> ## Documentation Index
> Fetch the complete documentation index at: https://docs.getversive.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Getting the most from AI tests

> Features worth turning on, and how experienced teams run their tests

## Set up personas like a qual study

* **Build personas from real research data.** Upload interview transcripts, survey exports, or research docs and let Versive [distill them](/ai-tests/personas) — a persona grounded in real quotes and numbers behaves far more specifically than one written from imagination.
* **Treat personas like human qualitative research.** You can learn from a sample size of one — a single run will surface real issues. But aim for **3–5 executions per persona** so the aggregated summary can group findings and show you which ones are stable versus one-off.
* **Test with more than one persona.** Results are attributed per persona, so running your skeptical novice alongside your power user shows you where their experiences diverge.

## Scope the test tightly

* **Test small flows and tasks, not open-ended tours.** "Sign up and invite a teammate" produces attributable findings; "explore the app" produces noise. Run several small tests instead of one big one.
* **Keep Figma files clean.** Import only the frames you actually want tested, in the order you want them seen — stray exploration screens and dead-end frames read as usability failures.
* **Have placeholder content? Turn on the content-accuracy flag.** The [*ignore content accuracy*](/ai-tests/figma-and-image-tests#test-configuration) setting tells the AI to disregard lorem ipsum and dummy data and focus on structure and flow — without it, placeholder content gets flagged as confusing.

## Pick the right mode

* **Conversational tests simulate real user interviews.** Use them when you want reactions, expectations, and the *why* behind problems — and add a per-screen interview script for the screens you're most worried about.
* **Expert audits give design-review-style feedback.** A systematic pass against Nielsen's heuristics (or your own design-system rules as [custom heuristics](/ai-tests/test-modes#expert-audit-heuristic-evaluation)) — great before design reviews.
* **Use SUS to compare across concepts.** Enable the [SUS questionnaire](/ai-tests/test-modes#sus-questionnaires) on a fixed test setup and re-run it per concept or design iteration — the 0–100 score gives you a consistent yardstick. Read it relatively (A vs. B), not as an absolute benchmark.

## Iterate inside the project

* **Add more test runs anytime.** A project isn't one-and-done — after reading the first results, add runs with new personas or more executions to the same project and regenerate the summary.
* **Test iterations in the same project.** When the design changes, run a new round of tests in the same project with the same personas — comparing rounds shows you whether the fixes actually landed.

## Know what AI tests are best at

AI tests excel at uncovering **usability, product, and design improvements** — confusing flows, unclear copy, missing affordances, heuristic violations — in minutes instead of weeks. Treat findings as hypotheses ranked by likelihood: for higher-risk decisions, or research into real-world behavior, attitudes, and willingness to pay, validate with real participants in a [study](/studies/overview).