Synthetic users for website testing: how AI personas find what your analytics miss

What are synthetic users for website testing?

A synthetic user is an AI agent briefed as one of your customer personas. It opens your site in a real browser, clicks, scrolls and reads like a customer would, and thinks out loud along the way. For each session you get the path it took, its quotes, a screenshot of every finding and a ranked list of what to fix.

Three terms get mixed up here:

• Synthetic users, AI personas and digital twins usually mean the same thing in website testing: AI personas built from your customer data that act on your real pages.
• Synthetic monitoring is different. Scripted bots check that a page loads and a checkout works. They tell you the site is up, not whether a person would understand it, trust it and buy.
• AI survey tools simulate interview answers. Synthetic users act on your actual pages, so every piece of feedback is tied to what they saw on screen.

What can synthetic users find that analytics can't?

Analytics counts behavior. It tells you that half your visitors leave the pricing page, not that they can't work out the monthly cost after the promotion ends. Synthetic users give you the reason, in the customer's words, with the screen they were looking at.

In customer projects, that has meant:

• Things that are broken, and flows that make no sense. In every customer project so far, personas found something broken, such as missing content on a page, and flows that don't add up, such as a form that asks for information the customer doesn't have.
• Gaps nobody on the team could see. An Italian persona found that MIRON's live chat was still in English. A Kenyan persona hit a language issue on a reading platform that the team hadn't caught, because nobody on it browses the site as a reader in Kenya.
• Friction behind a login. At a top European telco, 35 personas ran 140 sessions through 4 customer portal journeys and surfaced more than 100 issues that page analytics alone couldn't explain.
• Moments that break the flow. On the same reading platform, an audio player that opened in a new window interrupted listening. The team moved it onto the page.

When should you use synthetic users?

Use them wherever you'd want customer feedback but can't get it fast enough, or at all.

As MIRON put it: "You don't need a huge research project to start getting value from it."

How does a synthetic user test work?

A run takes about 5 minutes, not weeks, and a test follows the same six steps every time:

1. Build personas from what you already know. There are three ways in: upload raw data, such as interviews, survey answers or NPS verbatims; use persona descriptions you already have, uploaded or typed in; or, if you have nothing yet, let Maskerade infer personas from your current website.
2. Write one clear task. Say what to do, such as "find a plan that fits your household and sign up", not how to do it.
3. Do a sample run. Send a few personas through first, so a login wall or a bug doesn't spoil the full run.
4. Use at least 20 personas per run. With at least 6 per segment, built as variations of your core personas. Each persona finds its own path, so volume is what covers the journeys.
5. Review with people who know the business. Personas know what customers feel, not your regulations or what you tried last year.
6. Fix, then re-test with the same personas. Because the personas didn't change, any difference comes from your site.

More best practices for each step can be found in Synthetic users: 13 best practices from real customer projects.

How accurate are synthetic users?

In our own study, synthetic users found 8 of the 10 core findings (80%) of a moderated focus group on the same website, and contradicted none of them. They also found 40% more issues, which two other focus groups confirmed. What didn't hold up were wishes for a future redesign, not problems with the current site. The method is in How accurate are synthetic users? Our focus group study.

Outside research points the same way. Stanford researchers found that AI agents built from in-depth interviews matched people's survey answers 83% as well as the people matched their own answers two weeks later, against 74% for agents built from demographics alone (Park et al., 2024, revised 2026).

Critics are right that thin personas drift toward average answers, as a 2026 NN/g study shows. That risk shrinks when personas are built from your own research and act on real pages. How we build synthetic users, and the research behind it explains the structure we use.

How do synthetic users compare with other methods?

They fill the gap between what analytics measures and what a user study explains.

The best setup combines them: synthetic users find and filter problems fast, and for big decisions you add focus groups, one-to-one UX research, analytics data and A/B tests.

Where do synthetic users fall short?

• Future wishes. They are strongest on the site as it is today. Treat what they say about features that don't exist yet as hypotheses.
• Business context. They don't know your regulations, technical limits or past experiments. Your team does.
• Thin input. Personas are only as good as the research behind them. If your data leaves a group out, so will they.
• Bot protection. Heavy bot protection can stop a browsing agent, so check access before a big run.

Is testing with synthetic users GDPR-safe?

Personas are built from research about groups of customers, not from tracking individual visitors, and you never need to enter personal data about real people.