Hypothesis statement: we believe that, we will know

A hypothesis statement writes a product belief and the condition that would refute it. The Lean UX format, ranking by risk, and avoiding late metrics.

info Quick Definition
A hypothesis statement is a sentence that writes a belief about the product in two parts: what it expects to achieve for users and the business, and the observable condition that would show that belief is false. It turns an idea or a requirement into something that can be tested.

What is a hypothesis statement?

It is the bridge between an idea and an experiment. Instead of writing “add a help button on the payment screen”, the team writes what it believes will happen if it does, for whom, and how it will know if it is wrong.

Lean UX is the approach of Jeff Gothelf and Josh Seiden. According to O’Reilly, it rests on three foundations: user experience design, agile development, and Lean Startup by Eric Ries, with its build-measure-learn loop. In that approach each design is treated as a hypothesis to validate with customer data, and the hypothesis statement is the format for writing it down.

What is the format?

Gothelf proposes a two-part structure. The first part states the belief:

We believe that [business outcome] will be achieved if [users] attain [benefit] with [feature].

The second is the falsification line, the condition that would tell you the belief is wrong:

We’ll know we’re wrong if [observable behavior] doesn’t reach [level] by [date].

The second part is what makes the hypothesis useful. If you cannot write what result would disprove it, it is still an opinion.

Illustrative example: “We believe app retention will rise if new users complete their profile with the suggestions feature. We’ll know we’re wrong if fewer than 40% of new users complete their profile within the first two weeks.” The hypothesis has an outcome, a group, a benefit, a feature, a threshold, and a deadline.

To write your own, fill in the blanks in this order:

  1. Business outcome. Which business number should move (retention, conversion, support tickets).
  2. Users. The specific group, not “everyone”.
  3. Benefit. What that person can now do that they could not before.
  4. Feature. The solution you will test, just one.
  5. Behavior, level, and date. What you will observe, at what threshold, and by when.

Why does it matter to a designer?

Because most design debates are settled by preference: “I like this version better”. A hypothesis changes the question to “what behavior would show this version works”, and that lets you design the test alongside the screen.

It also helps prioritize. A designer who writes the hypothesis before drawing knows which screen is needed to test it and which can wait. And it gives product, engineering, and business a shared language: everyone sees the same bet and the same failure condition.

How do you choose what to test first?

Rank assumptions by risk: which ones would most change the plan if proven false. UIE’s seminar on hypotheses in Lean UX puts it this way: do not try to test every idea, test the riskier ones. And start with the simplest experiment that gives you learning.

Some practical rules:

  • One hypothesis per feature, so a result is not attributed to something else.
  • The cheapest experiment first. Gothelf notes that, for AI features, the cheapest version is usually a prompt and a model with no interface, evaluated on real inputs before building the screen.
  • One experiment at a time when possible, so results do not contaminate each other.

How do you avoid traps when evaluating?

The most common error is choosing the metric after seeing the data. Wikipedia defines vanity metrics as measurements that give the rosiest picture possible but do not reflect the real drivers of the business. Be wary of any metric that appears only when the hypothesis works.

That is why the falsification line is written before building. Gothelf’s article on turning a hypothesis into an AI eval insists on agreeing what success looks like before seeing results, so the team does not reverse-engineer the story afterward: a written prediction is not quietly revised. If the threshold was 40% and the result is 38%, the hypothesis is in question, even if the team would prefer the result.

A nuance for design: rely on real data from your users, not assumptions. Gothelf makes the point for AI evaluations: made-up inputs mostly tell you how the model handles your own assumptions. If you have no reference point for the threshold, a pilot test can help set it.

When should you NOT use hypotheses?

If the decision is already fixed for legal or safety reasons, a hypothesis changes nothing. Nor is it useful for low-risk changes that can be reverted in a day, where writing the hypothesis costs more than the experiment.

There are also practical limits: prototyping is not always feasible, a test can cost more than it returns, and a team without practice may misread the results. A badly written hypothesis gives a false sense of rigor.

Example

Illustrative example: a course platform team suspects that students drop out of lesson three because the video is too long. It writes: “We believe students will finish more lessons if the video is split into five-minute segments. We’ll know we’re wrong if the share of finished lessons does not rise by at least five points within three weeks.” They test the segmented version with half of the students, with the threshold set beforehand. In the end, the increase is three points. The hypothesis is considered unconfirmed, and the team looks for another cause before changing the design of every lesson.

Common mistakes

  • Writing a hypothesis without a falsification line, so any result looks like a success.
  • Mixing several features into one hypothesis.
  • Changing the threshold after seeing the data.
  • Choosing the metric that did improve and forgetting the one that got worse.
  • Turning the hypothesis into a requirement with a delivery date, when it should be a question.
  • Writing many hypotheses at once without ranking them by risk.

To validate the belief with people before building, concept testing is the natural method. If the experiment needs to simulate a response, Wizard of Oz lets you test the interaction without building it. And to follow up on the result, the post-launch review helps compare the goal with what happened.

Sources

Frequently asked questions

What is a hypothesis statement?
It is a sentence that writes a belief about the product in two parts: the outcome expected for the business and users, and then the observable condition that would show the belief is false. In Lean UX it turns requirements into experiments.
What is the most common format?
Jeff Gothelf proposes ‘We believe that [business outcome] will be achieved if [users] attain [benefit] with [feature]’, followed by ‘We’ll know we’re wrong if [observable behavior] doesn’t reach [level] by [date]’.
How do I choose which hypothesis to test first?
Rank assumptions by risk: the ones that would most change the plan if they proved false. UIE advises not trying to test every idea and focusing on the riskier ones, with the simplest experiment that yields learning.
What mistake should I avoid when evaluating a hypothesis?
Choosing metrics after the experiment in order to declare the hypothesis met. Set the success criterion and the failure condition in advance, and do not change them after seeing the data.