Theresa Wilkinson

Articles & Publications

Can AI Build a Usable Website?

I wanted to find out whether AI website builders could create a working prototype for usability testing. The project was a fictional home remodeling company, Leonard Remodel & Build, in Columbus, Ohio. I needed a complete website with content, photos, navigation, and interactions that would allow participants to complete realistic tasks.

The results varied. Some attempts stopped before I had anything I could test. One builder produced a prototype I could view inside its workspace, but I could not open its published links. Another produced a published, responsive prototype that I selected for participant testing.

Getting a website generated was only part of the work. I still needed to verify access, content, layout, interactions, and selected color combinations. And a prototype being ready for testing did not mean it was usable. Participant testing has not yet been completed.

What I Asked the Tools to Build

I attempted to use Google AI Studio, Figma Make, Wix, Bolt, and Lovable. My prompts evolved as I worked through the tools, so this was not a controlled comparison using identical prompts and settings.

The goal was to have AI create the company information, website text, images, and design. The prototype needed to support tasks such as finding kitchen remodeling services, looking at past projects, checking the company's reputation, contacting the company, learning about its process, and finding its background and experience.

The more detailed prompt used with Bolt and Lovable specified visible top navigation: Home, Services, Process, Portfolio, About Us, and Contact. It also requested sample projects, reviews, team information, responsive design, readable text, strong contrast, and WCAG 2.1 compliance. Links, buttons, and a sample contact form needed to work, with no real submissions, payments, or login.

I included the planned testing tasks so the tools had information about what participants would need to find and do. That gave me a concrete basis for reviewing the output: did the prototype contain the information and interactions needed for the study?

Google AI Studio: Code Without a Working Preview

The Google AI Studio output appeared to be a code project rather than a simple HTML prototype I could open and use. The preview displayed "Waiting for content." I could not locate or open a working prototype.

I attempted to locate and access the output. Getting further would have required technical setup and more technical knowledge than I expected. I did not have a working website to evaluate for the planned test.

Figma Make: The Attempt Stopped Before Generation

Prompt limits and a payment requirement prevented me from continuing with Figma Make. I shortened and revised the prompt to fit the tool's requirements, but no prototype was generated or evaluated.

This was an access and setup outcome. I cannot draw conclusions about the quality of a Figma Make prototype from this attempt because I never reached that point.

Wix: A Site Flow, With the Website Still to Create

Wix created a site flow, but it did not produce something I could use for the planned usability test. I would have needed to create the entire website, all the website text, and the images.

That left substantial work for me. My goal was to evaluate whether AI could create the working prototype, including the material participants would need to complete the tasks.

Bolt: Useful Details, but Published Links Did Not Open

Bolt created a prototype with project photos and expandable portfolio details, including client quotes. I thought the photos were exceptional. I also liked the status messages showing what the tool was doing during generation.

Trust information and credentials appeared in two prominent places on the website. The expandable project details and client quotes were features I preferred over Lovable's portfolio presentation.

However, I could only view the prototype inside the builder. Neither published link opened for me, which prevented me from using it for the planned participant test.

My review also identified issues that would have needed revisions. White text on the orange background failed the color-contrast check. Some text did not align correctly. Large sections and excessive spacing required substantial scrolling, and descriptive content needed more detail.

The human work included troubleshooting the published links, capturing screenshots inside the builder, and checking color contrast. Content, contrast, and layout revisions would also have been needed.

Lovable: A Prototype I Could Take Into Testing

Lovable produced a published website prototype with services, portfolio, process, company background, and contact information. I could open it on my desktop and phone, and its layout was responsive. The working published link made it the most practical option for my planned unmoderated usability test.

There were still improvements to consider. Descriptive content needed more detail. Trust information and credentials, such as licenses, insurance, and accreditation, could be more prominent. I preferred Bolt's expandable project details and client quotes. Lovable's footer was smaller than Bolt's.

I checked access and responsive behavior and identified content improvements for consideration. These were findings from my own review. Participants still need to evaluate how well the website supports navigation, comprehension, and trust.

What This Means for UX Researchers

These attempts showed that generating website code or a design did not consistently produce a prototype ready for usability testing. Access and setup difficulties could stop the process before participants ever saw the output. A working prototype still required human verification.

For researchers considering this approach, I would recommend:

Requesting accessibility compliance did not remove the need to check the output. Bolt's failed color combination was a concrete example of why that verification matters. My accessibility checks focused on selected color combinations; they were not a complete accessibility evaluation.

What I Can Conclude So Far

This study reflects one researcher's experience. The prompts were not identical across all tools, access difficulties limited some evaluations, and participant testing has not yet been completed. The results do not establish whether a paid plan would have improved the outcomes.

Lovable gave me a published, responsive prototype I could take into testing. Bolt offered portfolio features I preferred, but its published links remained inaccessible to me. The other attempts did not give me a prototype I could use for the planned study.

The next step is to test Lovable with participants. That will help answer the question that generating the website cannot answer on its own: can people find what they need, understand the information, and use the website successfully?

View LinkedIn Version

Back to Portfolio

Contact

Email: theresaw@columbus.rr.com

LinkedIn: theresa-wilkinson