Just-Noticeable Difference: The Smallest Product Change Users Can Actually Perceive (2026)

A change can be real, statistically significant, and completely invisible to the people using your product. The just-noticeable difference (JND) is the smallest change in a stimulus that a person can reliably detect, and it is a property of human perception, not of your analytics pipeline. It scales with the size of the thing being changed: shaving 100 ms off a 400 ms interaction is dramatic, and shaving the same 100 ms off a 4-second interaction is nothing at all. If you ship improvements below the JND and then run a survey asking whether people noticed, you will get a flat result and conclude that users do not care — when the truth is that they could not tell.

This guide covers what the JND is, the arithmetic that predicts it, the peer-reviewed numbers you can use as a starting estimate, and how to measure your own thresholds with a small number of participants using AI-moderated interviews.

The one rule: perception is proportional, not absolute

Ernst Weber established in the 1830s that the smallest detectable change in a stimulus is a roughly constant proportion of the starting stimulus, not a constant amount. Written out, the rule is:

delta_I / I = k

where I is the starting intensity, delta_I is the smallest detectable change (the JND, also called the difference limen or DL), and k is the Weber fraction — a constant for a given kind of judgement.

Everything commercially interesting follows from that denominator:

• A price rising from €10 to €11 is a 10% change. The same €1 on a €400 annual plan is 0.25%, and it is invisible.
• A page that renders in 300 ms instead of 400 ms is a 25% improvement. The same 100 ms taken off a 3-second render is 3.3%, and nobody will report it.
• Adding one extra field to a two-field form is a 50% increase in effort. Adding one field to a twenty-field form is barely detectable — which is exactly how twenty-field forms come to exist.

The practical consequence is that you cannot budget product improvements in absolute units. A latency budget expressed in milliseconds is not a perception budget. A perception budget is expressed as a percentage of the current value.

The numbers you can start from

You do not have to guess at the Weber fraction for time. Rammsayer, Borter and Troche (Frontiers in Psychology, 2015, volume 6, article 1626) ran auditory and visual duration-discrimination tasks with 46 participants at two standard durations, 50 ms and 1000 ms. Their reported difference limens:

Read the third row again, because it is the one that matters for software. At a one-second reference, the average participant needed a change of roughly 206 ms before they could reliably tell two visually marked durations apart. A 100 ms improvement on a one-second interaction is less than half of that threshold. It is a genuine engineering win, it will show up cleanly in your telemetry, and the typical user cannot perceive it.

The authors also report that the visual-auditory gap closes once you control for the sensory-automatic timing mechanism: performance on longer intervals was, in their words, "virtually identical as indicated by adjusted mean Weber fractions of 0.175 and 0.173 for auditory and visual intervals, respectively." So a Weber fraction somewhere near 0.17 to 0.21 is a defensible planning assumption for one-second-scale durations, until you measure your own.

Treat these as a starting prior, not as your answer. Weber fractions vary by modality, by task, by how much attention the person is paying, and by how the comparison is presented. Two builds compared side by side are much easier to tell apart than two builds separated by a week.

Difference thresholds are not the same as absolute thresholds

There are two distinct perceptual questions and product teams routinely merge them.

An absolute threshold asks: at what point does this become perceptible at all? Jakob Nielsen's three response-time limits, published in Usability Engineering in 1993 and drawn from R. B. Miller's 1968 work on conversational transactions, are absolute thresholds. Nielsen gives 0.1 second as "about the limit for having the user feel that the system is reacting instantaneously", 1.0 second as "about the limit for the user's flow of thought to stay uninterrupted, even though the user will notice the delay", and 10 seconds as "about the limit for keeping the user's attention focused on the dialogue".

A difference threshold asks: how much do I have to change this before anyone notices it changed? That is the JND, and it is proportional.

You need both, and they give different advice. Absolute thresholds tell you which side of a qualitative boundary you are on — instantaneous, uninterrupted, or lost. Difference thresholds tell you whether an incremental improvement is worth announcing. A change from 1.4 s to 1.1 s may cross no absolute boundary at all and still be perceptible, because it is a 21% reduction. A change from 9.0 s to 8.4 s is neither: below the difference threshold, and on the wrong side of every absolute limit anyway.

This is not statistical power

The most common confusion in this area is between the smallest change your sample can detect and the smallest change your users can detect. They are unrelated quantities and they move in opposite directions.

• The minimum detectable effect (MDE) is a property of your study design: sample size, variance, and the alpha and power you chose. Recruit more people and it shrinks toward zero.
• The just-noticeable difference is a property of human sensory and cognitive machinery. Recruiting more people does not move it at all. It changes only if the reference level changes.

This produces two failure modes worth naming:

1. Statistically certified invisibility. With 40,000 sessions you can certify a 12 ms improvement at p < 0.001. No user will ever perceive it. The finding is true and worthless as a communications claim.
2. Perceptible but uncertifiable. With 60 participants you may be unable to reach significance on a change that half your users can plainly feel. The effect is real; the study is too small to prove it.

If you are sizing a study, read Statistical Power and Minimum Detectable Effect alongside this article. That one tells you how many participants you need. This one tells you whether the change is worth asking about in the first place. Design the change against the JND, then size the study against the MDE.

How to estimate a JND without a perception lab

Classical psychophysics estimates a threshold by presenting many carefully controlled stimulus levels to a handful of trained observers. You almost certainly cannot do that. You can do something considerably cheaper that still beats guessing.

Step 1: pick one continuum and hold everything else fixed. Latency, price, copy length, notification frequency, contrast, step count. Changing two at once gives you a number that means nothing.

Step 2: build a ladder around the current value. Five levels, spaced geometrically rather than arithmetically, because Weber's law is proportional: for a 1000 ms baseline, try 1000 / 1150 / 1320 / 1520 / 1750. Geometric spacing puts equal perceptual distance between adjacent rungs; arithmetic spacing wastes your top rungs.

Step 3: use forced choice, never free report. Show a participant two versions and require them to pick which one is faster, cheaper, or heavier. Do not ask "did you notice a difference?" — that question measures willingness to say yes at least as much as it measures perception, which is the subject of the companion article on signal detection.

Step 4: include a catch pair where nothing differs. If a participant confidently picks a winner between two identical builds, you have learned something important about the rest of their answers.

Step 5: find the level where accuracy crosses about 75%. By convention the JND is the difference that is detected reliably rather than always — 75% correct in a two-alternative task is the standard anchor, because 50% is pure chance. The mechanics of converging on that level efficiently are covered in threshold testing.

Step 6: ask why immediately after. The number tells you where the threshold is. The follow-up conversation tells you whether crossing it matters — a change can be perceptible and unwelcome, or perceptible and irrelevant.

Running this in Koji

Traditional survey tools make threshold work awkward because they are built to collect one answer per question and stop. A perception study needs a comparison, a confidence judgement, and a reason, and the reason has to be probed rather than typed into a text box that most people leave blank.

Koji's structured questions cover the full ladder in a single study. All six types are first-class:

• single_choice for the forced-choice comparison — which of these two builds felt faster.
• yes_no for the catch trials and for confidence gates.
• scale for magnitude judgements — how much faster, on a 1 to 7 scale — which turns a binary detection into a graded response you can plot.
• ranking for ordering three or more variants at once, which extracts more information per participant than a series of pairs.
• multiple_choice for attributing the perceived change to a cause, when you want to know whether people credit the right thing.
• open_ended for the reason, where Koji's AI interviewer asks its own follow-up questions instead of accepting "it just felt faster" and moving on.

That last point is the part no survey tool replicates. When a participant says a build felt quicker, the AI interviewer probes what they were doing, where they felt the wait, and whether anything else about the experience changed — automatically, in every session, without a moderator on the call. Voice and text both work, so you can run the comparison the way the product is actually used. The analysis arrives as a report rather than a spreadsheet, with the distribution of forced-choice answers per level sitting next to the quotes that explain them.

The practical result is that a threshold study that would have taken a scheduled lab and two weeks of moderation runs overnight, with the follow-up depth of a moderated session and the sample size of a survey.

What to do with the answer

Once you have a defensible JND for a continuum you care about, three decisions get easier.

Ship-or-hold. If an improvement lands below the threshold, it is still worth shipping for compounding reasons — three sub-threshold wins stack into a supra-threshold one — but it is not worth a changelog headline, a campaign, or a satisfaction survey. Measuring it with a perception question wastes participants.

Batching. Accumulate sub-threshold improvements and release them together so the perceived jump clears the threshold in one step. Users experience one noticeable improvement rather than four invisible ones, and the same engineering work produces a measurable attitude change.

Degradation budgets. The rule runs in reverse. If a new feature adds latency, you can compute how much you can add before anyone notices — and the answer scales with the current baseline, which means the fast parts of your product are the expensive places to spend it and the already-slow parts are where nobody will care.

The uncomfortable corollary of that last point is worth stating plainly: the better your product gets, the smaller the absolute improvement anyone can perceive. A team that has already cut a flow from 4 s to 800 ms faces a JND of roughly 170 ms on the next round, where at the start it was 800 ms. This is the perceptual equivalent of diminishing returns, and it is a reason to move on to a different continuum rather than to keep optimising a fast thing. It is also why a threshold has a shelf life: the reference level your users judge against moves as they adapt to what you shipped, which is the subject of the zero point moving.

Frequently asked questions

What is a just-noticeable difference in plain language?

It is the smallest change in something — a price, a delay, a size, a volume — that a person can reliably tell apart from the original. Reliably usually means detected about 75% of the time in a two-option comparison, since 50% is what pure guessing produces. Below the JND, the change is real but imperceptible.

How is the JND different from the minimum detectable effect in a survey?

The minimum detectable effect describes your study: it is the smallest difference your sample size and variance let you certify statistically, and it shrinks as you recruit more people. The just-noticeable difference describes your users: it is the smallest difference a human can perceive, and recruiting more people does not change it. A change can be statistically significant and perceptually invisible, or obviously perceptible and statistically unproven.

What Weber fraction should I assume for app performance?

For durations around one second, published duration-discrimination work puts adjusted Weber fractions near 0.17 to 0.21, meaning a change of roughly 17% to 21% before the average person reliably notices. Rammsayer and colleagues reported mean difference limens of 206.4 ms (visual) and 141.5 ms (auditory) at a 1000 ms standard with 46 participants. Use that as a planning prior and measure your own, since your task, context and attention conditions differ from a laboratory timing task.

Does Weber's law hold for prices and other non-sensory judgements?

Approximately, and usefully, but with more exceptions. Proportional sensitivity is a good first model for price, effort and quantity judgements, which is why a 10% increase reads as similar across very different absolute prices. It breaks down near psychologically loaded boundaries — a rise from €99 to €101 is only 2% but crosses a round number, and people notice round numbers far out of proportion to the arithmetic.

How many participants do I need to estimate a threshold?

Fewer than a typical attitude survey, because each participant contributes many comparisons rather than one opinion. Twenty to thirty participants giving eight to twelve forced-choice comparisons each is enough for a working estimate of a single threshold. What matters more than headcount is that the comparisons are forced-choice, that the levels are geometrically spaced, and that you include catch trials with no real difference.

Can I just ask people whether they noticed the change?

Not on its own. A direct "did you notice?" question mixes two separate things: whether the person can actually detect the change, and how willing they are to answer yes. You can move the second one substantially with wording, incentives or a changelog announcement while the first stays exactly the same. The fix is to measure the same question against a group that received no change, which is the subject of the signal detection guide linked below.

Related Resources

• Structured Questions in AI Interviews — the six question types used to build a forced-choice comparison ladder.
• Did Users Actually Notice? Sensitivity vs Criterion — why a raw "did you notice" percentage is uninterpretable without a control.
• Threshold Testing: Finding the Level Where Users Start to Notice — the efficient way to converge on a threshold.
• Statistical Power and Minimum Detectable Effect — sizing the study once you know what change is worth measuring.
• Scale Questions in AI Interviews — collecting graded magnitude judgements alongside binary detections.
• The Zero Point Moved: Adaptation and Response Shift — why a threshold measured last year may not hold this year.