Answer Length in the Assistant: Short, Complete and Sourced

An answer in the chat can fail in two ways. If it is too long, the visitor reads the beginning and skips the rest, often exactly the part where the condition sits. If it is too short, the first sentence is correct, but the deadline, the exception or the prerequisite is missing, and the visitor acts on half the information. This article shows how to set the length per assistant and which details should not drop out in any version. It explains when a list works better than a paragraph and when the assistant should point to the source instead of reproducing it. Finally, it covers how to use ratings and conversation analytics to check whether the length fits.

Enlarge graphic

Labels in the illustration

• Question in the chat: Can I still return the jacket?
• Part of the answer
• Too short
• Short and complete
• Too long
• Answer to the question
• included
• in the first sentence
• in sentence three
• Deadline and its start
• missing
• included
• included
• Condition
• missing
• included
• far down
• Exception
• missing
• own sentence
• easily missed
• Source named
• missing
• below the answer
• buried in the text
• Filler and repetition
• none
• none
• several
• Effect on the visitor
• acts wrongly
• can act
• stops reading early
• fine
• missing
• present but badly placed or superfluous

Three answers to the same question

Why the length of an answer matters

A chat answer is read differently from a page. It usually has no subheadings to skim by; the visitor typically reads from the top until they believe they have the answer and then stops. Whatever comes after that often no longer reaches them. Length therefore decides not only how long reading takes, but also which details arrive at all. A long answer that opens with background and only names the condition in the fourth sentence is shorter in effect than intended: the background gets read, the condition does not.

The answer that is too short fails more quietly. To the question “Can I return the jacket?”, “Yes, you can” is correct and still incomplete if returns are only possible within a deadline, only for unworn items or not for made-to-measure goods. The visitor has information that is true and still acts wrongly. The complaint then arrives not in the chat but weeks later in the inbox, and it cites what the assistant said. The drawing above shows both mistakes for the same question: the answer that is too short lacks the deadline, condition, exception and source. The one that is too long contains them, but puts the actual answer, the condition, the exception and the source in places that are rarely read.

Length, wording and tone: three settings

Three properties of an answer are often lumped together, although each answers a question of its own. Wording decides whether an answer is understandable: short sentences, familiar words, explained technical terms. That is the subject of the article on plain language in the assistant. Tone decides how a business sounds, formal or warm; this is described in the article on the assistant's brand voice. Length decides how much the answer contains: which details it selects, in which order it names them and what it leaves out. An answer can be worded simply and still be too long, or short and friendly and still incomplete.

Setting the length per assistant

The right length depends less on the individual question than on what the assistant is used for. Someone asking about delivery at checkout wants one line and to move on. Someone on the careers page who wants to know how the application process works will read several sentences if they make the steps clear. That is why the default length is set per assistant and not per answer. Deviations are governed by the type of question: a yes-or-no question gets the yes or no first, a question about a process gets the steps, a question with conditions gets the conditions. How shop, support and careers can be set up as separate assistants in one account is described in the article on several assistants in one account.

In XICBOT, every version of an assistant carries a “Tone” field with four values: factual, friendly, brief or detailed. Two of them govern length explicitly. With “brief”, the instruction asks for at most three sentences and no restating of the question; with “detailed”, the assistant is to explain context and stay on topic. The four values are mutually exclusive: choosing “brief” means giving up the “friendly” setting. How the assistant comes across, for example patient or warm, can additionally be described in the version's “Role” field, which also goes into the instruction. Because the setting belongs to the version, several assistants in one account can answer at different lengths, and every change creates a new version while the previous one is kept. The rules on sources, topic boundary and handover sit in the instruction independently of this. If a question touches one of the fixed restricted topics or a topic that is to be handed over, a fixed layer appends the mandatory notice to the answer, whichever tone is selected. How an instruction is structured as a whole is described in the article on the system instruction and its boundaries.

Details that are not cut

Cutting is a selection. So that it does not turn out at random, the assistant needs a fixed list of what stays in at every length. What these details have in common is that the visitor bases a decision on them: they send something back, they pay, they cancel, they come by. If one of them is missing, the answer is not short but incomplete in exactly the place that matters.

• Deadlines with a reference point: not just “within the deadline”, but when it starts and when it ends, as stated in the source.
• Conditions: what has to be met for the information to apply, such as unworn, with a receipt or after prior registration.
• Exceptions: which cases the information does not apply to, such as made-to-measure items, clearance goods or certain countries.
• Costs and who bears them: whether something is free of charge, and if not, for whom it is not.
• Form and channel: whether in writing, via a form or with a signature, where the source specifies this.
• The next step: what the visitor can do now and where to turn if their case does not fit.

What the law says about brief information

For businesses that inform consumers in the chat, the question of length is also one of unfair competition law. Section 5a(1) UWG, the German Act against Unfair Competition, opens with the words that a person also acts unfairly who misleads a consumer or other market participant by withholding essential information from them, and then names two requirements: the consumer needs the information, according to the respective circumstances, to take an informed decision, and withholding it is likely to cause them to take a decision they would not otherwise have taken. Subsection 2 defines withholding broadly; under no. 2 it also covers the provision of essential information in an unclear, incomprehensible or ambiguous manner. An answer that shortens a condition so far that it becomes ambiguous is therefore not just a matter of readability. At the same time, the law takes the medium into account:

Whether a chat window is a means of communication with limited space within the meaning of the provision has not been settled. Such a lack of space tends to be judged by whether the details could objectively be presented, not by the business's own decision to keep answers short; a chat answer can be read further and has no fixed length. A source reference or a pointer to the full terms can be a measure to make information available by other means and, under Section 5a(3) UWG, is taken into account in the assessment. The provision is not a licence to leave things out: it names circumstances to be taken into account, not an outcome. A model of which details the legislator considers indispensable even where space is limited is provided by Article 246a Section 3 of the Introductory Act to the German Civil Code (EGBGB). If a distance contract is to be concluded via a means of distance communication with limited space or limited time, the five items (Art. 246a Section 3 EGBGB) listed in sentence 1 have to be provided through it, two of them only where applicable; they include the total price and, where applicable, the conditions, time limits and procedure for exercising the right of withdrawal. Under sentence 2, the trader has to make the further details accessible to the consumer in an appropriate way.

Whether this provision applies to a chat assistant depends on whether a distance contract is to be concluded via the chat; that is a question of legal classification in the individual case. It still works as a checklist: characteristics, provider, total price, withdrawal terms as well as term and cancellation are details that even a short answer should not lose when the question touches on them. What applies to price information in the chat is described in the article on prices in the chat.

List or paragraph?

Whether an answer comes as a list or as a paragraph influences how long it feels and what gets through. A list makes parts visible that would disappear in a sentence, and suits anything the visitor works through or ticks off. A paragraph keeps connections together: because, therefore, unless. Breaking reasoning into bullet points loses exactly those connections. In the narrow chat window, one line break per item is enough for a list; elaborate formatting is not needed.

• One sentence before the list answers the question; the list supplies the details.
• One item, one line: anything that runs over several lines in the window is usually two items.
• Few items: if the list gets long, the rest belongs in the source the answer points to.
• No nested lists; sub-items are hard to tell apart in the narrow window.
• Name conditions next to the item they apply to, not collected at the end.
• If the visitor asks several questions in one message, each gets its own part, in the order in which they were asked.

Quote or point to the source

An assistant that answers from your documents has three options: reproduce the decisive sentence verbatim, summarise the content, or name the core and point to the source. Which one fits depends on how much hinges on the wording and how long the source is. A single sentence with a deadline or condition is best taken over verbatim, because any rephrasing can shift the meaning. A long set of rules such as the terms and conditions or the privacy notice cannot be reproduced in the chat; here the answer names the point that was asked about, with its conditions, and the location for the rest.

With the default setting, XICBOT answers exclusively from your documents and backs statements with the number of the excerpt used. Below the answer there is a “Sources” list with the name of the document and the heading of the section; the location therefore takes up no space in the answer text. The operator can relax this setting per version. An answer without a source reference is not blocked, which is why it pays to count source binding. In our own measurement series of 11 September 2026, 42 of 48 answers carried a source reference; the count covered questions whose answer is in the knowledge base, and six remained without one. Method and limits are set out under measurements. How the reference should look in the narrow window is described in the article on citing sources in the chat.

For information that should go out exactly as you have worded and checked it, for example on returns or cancellation, you store a maintained question-and-answer pair. If the question matches such a pair, it goes out without a model call. For these questions you thus set the length yourself: what is in the pair goes out in its wording and is not rephrased; if the question touches one of the fixed restricted topics or a taboo, the fixed layer appends the mandatory notice, as with every answer, and hands the conversation over. The pair itself should therefore contain all conditions and the words visitors really use when they ask. How pairs and sources are maintained is described on the knowledge base page.

Reading on a phone

On a phone the chat window is narrow, and while the visitor is typing, the keyboard takes up part of the screen. An answer that fits into a few lines at a desk needs several times that space there and pushes the visitor's own question out of view. Depending on the window, the view scrolls along with the answer, and the visitor sees its end first or has to scroll back to find the beginning. This leads to a simple rule: whatever is not visible at first glance on a phone, the visitor has to work for. The first line therefore carries the answer, not the introduction.

• The answer is in the first sentence; reasons and details follow.
• Short paragraphs with a blank line between them instead of a solid block of text.
• No tables in the chat: columns wrap in the narrow window and become hard to read.
• Numbers, dates and times as digits, so that the eye finds them.
• At most one follow-up question, and at the end, not several scattered through the text.
• The location below the answer, not as a separate paragraph in the middle of the text.
• Check on a real phone, not in a wide window at a desk.

Whether the length fits: measure, do not guess

Whether answers are too long or too short is shown by live operation, not by how it felt when writing. The most direct feedback comes from the visitor: in the XICBOT widget they can rate every answer with yes or no, and the analytics show the ratings in cards of their own. In the conversation view, the rating sits next to the answer it belongs to. A no on its own does not say whether the answer was too long, too short or simply wrong; the surrounding conversation often shows it. How ratings are collected and interpreted is described in the article on feedback in the chat.

Conversation analytics groups visitors' questions by topic, follow-ups included. If follow-ups pile up on one topic, that shows which piece of information carries one condition too few. Whether not-helpful ratings pile up there as well is shown by the conversations listed under the topic; the rated answer is then the first thing to check. After that, change one setting at a time: the version's tone, the section in the source or a maintained question-and-answer pair. After that, ask the same test questions again, in the same wording, and compare in the next period. What the analytics count is set out on the conversation analytics page; fixed test questions are described in the article on test cases before go-live.

What XICBOT provides for this

In XICBOT the building blocks for this sit in different places, but they work together. The “Tone” field sets per version whether an assistant answers factually, warmly, briefly or in detail. Mandatory sourcing keeps the answer tied to your documents and puts the location below the text instead of writing it into it. Maintained question-and-answer pairs fix the wording where nothing should be rephrased. Ratings by visitors and the analytics show where an answer does not land. During setup we test typical questions from your customers, check the answers against the sources and sharpen wording and tone. If you would like to work out which length suits your use, get in touch via contact.

Frequent questions