We were stuck on a genuinely expensive decision: build a native mobile app, or pour the same effort into making our mobile web experience excellent. The kind of decision where being wrong costs two quarters. I needed data to break the tie. The catch: our UX researcher, the person who is good at this, was on parental leave, and the decision could not wait three months. So it fell to me, a PM whose survey-writing instinct is, scientifically speaking, bad.
I know my weakness here precisely because I have been burned by it. Left to my own devices, I write questions that are accidentally leading ('How frustrated are you with our mobile site?' presumes frustration) or so broad they produce mush ('What do you think about mobile?'). A bad survey is worse than no survey, because it gives you confident, official-looking numbers that point the wrong way. I was genuinely worried about steering a two-quarter decision with garbage data.
What the prompt actually caught
I set the research goal to 'understand how users access our product on mobile and what friction they hit,' the decision to 'whether to build a native iOS and Android app next quarter,' and the respondents to 'active users who logged in on a mobile device in the last 30 days.' I ran the Design a Bias-Minimized User Survey from a Research Goal prompt on Gemini 3 Pro Preview, partly because I wanted the clean table layout it produces and partly because it is strong at holding a lot of context and keeping the whole survey coherent.
It returned eight questions in a tidy table, and the bias note column was worth the entire exercise. My instinct would have produced question three as 'How frustrating is our mobile site to use?' The prompt's own bias note on a similar question flagged it: 'presupposes frustration; replace with a neutral satisfaction scale with both anchors labeled.' It had caught, and fixed, the exact mistake I make every time, before I even made it. It also forced in a behavioral question, 'How often do you complete a full task on mobile versus switching to desktop,' which turned out to be the most useful question of the eight.
I want to dwell on that behavioral question for a second, because it taught me something about my own bad instincts. Left to myself, I write attitude questions: would you prefer, do you like, how satisfied are you. Those feel like research but they mostly measure how people imagine they would behave, which is famously unrelated to how they actually behave. The prompt is built to insist on at least one behavioral question, 'how often do you,' and that single question, about how often people switched to desktop to finish a task, told us more than the other seven combined. People are unreliable narrators of their own preferences but pretty reliable reporters of their own habits. I would never have known to lean on that distinction.
- 67 percent of mobile users said they used mobile web only because they 'did not know there was an app option,' which completely reframed the decision: our discovery problem was masquerading as a platform problem.
- The required open-text question surfaced three specific friction points we had not anticipated, the biggest being that login expired far too aggressively on mobile.
- The recommended-order section put the sensitive 'would you pay more for a better mobile experience' question last, after trust was established, which I would have buried in the middle and tanked the response rate.
- We decided not to build the native app, fixed the three friction points instead, and got a measurable mobile-completion lift for a fraction of the cost.
The honest part
When our researcher came back from leave, I showed her the survey, slightly nervous, like a kid showing a drawing. She read it, changed exactly two words, and said the structure was solid. For someone who is genuinely not a researcher, that was the highest compliment available. I want to be clear that the prompt did not replace her; it gave me a competent baseline that her two-word edit then made excellent. It is a great floor, not a ceiling.
I will also note: I left this one unverified on the marketplace on purpose for a while, because survey design is a craft and I wanted real usage before I vouched for it hard. It has earned its keep. If you are a non-researcher who occasionally has to run a survey that matters, grab the Design a Bias-Minimized User Survey from a Research Goal prompt on Prompt Dock and run it before you write a single question yourself. When the responses come back, I pipe the open-text answers straight into my Cluster Raw User Feedback into Ranked Themes with Verbatim Quotes prompt and the analysis is done by lunch.