-29

## Start from the decision

Before you write a question, name what you will do differently depending on the answer: change a price, build one feature before another, keep or drop a plan. Every question should move that decision, so cut any that only satisfy curiosity. Most studies need three to six questions. Measure first, with a choice or a rating, then ask why.

**Design with Alpha** writes a first draft from your brief and the customers you chose. Read every question before you start the study: you can reword, reorder, add, or remove any of them.

## Pick a type for each question

| Type | Use it to | Avoid |
| --- | --- | --- |
| Multiple choice | Force the real trade-off between the options a customer has, keeping what they do today included | Options that overlap or leave someone out |
| Checkboxes | See everything that applies, then ask which one matters most | Measuring priority, since everything gets ticked |
| Linear scale | Measure how likely, how satisfied, or how much, with every point labelled | Agree or disagree statements no decision depends on |
| Matrix | Rate a few items on the same scale, side by side | More than five or six rows |
| Ranking | Put a short list in order | Ranking more than five or six items |
| Paragraph | Hear the reason behind an answer, in the customer's words | "Any other comments?" |
| Computer task | See whether a twin can do something on a real website | A task with no clear sign of success |

## Keep the answers honest

- **Give a way out.** Choose **Add “Not sure”** under the options. Results list it after the real answers.
- **Let customers name something else.** Choose **Add “Other”** to add a write-in option that captures what they name there.
- **Shuffle option order** on a list with no natural order, so no option wins just by coming first: each twin sees its own order. Leave it off for an ordered list, such as how often or price bands. *Other* and *Not sure* stay in place.
- **Ask before the subject is shown** when you want unprompted recall: the twin answers before it sees your brand or your material.
- **Add customer context** for facts several questions share, such as a price or the situation. Write one short fact per line: each line shows as a bullet point.

Each twin is told to answer only from its customer's profile and what your study shows it, so "I don't know" and "that doesn't apply to me" are real answers worth reading.

## Test your twins with a backtest

A backtest asks your twins about a decision your business already made, so you can see how close they come to what really happened. Choose **Backtesting** as the type when you create a study. Describe the situation as your customers faced it, the day the decision was made, and what happened after it, in your own words. Point11 reads each metric that moved, and by how much, from what you write.

Under **Results**, each outcome you observed sits beside what the twins predicted and their margin of error, which narrows as more twins answer. Under **Calibration**, each run lists the model that played the twins, which twins answered, and how close it came; choose one to show its figures. To try again, pick another model or other customers' twins and **Run** the same test, or **Find the best model** to try every model on the same twins at once. The study itself sits last, closed until you open it.