Yes, you can do meaningful UX research without ever interviewing a user, by reading the behavioral and indirect signals your product already generates and by running AI synthetic users through the flows you can't recruit real people for. This guide covers the signals you already have, how unmoderated and automated testing fills the gaps, where AI synthetic users fit, and how to combine all of it. It is written for pre-launch teams, small teams with no budget, and B2B products where the real users are gatekept. The short version: you can learn a lot without talking to users, but treat everything you find as a hypothesis, not a verdict.
What signals do I already have?
Even with zero interviews, your product is generating evidence. Analytics and session-replay tools show where people hesitate, rage-click, and drop off. Your support and customer-success teams already hold a backlog of the exact friction users hit most, so ask them first. App-store reviews, and your competitors' reviews, tell you what people praise and what makes them quit.
Public communities are the most underused source. Eleken recommends mining forums and communities like Reddit and Quora for unfiltered pain points, noting that Reddit posts "often reveal unfiltered user frustrations, giving a more authentic look at what people really think." None of this needs a recruiting budget. For a wider survey of what to reach for, see our roundup of the best UX testing tools.
What about unmoderated and automated testing?
When you do want task-based feedback but can't sit in a room, unmoderated testing is the next step. Respondent defines unmoderated UX research as gathering insights without a facilitator, where participants complete tasks and give feedback on their own, in their own environment.
This matters because you can recruit a general panel for an unmoderated test even when your specific users are off-limits. A panelist may not be your exact buyer, but they will still surface broken inputs, confusing labels, and dead-end flows. Automated checks catch the rest: accessibility failures, slow pages, and steps that simply don't work.
Where do AI synthetic users fit?
An AI synthetic user is a persona given a profile and a goal that navigates your product the way a matching person would, then reports what felt confusing. This is the option that lets you test a flow you cannot recruit for at all, like a gated B2B onboarding or a product with no traffic yet.
Used well, they are a fast way to generate hypotheses before you spend on real research. Nielsen Norman Group says a synthetic user can be useful if researchers treat the output as a hypothesis to guide future research, not as a finished answer. If you are new to the idea, our explainers on synthetic personas and AI usability testing go deeper.
What are the limits?
Be honest about what synthetic users can't do. The same Nielsen Norman Group analysis notes that synthetic research cannot produce behavioral data, because AI can't actually use a product like a human does. The output is reasoning about your product, not a recording of someone struggling through it. AI chatbots also tend toward sycophancy, so they can agree with your framing instead of pushing back.
There is also a range problem. A later Nielsen Norman Group study found the standard deviation in synthetic data was consistently lower than in human data, which means AI underrepresents the full spread of real opinions. The same study allowed that "the gaps were narrow enough that their directional accuracy might still be useful in exploratory research or early-stage testing." Directional is the right word. Use it for early signal, not for a launch decision.
How do I combine all of this?
Triangulate. Start with the signals you already own, analytics, support tickets, reviews, and forum threads, to form a short list of suspected friction points. Then run AI synthetic users through those exact flows to see whether an outside actor hits the same walls and to generate hypotheses you hadn't considered.
Where the stakes are high, validate the top hypotheses with an unmoderated test on a recruited panel. No single method is trustworthy alone, but stacked together they cover behavior, opinion, and edge cases without a single interview on your calendar.
How do I try it?
Swarm runs AI synthetic users through your product like real users, surfacing friction, drop-offs, and confusing steps before launch, even when you have no traffic and no one to recruit. It works in the browser, your terminal, or as an MCP server in Cursor and Claude Code so the model that wrote your code can test it too.