Nobody wants to talk to your chatbot. And that’s okay.

Why people treat chatbots like a search bar and what that means for good conversation design.

5 min readMin. Lesezeit

Titelbild zum Artikel "Niemand will mit eurem Chatbot reden. Und das ist okay."

In shortKurz gesagt

Most users don’t have conversations with chatbots. They type keywords, skip greetings and pleasantries, and want an answer as fast as possible. Good conversation design takes this behavior seriously. It relies on direct answers, clear expectations and dialogs that get users to their goal without detours.

How do people really interact with chatbots?

If you read chatbot logs, you rarely see full sentences. You see inputs like these:

  • "invoice wrong"
  • "cancel contract"
  • "opening hours"

No greeting, no “I have a question about…”, often not even punctuation. Users treat the input field like a search bar.

We don’t just see this in our own projects. In April 2026, the Nielsen Norman Group published a usability study on AI chatbots on company websites. The finding was clear. Almost nobody greeted the bot, hardly anyone said thank you, and participants jumped straight into their question.

What’s interesting is how the behavior develops. Many people still start with a complete, grammatically correct sentence, almost as a test. As soon as they notice that the bot understands typos and half-finished inputs, their prompts get shorter. Typing takes effort, and nobody puts in more than necessary. The more users trust the bot, the shorter they write.

Short, fragmented inputs don’t signal a lack of interest or poor usage. They show that the bot works. And they are the normal behavior you need to design for.

Is that user error?

No. But this is exactly where many chatbot projects take a wrong turn.

The typical reaction to keyword inputs is to try to “train” users. The bot then says “Please phrase your question as a full sentence.” Or it replies to “invoice wrong” with “Sorry, I didn’t understand that”, even though the intent is crystal clear.

That’s like a shop turning customers away at the door because they didn’t greet the staff politely enough.

UX research has a basic principle for this. Systems adapt to people, not the other way around. Anyone who lands on your website wants to get something done. Nobody is there to chat. They’re looking for an answer.

In our projects, we put it like this: Design for the behavior you see in the logs. The behavior you wish for isn’t coming.

What does this mean for conversation design in practice?

From our project experience, backed by the NN/g findings, six things make the biggest difference.

1. Treat keywords as complete input

“cancel contract” is a complete, unambiguous intent. A well-designed bot doesn’t ask for a full sentence. It goes straight into the process: “You’d like to cancel your contract. Which contract is it about?” Intent recognized, next step, done.

Technically, this is a solved problem. Modern language models reliably understand fragmented inputs and typos. The real work is in the design. Which follow-up question is the right one? What information does the bot need next?

2. Cut small talk and filler

“Great question!”, “Happy to help! 😊”, “I hope this answers your question”. All of this makes the answer longer and adds nothing. In the NN/g study, participants liked the bots that skipped these phrases the most. For most people, a bot is a tool. And you expect results from a tool, not compliments.

That doesn’t mean the bot has to sound cold. Personality shows in word choice and tone. It doesn’t need filler sentences that stand between the user and the answer.

3. Answer first, details on request

Web writing has followed the inverted pyramid for decades, with the most important information first. For chatbots, NN/g goes one step further and suggests a truncated pyramid. The first answer contains only what was asked, plus the caveats without which it would be wrong. Background, edge cases and next steps move into follow-up prompts that interested users can click.

The reason is simple. The chat window is small. Even a medium-length paragraph looks like a wall of text there, and an answer that fills the whole window overwhelms people before they read it. In the study, participants dropped off at exactly these points.

In practice, this means short sentences, paragraphs of two to three lines at most, and lists and highlights when an answer has to be longer. A precise answer with three clickable follow-ups beats any catch-all reply.

4. Give specific answers instead of dodging

Users notice immediately whether an answer is really meant for them or whether they’re getting boilerplate. “That depends on various factors, please check the details on the booking page” is the fastest way to lose trust. In the NN/g study, participants switched straight to human support after answers like that.

If the bot can’t give an exact answer, an honest range is almost always better than a referral. “Depending on scope, costs usually fall between X and Y” beats “It varies from case to case”. Specific answers are also usually shorter, because they don’t need to cover every possibility.

5. State limits immediately and clearly

The most costly mistake in the study was bots that didn’t clearly say what they couldn’t do. One participant repeatedly tried to search for offers through a chatbot. The bot couldn’t do that at all, but it buried this information in long help texts. The result was misunderstandings, wasted time and frustration.

If the bot can’t do something, that belongs in the first sentence. “I can’t access contracts, but I’ll connect you with someone who can” is a good answer. “I’ll connect you with a person” at the right moment isn’t the bot failing. It shows the experience is working.

6. Set expectations in the first few seconds

Users decide within the first one or two interactions whether the bot is worth their time. The opening therefore has to answer two questions right away. What can this thing do? And can I reach a human if I need to?

“Hi! I’m your digital assistant and happy to help with all your questions” answers neither. It’s better to say specifically what the bot can do, for example “I can help with invoices, contract questions and service outages”. It should be just as honest about what it can’t do. That prevents the misunderstandings from point 5 before they happen.

What does the difference look like in practice?

We see this pattern again and again (simplified example):

Before:

User: invoice wrong

Bot: Sorry, I didn't understand that. Please describe your request in a full sentence.

User: my invoice is wrong!!

Bot: I'm happy to help! What can I do for you? 😊

User: (closes the window)

After:

User: invoice wrong

Bot: This is about an incorrect invoice. Is it the current one from July or an older one?

User: july

Bot: Got it. What's wrong with the July invoice? (Buttons: Amount too high / Unknown item / Charged twice / Something else)

Same technology, a completely different experience. The design makes the difference. The model barely plays a role.

How can you tell a bot that actually works for people?

A bot functions when it processes inputs correctly and delivers stable answers. A bot works when people choose to use it a second time. Here’s how to recognize it:

  • Users reach their goal in a few steps, even with keywords.
  • Answers are short, specific and easy to scan. Details are available on request.
  • The bot says in its first sentence what it can and can’t do.
  • Dead ends lead to alternatives instead of loops.
  • Switching to a human is possible at any time, and the bot actively offers it.

In other words, the bot respects its users’ time and their real behavior. Every word in a bot response has to earn its place.

A recent study shows what happens when this is missing: Why 74% of companies roll back their AI agents.

Conclusion: less chat, more answers

That nobody wants to talk to your chatbot isn’t bad news. It just means people want results, not conversation. The bots users trust and come back to treat every interaction as what it is: something to get done quickly. Accept that, and you’ll build bots that adapt to people. Those are the bots that get used.

Your bot runs fine technically but nobody uses it? We deal with exactly these cases all the time.

Get in touch →

FAQ: How people use chatbots

Why do users type only keywords into chatbots?

Because typing takes effort and the input field feels like a search bar. Many people still start with a full sentence, almost as a test. As soon as they notice the bot understands typos and half-finished input, their input gets shorter. Keywords don't signal a lack of interest. They show that the bot works.

Should a chatbot ask users to write full sentences?

No. "cancel contract" is a complete intent. A good bot recognizes it and asks for the next step right away: "Which contract is it?" Systems have to adapt to people, not the other way round.

How long should a chatbot answer be?

As short as possible. The first answer contains only what was asked, plus the caveats without which it would be wrong. Background, edge cases and next steps come on request, for example as clickable follow-up prompts. Paragraphs run to two or three lines at most, because in a small chat window even a medium-length paragraph looks like a wall of text.

Are buttons in a chatbot bad conversational UX?

No. In the article's example, the bot asks a short follow-up question and then offers buttons: "Amount too high", "Unknown item", "Charged twice", "Something else". Users move on with one click instead of having to spell out their problem. Details on request also work well as clickable follow-up prompts.

How can you tell a chatbot has poor conversation design?

It answers a clear intent like "invoice wrong" with "Sorry, I didn't understand that", asks for full sentences, wraps answers in filler and walls of text, dodges with referrals, hides what it can't do and offers no way to reach a human. A good sign, on the other hand, is when people choose to use the bot a second time.

Does a better language model solve these problems automatically?

No. Modern language models understand inputs better. However, response length, dialogue flow, expectation management, and the path to human interaction remain design decisions. A strong model with poor design results in an eloquent bot that is still frustrating to use. And even the best model makes mistakes sometimes, we explain why language models will always hallucinate in a separate article.

QuoteZitat

/ FAQ

The key points of this article at a glance. Your question isn't here? Then just chat with our bot!Das Wichtigste aus dem Artikel auf einen Blick. Eure Frage ist nicht dabei? Dann chattet einfach mit unserem Bot!

Ihr habt eine Projektidee?

Lasst uns reden und schauen, was wir gemeinsam machen können.

Kontakt

You have a Project Idea?

Let's talk and see what we can do together.

Get in Touch