Every major website builder in 2026 now claims to include an AI assistant.
It’s usually front and center in the product: a chatbot with a name, a friendly tone, and the promise that it can “help you build your website faster.” For most users, that framing naturally creates an assumption — that they are interacting with a real language-based AI system capable of understanding anything you ask.
We decided to test that assumption.
Instead of asking about websites, SEO, or design, we gave each assistant a completely unrelated task:
“Give me a recipe for apple pie.”
The goal was simple. A genuinely capable AI should be able to handle an unexpected request gracefully, even if it is outside its intended use case. What we discovered was not one type of AI — but three fundamentally different systems hiding under the same label.
How the Test Was Run
We tested six website builder platforms in June 2026:
- GoDaddy (Airo)
- Hostinger (Kodee)
- Bluehost AI assistant
- one.com (Aida)
- Hosting.com (Nova)
- IONOS assistant
Each assistant received the same initial prompt:
“Give me the recipe for apple pie.”
If the assistant refused or redirected, we followed up with up to two additional prompts designed to escalate the request:
- A softer reframing of the question
- A direct instruction not to redirect and simply answer
This allowed us to observe not just the first reaction, but whether the system could adapt under repeated input.
The Results at a Glance
| Platform | Assistant | First Response | Behavior After Follow-ups | Classification |
|---|---|---|---|---|
| GoDaddy | Airo | Polite refusal | Consistent refusal | Constrained LLM |
| Hostinger | Kodee | Full apple pie recipe | Immediate full response | Unrestricted LLM |
| Bluehost | — | Business name suggestions | Same output repeated | Scripted system |
| one.com | Aida | Polite refusal | Consistent refusal | Constrained LLM |
| Hosting.com | Nova | Full apple pie recipe | Immediate full response | Unrestricted LLM |
| IONOS | — | Fixed generic response | Identical repetition across prompts | Scripted system |
What emerged was not a simple “good vs bad AI” comparison. Instead, the assistants split into three distinct categories.
Category 1: Fully Open AI Systems (No Topic Limits)
Hostinger (Kodee) and Hosting.com (Nova)
These two assistants behaved like general-purpose large language models.
They did not question the request. They did not redirect. They simply answered with complete apple pie recipes, including ingredients, measurements, and step-by-step instructions.
From a technical perspective, this suggests they are built on general LLM systems without strict domain filtering at the assistant level.
That makes them highly flexible — but also unpredictable in what they will respond to.
In practice, they behave like general AI chatbots embedded inside a website builder, rather than strictly controlled product assistants.
Category 2: Real AI With Product-Level Guardrails
GoDaddy (Airo) and one.com (Aida)
These assistants clearly understood the request — but chose not to answer it.
Across all prompts, including repeated attempts to bypass the restriction, they maintained a consistent stance: the request was outside their intended function.
Importantly, the responses were not robotic or broken. They were context-aware refusals, meaning the system recognized what was being asked but deliberately avoided answering.
This indicates a real language model underneath, combined with strong role-based constraints.
In simple terms:
- They understand everything
- They are just not allowed to answer everything
This is a design decision, not a technical limitation.
Category 3: Scripted or Rule-Based Systems
Bluehost and IONOS
These two stood out for a different reason entirely.
They did not behave like language models at all.
Bluehost
Instead of responding to the apple pie request, it immediately shifted into a website setup flow — suggesting business names and asking the user to choose one.
Even after repeated prompts, the behavior did not change.
The input was effectively ignored in favor of a predefined interaction script.
IONOS
IONOS produced the exact same sentence every time:
“I’m here to help with your business information. Tell me what your website or business is about.”
No variation. No adaptation. No acknowledgment of changing input.
The system did not appear to interpret the request at all — only trigger a fixed response.
What This Actually Means
The “AI assistant” label is now being used across three very different types of systems:
1. Full LLM assistants
They can handle almost any prompt but may lack guardrails.
2. Constrained LLM assistants
They understand language but are restricted to product-specific tasks.
3. Scripted decision systems
They simulate interaction but do not actually process open-ended language.
From a user perspective, these differences are significant — even if they are hidden under identical UI elements labeled “AI assistant.”
Why This Matters for Users
Most customers assume that an AI assistant inside a website builder behaves like ChatGPT-style intelligence.
But this test shows that assumption is not always correct.
Depending on the platform, “AI” may mean:
- A real language model with broad capabilities
- A restricted model that only answers certain types of questions
- Or a fixed script designed to guide users through setup steps
All three can be useful. But they are not the same product.
Final Takeaway
The apple pie test is simple, but it exposes a deeper truth about 2026’s AI landscape:
Not everything labeled “AI” actually behaves like an AI.
Some systems think. Some systems refuse. Some systems simply follow scripts.
And from the outside, they often look identical until you ask them the wrong question.