Maze review: AI-assisted user testing for product teams in 2026

Maze is a continuous product research platform: you build a test from a prototype, a live URL, or a survey, send a link to participants, and get structured results back without scheduling a single moderated call. The pitch to product teams has stayed consistent since the company’s early days — turn research into something a PM or designer can run in an afternoon, not a multi-week study handed off to a dedicated researcher. What changed in 2025 and 2026 is the AI layer Maze wrapped around that workflow. We spent time with the test builder and the reporting view to figure out where that layer earns its place and where it quietly gets in the way.

What Maze actually does

Strip away the marketing and Maze is a few distinct tools sharing one dashboard. You can run unmoderated usability tests on a Figma prototype or a deployed site, with task success, misclick paths, and time-on-task captured automatically. There are surveys with branching logic, card sorting and tree testing for information architecture, and short-format five-second tests for first-impression checks. A separate recruitment product lets you pull participants from Maze’s panel when you don’t have your own list, billed per response.

The through-line is that everything is unmoderated and asynchronous. You are trading the depth of a live interview — where you can ask “wait, why did you do that?” in the moment — for volume and speed. That trade is the whole reason teams reach for Maze instead of a calendar full of Zoom calls. A test that would take two weeks to schedule and run as moderated sessions can collect 30 to 50 responses overnight.

The reporting is where Maze has always been stronger than DIY survey tools. Results roll up into a shareable report with heatmaps, completion funnels, and clip reels, so the output is something you can drop into a stakeholder review rather than a CSV nobody opens.

Where the AI features help — and where they don’t

Maze’s AI features cluster into three jobs: drafting questions, generating follow-up prompts during a session, and summarizing results after.

Question drafting is the most useful of the three. You describe what you’re trying to learn and Maze proposes a test structure — tasks, follow-up questions, and rating scales. For a PM who runs research occasionally and isn’t fluent in survey design, this removes the blank-page problem and nudges you away from leading questions. You still need to edit the output, because the generated wording tends toward the generic, but starting from a draft beats starting from nothing.

AI follow-up questions react to an open-text answer in real time, asking a participant to expand when they give a thin response. This is the feature that comes closest to closing the gap with moderated research. In practice it works when the participant wrote something concrete and stalls when they didn’t — an AI prompt asking someone to “tell us more” about a one-word answer rarely produces gold. It nets you a bit more signal per session without you sitting in the room.

AI summarization is where you should keep your guard up. Maze will cluster open-text responses into themes and write a synthesis. For a quick read on 40 responses, that’s a genuine time-saver. But a summary is a compression, and compression loses the outlier — the one participant who described a workflow you’d never considered. The themes an AI surfaces are the frequent ones, which are often the ones you already suspected. The insight that changes your roadmap is usually rare, and rare is exactly what averaging filters out.

The pattern across all three: AI helps most at the edges of the workflow — setup and first-pass triage — and helps least at the part that actually requires judgment, which is deciding what a finding means for your product.

Pricing and whether it fits your team

Maze offers a free tier with a capped number of studies and responses, paid plans billed annually, and an enterprise tier gated behind a sales conversation. The recruitment panel is priced separately per response, and it adds up fast if you don’t bring your own participants — sourcing is frequently the largest line item in an unmoderated research budget, not the software seat.

The teams that get the most out of Maze share a profile: they ship frequently, they already have a list of users or customers they can recruit from, and they want research to be a recurring habit rather than a quarterly event. If you run research twice a year, the free tier or a manual survey tool will cover you. If a designer or PM is testing flows most weeks, the structured reporting and reusable test templates start paying for themselves in saved synthesis time.

Where Maze fits awkwardly: deep generative research and discovery interviews. Those want a moderated, conversational format, and bolting AI follow-ups onto an unmoderated test does not fully substitute for a researcher who can change direction mid-conversation.

Whatever you collect, the findings need a home your team will actually revisit. A research repository — tagged by feature, linked to the decisions it informed — is what keeps a study from being run once and forgotten.

The honest summary

Maze in 2026 is a competent unmoderated testing platform with an AI layer that is genuinely helpful for setup and first-pass synthesis, and genuinely risky if you let it make decisions for you. Buy it for the speed of going from question to structured result, and for reporting your stakeholders will actually open. Keep a human in the loop for the part that matters — reading the raw answers and deciding what they mean.