An AI-built website is ready for customers when its important journeys work, its messages reflect what actually happened, and someone can recover it when something goes wrong. A finished-looking homepage is one piece of that. Before launch, follow an enquiry, booking or account action through the entire system and check the result that the customer and business depend on.
AI tools vary considerably. Some create interfaces, others connect working services, and many can help with both implementation and testing. The useful question is what your particular project does. The Canadian Centre for Cyber Security's guidance on vibe coding recommends review and testing appropriate to the application's risk. Its guidance recognises benefits as well as limitations.
The walkthrough below uses a fictional renovation company's website with a quote form and a private enquiries dashboard. It is an illustrative review plan, not a report of tests performed on a real client's site. You can apply the same reasoning to an AI-generated website or one built without AI.
Start with the promise your website makes
Write down the three actions that matter most to the business. For our fictional site, visitors should understand the service, request a quote and receive an accurate explanation of what happens next. Staff should be able to retrieve that request.
Turn those promises into observable outcomes:
| Journey | Evidence I would ask to see |
|---|---|
| Request a quote | The submitted details reach the intended business system and can be found again |
| Correct a mistake | The visitor understands the problem and keeps the information already entered |
| View an enquiry | An authorised staff member can open the correct record |
| Recover from a failure | The visitor has a usable next step and the owner knows where to investigate |
This keeps the review focused. A small brochure site may need only reliable contact details and a working form. A customer portal needs additional checks because it stores information and separates users. The checklist should follow the actual product.
Check whether the contact form really works
Submit a clearly labelled test enquiry to an inbox or account you control. Use a unique phrase, such as “Launch check: kitchen enquiry”, so you can trace the same submission. Avoid sending tests to customers or entering genuine sensitive information.
Follow the result through four places: the form, the receiving service, the stored enquiry if there is one, and the intended notification. Record the time and any reference identifier. A green confirmation alone does not establish that the owner received the message.
Also establish what success means. If the system has saved the request and queued a notification, “Your request has been received” can be accurate. “Your email has arrived” needs different evidence. If no durable record exists, you need to know how an unsuccessful notification is handled before relying on the form for leads.
For a focused troubleshooting sequence, see why a contact form can show success without sending email.
Next, try an invalid email address, an omitted required field and a longer realistic message. Check that the form explains the problem and preserves useful input. Browser validation is helpful, but MDN's form-validation guidance also requires validation on the receiving side.
Confirm where the app actually saves data
Our fictional dashboard might show a neat list of enquiries before any live connection exists. Alternatively, it might already use a functioning managed backend. Establish which situation applies by tracing a record rather than judging the interface.
Create one authorised test record and locate the matching record in the intended storage service. If the product should share data, retrieve it from a separate browser profile or another device, using an account authorised to see it. Ask a developer to identify the write operation and the read operation. Which deployment and environment receive them? Where would the business owner retrieve the record tomorrow?
Reloading alone is weak evidence of a shared backend. MDN documents that localStorage persists across browser sessions. Browser storage can be an intentional design choice, particularly for a personal tool, but it is different from a shared enquiries system.
A network response also needs interpretation: it could contain real records, demonstration fixtures or a mixture. The real-data versus mock-data checks explain how to establish a stronger chain of evidence. Sample data is useful during development; make its role explicit before customers rely on it.
Review failure states before polishing another animation
In a staging environment, repeat the main journey with a slow connection and a deliberately unavailable dependency. Watch what the visitor sees while waiting, after failure and when retrying.
For the quote form, I would ask:
- Does the submit button communicate that something is happening?
- Can a double click create duplicate enquiries?
- Does a failure preserve the message instead of clearing the form?
- Can the visitor retry without guessing whether the first attempt worked?
- Does the page offer another contact method when the service stays unavailable?
A developer should inspect how the interface decides an operation succeeded. For example, MDN explains that Fetch can return a response for an HTTP error; handling only a rejected promise is insufficient. The application must interpret its actual response contract as well.
The goal is a truthful experience. If the final outcome is uncertain, say so and provide a way to check it. Do not turn every uncertain response into a success message or repeatedly submit a potentially completed operation.
Test mobile use and accessibility as real journeys
Open the main journey on a phone, including with the keyboard visible. Try the menu, form, error messages and confirmation. Notice whether a sticky button covers the last field or whether a long address pushes content beyond the screen.
Then put the mouse aside. Can you move through the page, see where focus is, open and close a dialog, and submit the form? The WCAG keyboard criterion addresses keyboard operation. The form-label criterion addresses identifying the information users must enter.
Test a narrow viewport and zoom. WCAG's reflow guidance uses a width equivalent to 320 CSS pixels for vertically scrolling content, with exceptions for content requiring a two-dimensional layout. Ordinary enquiry fields should not become unreachable.
For dynamic messages, verify that assistive-technology users can discover the outcome; W3C's form-notification tutorial explains relevant patterns. These checks are a useful start, not a complete accessibility assessment. An automated score alone should not close the review.
Check permissions and private configuration
The fictional staff dashboard contains information that visitors should not see. Ask a technical reviewer to check the actual permission rules using test accounts and records in your own authorised environment.
For example, an ordinary account should not gain staff access simply because it knows the dashboard address. If customers have separate accounts, one test customer's private record should remain unavailable to another. OWASP's authorisation guidance recommends checking permissions on every request and applying least privilege. Hiding a button is not the whole permission system.
Review private server credentials and deployment configuration too. Know where secrets belong and who can change them. Some platforms intentionally provide restricted public identifiers or keys, so the presence of a key-like string is not automatically a leak; its permissions and intended use matter.
Have the reviewer record what was checked and what remains outside the review. For a site that handles payments or particularly sensitive records, the required work extends beyond this launch walkthrough.
Make maintenance and recovery part of the launch
Ask who owns the domain, hosting, email service and deployment account. Confirm that the business can access those accounts and that another person can find the operating instructions if the original builder is unavailable.
Choose a previous working release and describe how you would return to it. If the site stores important data, document the backup location, retention and restore procedure. Reverting website code does not automatically reverse a database change, so recovery needs to account for both.
For important data, restore a backup in an isolated test environment and verify that the expected records can be read. Record the result and any limitations. If recovery has only been documented, mark it as unverified until this check has been completed.
For our example, I would want an owner for failed enquiry notifications, a way to find stored requests and a short process for checking the form after a release. Decide which failures should produce an alert and where someone will actually see it. Keep customer content and credentials out of screenshots shared for troubleshooting.
This is practical operating work: who notices a problem, what evidence they need and what they can do next. A feature is easier to trust when its failure has an owner.
Decide what must be fixed before launch
Use the effect on customers to set priorities. The examples below are a suggested decision aid, not findings about your site.
| Finding | Suggested launch decision |
|---|---|
| Private enquiries are accessible to the wrong account | Fix and verify before exposing customer data |
| A quote form reports success without a traceable result | Fix the journey before depending on it for enquiries |
| The mobile form cannot be completed | Resolve before inviting mobile visitors to use it |
| A decorative animation is slightly uneven | Assess separately from the working customer journey |
Keep a short record of the issue, its owner, the expected behaviour and the evidence needed to close it. Recheck the affected journey after a fix. Avoid changing unrelated parts of the site while trying to confirm one repair.
A practical AI website launch checklist
Use the downloadable website launch checklist to record results. It includes customer journeys, forms, data, mobile use, accessibility, permissions and recovery. Mark an item as unverified when you do not have evidence; that is more useful than an assumed pass.
The best reason to launch is that the site's important promises have been demonstrated. AI can help write the implementation, explain unfamiliar code and prepare checks. Someone still needs to decide whether the evidence is sufficient for the customers and data the site will handle.
Related reading: Can I Export an AI-Built Website and Host It Elsewhere?
Sources
- Canadian Centre for Cyber Security: Managing security risks in vibe coding, 29 October 2026.
- MDN: Client-side form validation.
- MDN: Window.localStorage.
- MDN: Using the Fetch API.
- W3C: Keyboard accessibility, labels or instructions, reflow and form notifications.
- OWASP: Authorization Cheat Sheet.
Reviewed 13 November 2026.
Putting AI to work on your website?
Choosing a model is one part of a working feature. If your business needs a support assistant, content workflow or connection to an existing system, I can help define and build the interface, integrations and checks it needs.