Lesson 01: Chopping endive, laminitic Shetland ponies and other questions for the oracle

Since launching in 2023 we have taken vragen.ai live at more than forty organisations. Municipalities and healthcare institutions, but also trade publishers, industry associations, construction firms and technical service providers. Every onboarding teaches us something that is in no manual anywhere. In this series we line up ten of those lessons. We start with ourselves, because our own site turned out to be a good first lesson.

What happens when you put “Ask your question” above an AI field without saying what it is about? We know now. We read through every conversation on our own site this year, 2,389 questions in 1,589 conversations, and the odd ones turned out not to be the real problem.

This is what came in

• “Can a Shetland pony get laminitis in just one front leg?”
• “Can you come up with a pub quiz question about the GP practice in Ewijk”
• “Ford kuga sometimes makes 2 little beeps when I lock the car. Why”
• “Can two pizza ovens run off one circuit”
• “How do I get a chicken cart upright again without the wheels folding over”
• “Where is the first prison hotel in the world”
• “Give me the 5 cheapest charging stations on the A1 in France for a BMW i3 with 120 Ah”

Every single one of these on vragen.ai, our own site. Questions that have nothing whatsoever to do with us. And because we had no answer for them, those people politely got our fallback: please contact Timon Zuurmond (our account manager), here is his number and his email address. He has genuinely been called about this a few times.

First surprise, then discomfort

Our first reaction: what are people doing?

Our second reaction: maybe we were looking in the wrong place. It was not the model. It was our own homepage that was inviting these questions.

Look at it through the eyes of someone who does not know us. The name vragen.ai (Dutch for “questions.ai”) is pretty broad. It can sound like the oracle: anything goes here, they know everything here. Add a big “Ask your question” on the page, with “is the new search” underneath, and it is not that strange to think: okay, let's go. And people scan. Three words, two seconds, off they go. That mistake is easily made.

So people did exactly what it said. They asked a question. Just not about us.

A clearer entry point made an immediate difference

We rebuilt the story on our website so it is clear faster what vragen.ai is and who it is for. And we changed the entry point: from “Ask your question” to “Try vragen.ai”. That one line changes the expectation. Not because “Ask your question” is wrong in itself. But combined with the name vragen.ai and “is the new search”, the invitation became very broad indeed. “Try vragen.ai” makes it clearer what the conversation on our website is for.

Month	Conversations	Not about anything
January	154	55%
February	180	52%
March	207	53%
April	126	69%
May	169	85%
June	185	55%
July	369	27%
August (up to the 28th)	199	34%

From well over half to roughly a third. Not one rule was added to the model. All we did was make it clearer where people were and what they could use vragen.ai for.

1,589 conversations in eight months is not a huge number, by the way. We know. We are working on getting more people to find and use vragen.ai. Tips are welcome.

The same mistake, one layer up

May is the strangest month of the year: 85% of the conversations were not about anything. And it was all the same thing. “Where can I pick up my race number?” “Until what time can I collect my 5 km race number on Saturday?” “The 10k was cancelled, I'd like a refund.” “Is there a shuttle bus from Den Haag?”

We sponsor the Leiden Marathon. The marathon site runs vragen.ai, which answers these questions perfectly well. But anyone who reached us through our logo simply asked their question there.

Same assumption, different place. Here too, the context decided what someone asked. The page set an expectation, and the visitor acted on it.

The odd questions were not the real problem

Of the 2,389 questions, 1,025 were about something that has nothing to do with us. Annoying, but fixable with a clearer page.

The other 1,314 were about us. About vragen.ai, about pricing, about privacy, about integrations. And 294 of those got no answer. More than one in five people asked exactly the right question and were told we did not know.

That is not an AI problem. It is a content problem. The information simply was not on our website.

What those 294 were about:

What they wanted to know	Count	For example
What is this exactly	69	“Why are you different from asking chatGPT a question?”
What does it cost	35	“What are the prices of the different packages?”
How do I connect it	27	“Are there existing integrations with WordPress?”
How does a project run	20	“Which people and departments are involved in the implementation?”
Who is already using it	11	“Which customers do you have in education?”
Is it safe	10	“How does vragen.ai handle personal data of children under 16?”

And then some loose ones: “Do you replace the search function on a site?” “How do I get an API bearer token?” “Do you have pricing for non-profits?” “Where is vragen.ai hosted?” “If I give you all sorts of private data right now, what do you do with it?” “Yes, give me a reference municipality.”

Those questions point straight at what is still missing. They are our table of contents, really. The pricing, the WordPress plugin, the hosting and security: it is all there now, because somebody asked.

The questions decide what we write about

The unanswered questions have become our most important editorial tool. We read through them and ask the same thing every time: should our website answer this? If the answer is yes, we write it. Not as a loose line in a prompt, but as a page, because then anyone can read it and the agent can use it.

And we look further than the conversations in vragen.ai. What gets asked in customer meetings? What comes in through service tickets? What do people follow up on over the phone? Which questions keep coming back by email? Always with the same filter: does this belong on the site? If so, we write it.

We are not the only ones. The content team at Werken voor Nederland reads back their unanswered questions in exactly the same way.

That gets you three things at once. An agent that can tell people more about vragen.ai, a website that holds the information people want, and better findability on the topics people search for. That last one is a nice bonus, but the core stays the same.

We write for people, not for the model. The agent benefits along the way.

It works:

Month	Questions about us	Got no answer
January	88	99%
February	128	73%
March	171	20%
April	59	17%
May	59	27%
June	141	9%
July	494	6%
August (up to the 28th)	174	6%

From nearly everything to nearly nothing, without anyone once sitting down to imagine what visitors might want to know. They just told us.

Start with the context

If you put an AI conversation on your own site, first look at what you say around it. What is written above the input field? Is it clear where someone is and what they can ask about? The name, the copy and the place on the page together decide what someone expects.

Then look at the questions that get no answer. If a lot of them have nothing to do with your organisation, the context may not be clear enough. If the questions are about your organisation but the answer is missing, you know where your content falls short.

So a good AI conversation does not start with the AI alone. It starts with the context you give and the information you make available.