Audit your app
How to Get Honest Website Feedback Before You Launch
Get honest website feedback before launch: run an audit first, then watch about 5 real people try one task, and learn to read what they tell you.
The most useful website feedback comes from about 5 people who look like your real visitors, each asked to do one real task while you stay quiet. Run an audit first and use its report as your test plan, so each tester checks something specific. Then trust the problems that repeat.
After weeks of building, you read your own headline and see everything you meant. Asking outside eyes to look can feel risky, because they'll only see what's actually on the page. What follows is a small plan, in the order we'd run it.
What should an audit clear out before you ask anyone?
Start here, so no tester spends a session on something you could find alone. The report comes from an AI scoring a fixed rubric, so no person reads your site. You get a score out of 100, a Verdict label, and for each of the six product dimensions a What works note, Slop tells, and Top fixes ranked by impact.
An audit checks for signals, and people tell you what they make of them. So use the report as a test plan, and let what people do settle the question.
How do you turn an audit report into a test agenda?
Open the report before you recruit anyone. 3 of its fields turn straight into things for testers to confirm or kill. (Our guide on reading an audit report goes deeper.)
- Not confirmed lists what the audit tried but couldn't verify. Hand these to a person first. Give a tester a task that passes through each one, and write down what happens.
- Top fixes, ranked by impact are bets about what hurts most. Build a path that touches the first fix in each dimension. If 2 or more testers hit the problem, it stays near the top. If nobody does, move it down your list but keep it there, because 5 people can find problems but can't prove a page is fine.
- Slop tells are predictions. Each one says where the report expects a visitor to notice work that looks generated or careless. Copy each onto a card as a plain sentence, and after the round mark it confirmed, contradicted, or not seen.
Keep a fourth pile for anything testers hit that the report never mentioned. That pile is the reason to run a people round at all.
Who should you ask for website feedback?
Ask people who could really use what you're making. For one kind of visitor, a good mix is 3 people who match your real visitor, 1 who knows nothing about your topic, and 1 who'll use a phone instead of a laptop.
If your site serves clearly different groups, test people from each. Nielsen Norman Group (NN/g), a research firm that studies how people use websites, suggests 3 to 4 per group when you have 2 groups, and 3 per group when you have more (NN/g's article on testing with 5 users).
Start with anyone who has already talked to you about the problem, or a community you belong to.
What should each tester do?
Give every tester the same one main task, so you can count repeats: "Sign up for a shift," "find the price and start checkout," or "book a call." Then give each of the 5 testers one extra prompt, a different one each. Every prompt tests a different dimension of the rubric, so the round covers 5 dimensions.
| Tester | Extra prompt | Dimension it tests |
|---|---|---|
| 1 | "In your own words, what is this site for, and who is it for?" | Clarity and Promise |
| 2 | "Before each important click, tell me what you expect to happen." | Flow and Usability |
| 3 | Do the main task on their own phone, held the way they normally hold it. | Mobile and Access |
| 4 | "What would you have to see before you handed over your email or a card?" | Trust and Legitimacy |
| 5 | "What is this page asking you to do, and what, if anything, is holding you back?" | Conversion and The Ask |
Craft and Design Integrity gets no prompt on purpose. Testers volunteer taste comments without being asked, and those are the easiest to park. Its Slop tells still go on your prediction card.
One catch: an extra prompt asked of one person is a single reading. Treat it as strong only when it matches a card or a second tester says the same thing unprompted.
Which questions get you flattery instead of facts?
Show a friend your site, ask whether they like it, and you'll usually hear yes. As I put it in Build the Damn Thing, "because most people don't relish crushing someone's idea, we often never receive the truth ... until we've spent money and time building a bad product."
"Do you like it?" and "Would you use this?" invite that polite yes, and they ask people to predict their behavior, which is a different thing from watching it. Your own words can steer people too. NN/g's page on thinking aloud says prompts and clarifying questions are usually necessary, but from an untrained facilitator they can very easily change user behavior. The same page says that unless you blatantly bias users by putting words into their mouths, you'll still get reasonably good findings, even from a poorly run study.
So keep your questions neutral, and say less:
- Instead of "Isn't the pricing easy to find?" try "Where would you look for the price?"
- Instead of "Is it clear?" try "What is this page asking you to do?"
How do you run a 5-person round?
One person at a time. A video call works fine.
- Fix the obvious breaks first. Run the audit and write your prediction cards. Then click every button yourself (our guide on testing a vibe-coded app without being a developer shows how).
- Send the live link, and ask each tester to use their own device.
- Ask them to think aloud. That means saying what they see, what they expect, and what confuses them as they go. NN/g notes this doesn't come naturally, so expect to nudge.
- Stay quiet. Don't explain or defend the site. If they ask "what does this do?", answer "what would you guess?" If they go silent, ask "what are you thinking?"
- Write down what they do, and not only what they say: every pause, wrong click, and question.
- Finish with 3 questions: What was this for? What almost stopped you? What would you change first?
Keep each session to 15 to 20 minutes, and tell testers that up front. Our guide to website usability testing turns one of these sessions into 3 written tasks and a simple way to rank the problems.
Why 5 people and not 50?
NN/g's article says the best results come from testing no more than 5 users and running as many small tests as you can afford. It puts the typical share of problems a single user finds at 31%, averaged across the projects its authors studied, and says that after the fifth user you mostly see the same things again. It prefers 3 studies with 5 users each to 1 study with 15, so you can fix what you find and check the fixes worked.
Five people are good for finding problems. Measuring them is another job. If 3 of 5 get lost on the same screen, fix it. That doesn't tell you how many future visitors will get lost, and the same page points to 20 users for studies that need numbers.
How do you read mixed feedback?
Sort it with these rules:
- Behavior beats opinion. What people did matters more than what they said.
- Repeats come first. The same snag from several people goes to the top.
- Broken is broken. A button that fails for one person is still a fix.
- One-off taste can wait. See if anyone else says it.
- Problems are useful, fixes are optional. People are good at spotting where it hurts and weak at prescribing the cure.
Say 5 people try a volunteer sign-up site for a community garden, one we made up for this guide. Their task: sign up for a Saturday morning shift. Its report had flagged some of what they hit and missed the rest.
| What happened | How many | Did the report see it? | What we would do |
|---|---|---|---|
| Couldn't tell if newcomers were welcome | 3 of 5 | Yes, as a Clarity and Promise Slop tell | Fix first: say "No experience needed" on the first screen |
| Signed up, then wasn't sure it had worked | 2 of 5 | Yes, under Not confirmed | Likely fix: show a confirmation and say what happens next |
| On a phone, the date picker wouldn't open | 1 of 5 | No | Fix: it's a break, which outranks a preference |
| Said the site's green felt "too bright" | 1 of 5 | A low-ranked fix | Note it and wait |
| Asked for text reminders | 1 of 5 | No | Park it for later |
The first row outranks the rest: 3 people couldn't answer a basic question about the page, and the report had called it, so that prediction is confirmed. The date picker gets fixed even though only 1 person hit it, and it shows why the people round is worth running. The report never saw it.
Is it worth asking an online crowd to roast your website?
Some founders post a link and ask a community to "roast my website." The replies come fast and blunt, but the people replying may not be your visitors and aren't doing a task, so you get taste more than behavior. Run the round instead, or after it.
What do you do with the feedback once you have it?
Make 3 piles: breaks, repeated confusion, and everything else. Fix the first 2. Park the third.
Then change only a few things and run another round with 3 to 5 new faces. NN/g's article warns that a redesign can fix one problem and introduce another, so the second round is worth it. If you changed a lot, run the audit again first.
When people can say what your site is for and finish the one task you gave them, you're in a good place to launch. Our guide on when a vibe-coded MVP is good enough can help with that call. A kind "I would use this" still isn't a purchase.
Sources, both read 5 October 2026: NN/g's article on testing with 5 users (published 18 March 2000, updated 2 February 2024) and NN/g's page on thinking aloud (published 15 January 2012, updated 31 January 2024).
