What happens when your AI chatbot doesn’t have the answer?
Today we share tests from our users — the kind of questions people actually ask when they want to challenge an AI chatbot.
Here’s one particularly interesting example.
A tester asked the exact kind of question a real visitor would:
The user asks on our 500 pages demo cooking blog :
which ingredients can I use to cook under 20 minutes?
Fabio AI Chatbot (powered by ChatGPT 5.4 from @OpenAI answered :
If you want to cook a quick meal in under 20 minutes, there are a few wonderful ingredient combinations you can try based on our guides. For a carrot-based dish, you can use potatoes, garlic, bell peppers, tomatoes, carrots, and spinach. If you prefer cooking with onions, you will need parsley, salt, onions, canned tuna, potatoes, and tomatoes. Finally, another quick 20-minute option uses paprika, bell peppers, butter, potatoes, pasta, and tomatoes!
👉 The chatbot also provides direct URLs to the relevant “under 20 minutes” recipes.
What’s interesting here:
The AI didn’t find a single, ready-made answer to this exact question.
So instead of guessing or hallucinating, it synthesized multiple relevant pieces of content from the site.
Result:
No hallucination
A structured, useful answer
And actual navigation paths for the user
👉 It didn’t have a pre-written answer — but it built one from what was available.
This is where things get interesting for real-world usage:
not just answering questions, but intelligently bridging gaps in content.
Your turn to try :
On our 500 pages demo coocking blog
Curious how others are seeing this play out 👇
How does your AI handle “in-between” queries that don’t exactly exist in your data?
Replies