Prompt-based experimentation FAQ

The following article details common questions regarding prompt-based experimentation (PBX).

Which LLM is used for prompt-based experimentation?

Kameleoon currently uses openAi o3 for its prompt-based experimentation capabilities.

Which LLM does Kameleoon use for image generation?

Kameleoon integrates with Recraft.io for AI-powered image generation.

Can I use the images generated by Kameleoon AI?

Yes, all images generated by Kameleoon AI are free to use commercially. They are automatically uploaded to your Image Library within the Kameleoon platform for easy access.

Can I edit the code generated by Kameleoon?

Yes—however, for now, code editing is only available after duplicating the generated variant in our code editor. We’re actively working on an update that will allow you to edit code directly from the prompt-based interface.

Can I use the Graphic editor on a variant created with AI?

No, the Graphic editor cannot be used on AI-generated variants. While we understand some edge cases may benefit from the visual editor, our prompt-based experimentation is designed to handle all scenarios covered by the Graphic editor. To avoid conflicts, editing via the visual tool is currently disabled for AI-generated experiments.

Can I create any type of variant with prompts?

You can create any variant that can be managed with front-end code. However, if your prompt requires back-end logic or server-side changes, Kameleoon will not generate that code. In such cases, a developer familiar with your back-end system will need to step in.

Can I create multi-page experiments?

Yes, you can create multi-page experiments. Prompting changes across multiple pages is supported. You can simply browse to the pages of your choice and apply prompts directly. Kameleoon will automatically combine the code into the same variation

However, we strongly recommend using the Simulation Mode to validate the complete experience and ensure that the code generated across different pages works seamlessly together.

Can I prompt changes on elements that appear on hover?

Yes, you can—as long as the HTML code for the hover-triggered element is already present in the page’s DOM. If the element is dynamically generated only after the hover interaction (for example, injected by JavaScript at runtime), Kameleoon’s AI may not be able to detect or modify it accurately.

For best results, ensure that:

• The hover element exists in the initial HTML (even if hidden)
• The structure is stable and not created asychronously

I’ve created a variation on a product page and want to test it across several product pages — how can I do that?

Once your variation has been created on the first product page, simply navigate to the next product page. Then, click the Refresh icon located in the variation box. This will re-execute the AI-generated code on the new page.

Alternatively, we recommend using the Simulation Mode to test the full experience across all relevant pages and ensure the variation behaves consistently.

Why can the same prompt lead to different outputs when tried multiple times?

Our AI is non-deterministic, meaning it doesn’t produce one fixed result for a given prompt. Instead, it explores many possible ways to fulfill your request. The same prompt can therefore produce different outputs each time you run it.

This happens because:

• Controlled randomness: The model intentionally adds variation to avoid identical answers and encourage creativity.
• Prompt interpretation: Even small differences in context can lead the model to emphasize different details.
• Training diversity: The model has been trained on a wide range of examples, so it may draw from different patterns each time.

Can I use my own design or mockup to create a variant?

Yes. You can import your own mockup or design file using the Import a file feature. This feature helps guide the AI in building a variant that closely matches your visual input.

While you can create a variant using only a text prompt, prompts alone may not always lead to the exact result you expect, because the AI is non-deterministic—meaning the same prompt can produce different outputs each time. Factors such as the AI’s interpretation of the prompt, context, and visual complexity can all influence the outcome. Providing a design or mockup gives the AI a clearer reference, making it easier to achieve the result you have in mind.

Can I create a prompt-based experiment without installing the Kameleoon snippet?

Yes; however, there is an additional step required

• Create an experiment as usual from the app.
• You’ll be redirected to the URL you selected.
• The editor will not load automatically. Use the shortcut Shift+F4 (on PC) or fn+Shift+F4 (on Mac) to launch it.

Note: You will not be able to launch the experiment until the Kameleoon snippet has been implemented.

Does prompt-based experimentation work on any website? What are the current limitations?

Prompt-based experimentation can be used on most websites, including single-page applications, but there are some limitations. Kameleoon processes information with each prompt request. To help the AI understand the page’s content, Kameleoon sends context related to the page (HTML code, screenshots, for example). However, on some pages this context can be too large and exceed the current limit. When context exceeds this limit, the request cannot be processed and an error is displayed in the console, causing the prompt-based experimentation response to fail. Our team is actively working on resolving this limitation. If you encounter this error frequently, please reach out to your Customer Success Manager.

Can prompt-based experimentation handle complex experiments?

Prompt-based experimentation is highly effective when dealing with:

• Text or content changes (headlines, CTAs, disclaimers)
• Style updates (colors, fonts, layout tweaks)
• Banners and content insertions
• Simple interactive changes (button repositioning, links, modals)

However, prompt-based experimentation can also be used for multi-step flows, galleries, or custom interactive elements. For advanced use cases, a hybrid approach is best: use a prompt-based experiment to generate the initial implementation, then refine it manually if required. This way you get the speed of automation with the precision of custom development.

How reliable is prompt-based experimentation’s code quality?

Prompt-based experiments generate clean, secure, responsive, and accessible code. Often it produces more robust solutions than manual coding, especially for modern websites and single-page applications (SPAs). QA best practices (cross-browser testing, visual checks) are also followed in prompt-based experiments.

Can prompt-based experimentation generate a variant by retrieving the content from a different page than the one I’m prompting from?

Short answer: no.

Prompt-based experimentation only uses the context available on the page you are currently on, meaning you cannot prompt the AI to create a feature by loading content or code from another URL—PBX does not browse other pages to retrieve additional content.

That said, if you provide PBX with access to an endpoint or web service, it can generate a variant that loads and uses this data. For example, if you want to add an urgency tooltip showing how many times a product was purchased in the last 24 hours, and you provide the endpoint in your prompt, PBX can generate the variant with the code needed to call and display that information.

PBX cannot automatically pull code or features from a different page of your site. For instance, if you want to add an “Add to Cart” CTA on a listing page, but the logic only exists on the product page, PBX cannot fetch and replicate it. However, if you provide explicit guidance, such as “call this endpoint to add the product to the cart”, PBX may be able to generate the variant correctly.

In some cases (for example, if your listing page includes a quick-view overlay), the required logic may already be present on the listing page while the overlay is open. In that scenario, prompting PBX with the quick-view displayed increases the chance of generating a fully functional variant, since the relevant code context might already be available on the page.