How to do a UX audit: eight steps, a checklist, and which findings to trust
In three points
- A UX audit is a structured review of a live product that finds where users pay a cost and ranks what to fix. Most guides build it on an expert walk-through, which is the weakest evidence in the whole audit.
- Run it in eight steps: scope, behaviour data, what users say, the walk-through, accessibility, speed, the way out, and a report in which every finding names its source.
- Three of the eight steps use free or public data you can collect today: an accessibility scan, the speed your real visitors get, and a count of the screens it takes to leave.
1. What a UX audit is
A UX audit is a structured review of a live product that finds where users pay a cost in time, effort, errors or trust, and ranks what to fix first. It reads four kinds of input together: what users do, what users say, what an expert sees, and what technical checks report.
The term gets used for three different jobs, and they produce evidence of different strength.
| Method | What happens | What it produces |
|---|---|---|
| Heuristic evaluation | One or more experts walk the product against a list of principles | Judgment. Fast, cheap, and different from one expert to the next |
| Usability testing | Real people attempt set tasks while someone records what they do | Observed behaviour, from a small sample |
| UX audit | Existing data, user feedback, an expert walk-through and technical checks are read together | A ranked list of fixes, if each finding keeps its source |
Most published guides to a UX audit put the heuristic evaluation at the centre and treat analytics as background. This guide reverses the order, for a reason given in step 4.
2. When to run a UX audit
Run one when a number has moved and nobody can say why, before a redesign is budgeted, or after a run of releases has changed the product more than anyone has measured. The usual triggers are falling conversion, rising support volume, churn that the exit survey does not explain, and a new owner who has inherited a product without its history.
An audit is a poor fit for a product with few users. Below roughly 300 active users a month, retention figures are too noisy to read and five interviews will tell you more than any scoring. An audit also cannot answer whether the product should exist. It measures the cost of using it.
3. What you need before you start
- Read access to product analytics, with a funnel for the main task.
- The last 50 support tickets, and any reviews or survey answers from the last quarter.
- A phone and a keyboard. Several checks below are done by hand.
- A free accessibility scanner. Axe, WAVE and Lighthouse all work.
- PageSpeed Insights, for the speed your real visitors get.
- A spreadsheet with one row per finding and a column for where the finding came from.
4. How to do a UX audit in eight steps
Step 1. Name the task and the scope
Write down the one job users come to the product to do, and what counts as done, in their words. "Send an invoice and know it arrived" is a task. "Improve onboarding" is a project. Audit one main task end to end before you look at anything else, and list the releases that have touched it in the last six months.
Step 2. Collect behaviour before you open the product
Pull the numbers first, so that what you see later cannot colour what you count. For the main task you want four figures: how many who start it finish in the same session, how many recover after an error without leaving or contacting support, how many new users reach first value, and how many are still active a month later.
Compare each figure with an outside reference before you call it a finding. MeasuringU's set of 1,189 tasks from 115 usability tests averaged 78% completion. Baymard's average of 50 studies puts cart abandonment at 70.22%. A checkout that loses 70% of carts is normal for the category, and the finding is the step where your funnel drops more than the steps around it.
Step 3. Collect what users say, and keep it apart
Read the tickets, reviews and survey answers and sort them by the same task. Count, for each part of the product, how many people describe it as easy and how many as hard. Record these in a separate column from the behaviour figures and do not average the two.
They will disagree. In one MeasuringU analysis, the completion rate people reported for themselves ran almost three times higher than the rate verified from what they did. Where behaviour is poor and nobody complains, you have found the most expensive kind of problem, because no survey will ever raise it. RUCF calls that silent friction. Where behaviour is fine and people complain, the fix is in what the product says, and a rebuild would be wasted. That is phantom friction.
The team does not answer for the users. "Our customers find it intuitive" is the team's belief, and it goes in the audit as an assumption.
Step 4. Walk the product last, and label what you find
Now open the product and walk the main task as a new user would, on a phone. Note every point where you hesitate, every field you do not understand, every error message that does not say what to do next. Nielsen's ten heuristics are a reasonable list to walk against.
Then label every one of those notes as judgment. Experts disagree with each other far more than the confident tone of an audit report suggests. In a 1998 study, four evaluators analysed the same recorded test sessions and found 93 problems between them. Only 20% were found by all four, and 46% were found by just one. A 2014 study repeated the exercise with nine usability professionals, and each reported on average 33% of the problems the group found.
A walk-through is still worth doing. It is quick, and it tells you where to look in the data from step 2. A finding from it is a candidate until behaviour confirms it.
Step 5. Check accessibility as a gate
Run the scanner on every screen of the main task and count the serious failures. Then put the mouse away and complete the task with the keyboard alone.
Expect to find something. WebAIM's 2026 scan of the top million home pages found detectable failures on 95.9% of them, an average of 56.1 errors per page, with low-contrast text on 83.9%. A clean scan is also not a pass. In a 2021 study of more than 13,000 pages by Deque, which makes the axe scanner, automated tests caught 57% of the issues that full audits recorded. The rest needed a person.
Report the result separately from any overall score. Since June 2025 the European Accessibility Act has made accessibility a legal requirement for in-scope products and services in the EU, and a legal requirement should not be averaged with the quality of your onboarding.
Step 6. Read the speed your real users get
Open PageSpeed Insights, enter the page where the main task starts, and stay on the Mobile tab. The panel at the top shows field data: timings from real Chrome visits over the last 28 days, at the 75th percentile. That is observed behaviour, and it is public for any site with enough traffic. Ignore the 0 to 100 score further down, which comes from a single simulated load.
| Metric | What it times | Good | Poor |
|---|---|---|---|
| Largest Contentful Paint | Until the main content is shown | 2.5 s or less | More than 4 s |
| Interaction to Next Paint | From a tap or click to the next visual response | 200 ms or less | More than 500 ms |
| Cumulative Layout Shift | How far content moves while loading | 0.1 or less | More than 0.25 |
In the 2025 Web Almanac, 48% of mobile sites were good on all three. Speed is worth its own step because people rarely name it as the reason they left. When Google delayed search results by 100 to 400 milliseconds in an experiment, searches per user fell by 0.2% to 0.6%, and nobody files a complaint about four tenths of a second. If the page is slow, fix that before you redesign the flow, because some of the drop-off you counted in step 2 may be waiting.
Step 7. Count the way out
If the product has an account, a subscription or a trial, count the screens from deciding to join to being a customer. Then count the screens from deciding to leave to being out. Retention offers and exit surveys count as screens.
exit ratio = screens to leave ÷ screens to join
1 or less: pass up to 2: review above 2: fail
Check two more things on the way. Can a customer who joined online leave online, without a call or a chat they cannot skip? And is any paid extra selected in advance, or any charge shown for the first time at the last step?
This belongs in a UX audit for two reasons. A hard exit makes retention look better than it is, so the commitment figure from step 2 cannot be trusted until you have counted. It is also where regulators look. In a 2024 sweep by consumer authorities in 26 countries, 75.7% of 642 subscription sites and apps used at least one manipulative design pattern. The reasoning and the legal references are in section 14 of the RUCF manifesto.
Step 8. Write findings that carry their source
Every row in the spreadsheet gets a source class: observed, reported, inferred or assumed. Rank the fixes by how much cost each removes, discounted by how sure you are. Keep the list to three. A report that recommends twenty changes has not made a diagnosis.
5. Which UX audit findings to trust
Trust a finding in proportion to how it was obtained. RUCF, the framework I maintain, gives each class of evidence a published weight, borrowed in spirit from the way clinical medicine grades studies.
Fig. 01
What a finding is worth, by where it came from
In practice this changes the order of the report. A funnel drop that the tickets do not mention goes above a layout problem the auditor is sure about. And an audit built only from a walk-through should say so on its first page, with a confidence figure to match.
Evidence also expires. A finding about a screen that has since been redesigned describes a product that no longer exists. RUCF multiplies the confidence of a finding by 0.8 for every release that touches the part it describes, so after three releases it is worth about half.
6. UX audit checklist
Twenty checks, in the order of the steps above.
- The main task is written in the user's words, with a success condition.
- Releases that touched the task in the last six months are listed.
- Completion rate of the main task, same session, from the funnel.
- The single step with the largest drop is named.
- Share of users who recover after an error without leaving or contacting support.
- Share of new users who reach first value, and retention one month later.
- The last 50 tickets are read and counted by topic.
- Reviews and survey answers are sorted by the same task, in a separate column.
- Every place where behaviour is poor and nobody complains is marked.
- Every place where behaviour is fine and people complain is marked.
- The task is walked on a phone, as a new user, after the data is in.
- Each walk-through note is labelled as judgment.
- An automated accessibility scan is run on every screen of the task.
- The task is completed with the keyboard alone.
- Field speed is read for the entry page: loading, response, stability.
- What users say about speed is noted next to it.
- Screens to join and screens to leave are counted on a phone.
- Leaving works in the same channel as joining, and no charge is set in advance.
- Every finding has a source class and a date.
- No more than three fixes are ranked, each with the result that would show it worked.
The free RUCF scorecard runs a scored version of this list. It asks twelve questions about what your data shows and twelve about what users say, then seven optional ones on speed and the way out, and it returns a diagnosis with a confidence figure. It takes a few minutes and stores nothing.
7. What a UX audit report should include
A report someone can act on has six parts:
- Task and scope. What was audited and what was left out.
- Evidence inventory. Every source, with its date, sample size and class.
- Findings. Behaviour and perception side by side for each part of the product, never averaged.
- Gates. The accessibility result and the exit count, reported apart from any score.
- Three fixes. Each with the metric, direction and size of change that would show it worked, written before the work starts.
- Limits. What the audit could not see, and which findings rest on judgment alone.
The fifth part is the one most reports leave out. Without a criterion written in advance, the review of the fix becomes a matter of taste.
8. Can AI run a UX audit?
An AI assistant can do steps 3 to 7 on public material: read reviews and tickets, walk the public pages, run a scan, look up field speed, and read the help page on cancelling. It cannot watch your users, so step 2 still needs your data.
An assistant's opinion of a page is judgment, and it gets the same 0.30 weight as a human expert's. Its sorting of reviews into positive and negative is inferred evidence at 0.50 until a person checks a sample. An audit run from a URL alone should therefore come back with low confidence and a list of things to measure, and you should distrust one that does not.
The RUCF agent skill is a free download that makes Claude, ChatGPT, Gemini or Grok follow these rules. It keeps the two passes apart, tags each value with its source, runs the arithmetic in code and writes a report in twelve fixed sections.
9. Common questions about UX audits
What is the difference between a UX audit and usability testing?
Usability testing watches real people attempt tasks and records what they do. A UX audit is wider and reads existing data, user feedback, an expert review and technical checks together. A usability test is one of the strongest inputs an audit can have, because it is observed behaviour.
How long does a UX audit take?
From a day to several weeks, depending on how much data has to be gathered. A self-run audit of one main task fits in about a day when the analytics and tickets are to hand. A RUCF baseline with diagnosis runs three to four weeks, most of it spent collecting observed behaviour. The phases are on the methodology page.
Who should run it?
Someone who did not design the thing being audited, where possible. The people who built a flow know where everything is and have stopped seeing its cost. If the team must audit its own work, the behaviour data in step 2 is the protection, because it does not depend on who reads it.
How often should you run a UX audit?
After releases, not on a calendar. A part of the product nobody has touched for two years can keep its old findings. A part that was redesigned last week needs new ones.
Is a heuristic evaluation enough?
As a first look, yes. As the basis for a redesign budget, no. One evaluator finds a minority of the problems and a different evaluator finds a different minority. Use it to decide where to measure.
Notes and sources
The framework used here is RUCF v2.3, which I wrote and maintain under CC BY 4.0. Its weights and thresholds are calibrated judgments, published with their limitations in the manifesto. The benchmarks below come from the sources named, not from RUCF.
- The evaluator effect in usability studies: problem detection and severity judgments, Jacobsen, Hertzum and John, 1998
- How large is the evaluator effect in usability testing?, MeasuringU, summarising Hertzum, Molich and Jacobsen, Behaviour & Information Technology, 2014
- What is a good task-completion rate?, Jeff Sauro, MeasuringU
- How reliable are self-reported task completion rates?, Jeff Sauro, MeasuringU, 2015
- Cart abandonment rate statistics, Baymard Institute (average of 50 studies)
- The WebAIM Million: 2026 report on the accessibility of the top 1,000,000 home pages, WebAIM
- Automated testing identifies 57 percent of digital accessibility issues, Deque, 2021 (a study by the maker of the scanner)
- Directive (EU) 2019/882, the European Accessibility Act
- Web Vitals, Google, for the three speed metrics and their thresholds
- Web Almanac 2025, Performance, HTTP Archive
- Speed matters for Google web search, Jake Brutlag, Google, 2009
- FTC, ICPEN, GPEN announce results of review of use of dark patterns, Federal Trade Commission, 2024
Iordanis Passas is a creative director and type designer based on Syros, and the author of RUCF.
Related: UX metrics: how to measure user experience · Why expert review is the weakest evidence in UX · Accessibility is a gate, not a score · RUCF agent skill