How it works

A small, repeatable usability test. Run by AI agents.

Watch a few people use the site, find the few problems that hurt most, fix them with the smallest change, and test again. AI agents that act as your users run that loop for you in a real browser or phone: websites and web apps on desktop and mobile, mobile apps (Android and iPhone) on a phone. Each step below comes with an example.

The method

  1. First impression

    Before doing anything, each user looks at the page for a few seconds and answers: what is this, who is it for, and where would I click first? If the answers are wrong, the page is not saying what you think it says.

    Example

    A 5-second look at the home pageNo context · mobile
    What is it?
    “Invoicing software, I think.”
    Who is it for?
    “Accountants?” It is for freelancers
    First click?
    “Start free.”
  2. Tasks with expectations

    Each user gets a realistic goal, like “find the price for two nights” or “change my delivery address”. Before every click they write what they expect to happen. When reality differs from the expectation, that gap is the finding.

    Example

    Guest planning a weekend away · mobile Task: Find the price for two nights

    1. Expects“The price will show when I pick the dates.”
    2. DoesTaps Rooms, picks Friday and Saturday.
    3. DoesTaps “Check availability”. A form asks for an email.
    4. Says“I only wanted the price. Why do they want my email?”
    5. ResultGap between expectation and reality: that is the finding.
  3. Every problem, the worst first

    We rank everything we saw by how many sessions it hit and how badly. The report lists every problem, with the 3 worst at the top: that is what a team can actually fix before the next round.

    Example

    Problem 1 of 3 Severity: High Illustrative example

    Pricing is only linked in the footer

    4 of 6 sessions

    “Not under Product, not under Solutions… Do I have to book a demo just to see a price?”

    AI agent · Busy manager · desktop

    Proposed fix: Add “Pricing” to the top menu, between Product and Resources.

  4. The smallest fix

    For each problem we propose the least change that removes it: a label, a moved button, a visible price. Redesigns can wait. The fix comes as a prompt for your coding agent: for the top problems on every paid plan, complete and with before/after mockups on the Full study, Pro and Agency.

    Example

    Before
    ProductSolutionsResources
    Book a demo
    After
    ProductPricingResources
    Pro from $29/month Start free trial
    Example Fictional site, illustrative mockup.
    fix-prompt.md Example
    # Fix 1 of 3: add "Pricing" to the top menu
    Evidence: 4 of 6 AI agent sessions gave up looking for a price.
    
    ## Change
    - Add a "Pricing" link between Product and Resources.
    
    ## Check
    - Visible at 390px and 1440px without opening a menu.
    
    ## Do not touch
    Plan prices, checkout, analytics.
  5. Test again

    After you ship, the same profiles repeat the same tasks. The history shows what got better, what did not, and anything the change broke.

    Example

    Sessions blocked on the task “find the price”

    1. Round 14 of 6 Baseline
    2. Round 21 of 6 After adding “Pricing” to the menu
    3. Round 30 of 6 Fixed · 1 new minor issue

Session recordings

Every session is recorded. Watch it and judge for yourself.

The videos are not an extra: they are the evidence behind every finding. You see what the agent saw, step by step, next to what it was thinking at that moment.

Moments marked in this session

Illustrative example · Online store · first-time buyer · desktop · session 3 of 12
  1. Everything is recorded

    A video of the real browser or phone for every session, with the cursor or taps, plus a screenshot of each step.

  2. Synced with what it was thinking

    Next to the video you see what the agent expected, what it tried and what it said at that moment.

  3. Each finding points to the exact moment

    In the report, a finding links to 0:42 in session 3, and to the same stumble in sessions 5 and 9.

How we draw conclusions from the videos

  • A problem counts when several sessions show the same stumble, not because of one odd moment.
  • Severity comes from the consequence: blocked, major detour, doubt or minor.
  • A separate review step re-checks each candidate against the video and screenshots, and keeps it, weakens it or drops it.
  • Test limits (a click the automation rejected, a task the agent misread) are set apart, not blamed on your site.

And you can watch them yourself

Don’t take our word, or the agent’s: before you spend time on a fix, open the video and check.

User profiles, designed for each study

Before anything runs, we design user profiles from your business: what you sell, who comes to your site and what they are trying to get done. Each one has a role, a goal, a stage (discovering, comparing, deciding), the needs and objections it arrives with, and a short note on why it plausibly visits your site and what is invented.

  1. Start from your business

    Your site, what you sell and who buys it. A B2B tool, a local shop and a members’ club get very different people.

  2. One goal per profile

    Each profile comes to do something different, so together they cover the tasks that matter: compare, buy, get support, cancel, enter data.

  3. Modifiers only where they fit

    Traits like reduced vision or imprecise taps are added only when they match your real audience.

Some example user profiles

The procurement manager comparing plans, the member trying to cancel, the clerk registering a product: each profile arrives with a role, a goal and a device.

  • Online store
    First-time buyer Marta, 34 · Mobile

    Goal Check delivery costs before adding to the basket

    Modifiers In a hurry
    How they see a button

    “I’ll only add it if I can see what delivery costs first.”

  • Online store
    Returning customer Tom, 52 · Desktop

    Goal Track last week’s order

    Logs in with a test account

    Modifiers None: not needed for this audience

    “I’m already signed in. Where are my orders?”

  • B2B SaaS
    Procurement manager Priya, 45 · Desktop

    Goal Compare plans and find the contract terms

    Modifiers None: not needed for this audience

    “Is the annual price per seat or per company?”

  • B2B SaaS
    Developer Ivan, 29 · Desktop

    Goal Check the API docs and rate limits before recommending it

    Modifiers None: not needed for this audience

    “Rate limits… not on pricing, not in the docs index.”

  • Membership site
    Member Nora, 41 · Mobile

    Goal Cancel her membership

    Logs in with a test account

    Modifiers In a hurry
    How they see a button

    “There’s an upgrade button on every page. How do I cancel?”

  • Hiking club
    Visitor deciding whether to join Ray, 67 · Tablet

    Goal Find out what members get and what it costs

    Modifiers Reduced vision (simulated) Added because many of this club’s members are older.
    How they see a button

    “The grey text under each benefit is too faint for me.”

  • Back office
    Clerk Aisha, 38 · Desktop

    Goal Register a new product with its batch number

    Logs in with a test account

    Modifiers New to the tool
    How they see a button

    “The code is already in the header. Why type it again?”

  • Any site
    No-context visitor Sam, 26 · Mobile

    Goal Work out what this is from a link someone shared

    Modifiers None: not needed for this audience

    “Is this a shop or an agency? Nothing here tells me.”

Examples by type of business

  • B2B SaaS

    • Procurement manager comparing plans and contract terms
    • Developer checking the API docs before recommending it
  • Online store

    • First-time buyer checking delivery costs before adding to the basket
    • Returning customer tracking an order, logged in
  • Membership site

    • Visitor deciding whether to join
    • Member trying to cancel, logged in
  • Back office

    • Clerk registering a new product
    • Supervisor checking what is still pending

Other agents in a study

No-context visitor

Arrives from a search result or a shared link knowing nothing about you. Tells you whether your first screen explains what you do.

Returning and logged-in users

Use test accounts you create for us to try member areas, orders or settings. The model never sees the password, only that the login worked.

Optional modifiers

Physical and ability traits are not what a user profile is about. We add them only when they match your real audience: many sites don’t have older customers, so we don’t force them. When we do add one, we apply it to what the agent sees and does, not just to how it is described.

Optional modifiers
ModifierHow we simulate itWhat it tends to reveal
Reduced visionUsed when your audience includes older people or long reading sessions.Screenshots are blurred and their contrast is lowered before the agent sees them.Pale grey text, small icons, low-contrast buttons, text over images.
Imprecise tapsUsed when many visits are on a phone, on the move.Tap and click coordinates get a random offset of a few pixels.Tiny targets, links packed together, close buttons in corners.
In a hurryUsed when people arrive with a quick, concrete question.Fewer steps and less time. The agent skims and gives up after a few wrong turns.Buried calls to action, long pages before the point, slow feedback.
Low digital skillUsed when your customers rarely use the web.The agent avoids jargon, hesitates at unlabelled icons and does not assume hidden menus exist.Hamburger-only navigation, icon-only buttons, technical wording.

Devices are chosen the same way: desktop, tablet or mobile, depending on where your customers actually are.

What AI agents are, and what they are not

They are

  • AI models that operate a real browser or phone on your real site or app.
  • Recorded: video, screenshots and the reasoning behind each step.
  • Good at spotting what trips up a first-time visitor: unclear labels, hidden prices, dead ends.
  • Fast and repeatable, so you can test again after every fix.

They are not

  • Your customers. They have no budget, no history with you and no real stake.
  • A replacement for talking to people about needs, prices or trust.
  • Statistically representative. Six sessions show problems, not percentages.
  • Always right. Some findings will be wrong. The video lets you check.

Every report ends with a short script to confirm the top findings with three to five real people. That is where testing with AI agents hands over to research with humans.

Safety

  • Read-only on public sites

    Public tests are read-only by default: users navigate and read, and on third-party sites they never submit forms. With your permission, on your own site, they can sign up and confirm emails with UX Surgery test inboxes. They never send messages or place orders.

    ExampleLog: “Reached ‘Place order’. Stopping: this would create an order.”

  • Domain verification for apps

    Testing a web app behind a login needs proof that you control the domain, and test accounts you create for us. Mobile apps (Android and iPhone) run only from the build you upload.

    ExampleDNS record: TXT uxsurgery-verify=7f3c… (illustrative value)

  • No payments, ever

    Agents stop before any payment step. They never enter or submit card details.

    ExampleLog: “Card number field reached. Stopping before payment.”

  • EU data hosting

    Recordings and reports are stored in the EU, in Madrid, and deleted after 90 days.

    ExampleStudy run on 3 May → deleted on 1 August.

Security & privacy in detail

See where people get lost on your site.

Paste your URL. In a few minutes you get the 3 worst problems and one session video.

3 AI agent sessions on your site, the 3 worst problems and a video.