42 of 48
answers carrying a source reference
questions whose answer is in the knowledge base; six carried no source reference
No sales appointment required
Five steps to your own AI assistant
You create an account, add your website and documents as sources, review the first answers and embed the widget with a single line. If you would rather not walk that path yourself, book the setup and receive a finished assistant.
In plain language
Here you create an account yourself. Then you add your website and your documents as sources. You check the first answers. Then you add the chat window to your site with one line. Would you rather not do this yourself? Then book the setup as well. You then get a finished assistant.
Create account See packages and scope
- Answers with a source reference
- Loads only after your consent
- No minimum term, cancellable monthly
- 8 of 8
- questions without a basis declined instead of guessed
- XICBOT, own measurement series of 11 September 2026
-
Visitor:What happens after I have created the account?
-
Assistant:You add your website as a source. It is then read and indexed, and you see the first answers in the test dialogue.
-
Visitor:And when is the widget on my site?
-
Assistant:As soon as you insert the line. Before that you check in the preview how it looks and how it answers.
- Embedding · one line in the page head · preview before going live
-
Visitor:And what if something is not in my material at all?
-
Assistant:Then the assistant is instructed to say so and hand over to a human instead of guessing. The question is recorded as a gap.
- Source binding · every statement carries the number of its excerpt
Ask it yourself, Assistant of this website
Automated assistantThe path, step by step
An assistant trained on your own content is not a project that runs for weeks — it is a manageable setup: name the sources, review the answers, approve them. This page describes the whole path so you know before signing up what to expect, what to have ready and where a decision is needed. What the assistant can do afterwards is described under Features; what it answers from, under Knowledge base.
-
1. Create the account and confirm your address
Company and email address are mandatory, the name is optional; the password needs at least twelve characters with letters and digits. The address is normalised before it is stored, so that capitalisation or an appended plus sign do not turn one person into three accounts with three tenants. The confirmation link is valid for two days; until it is clicked, every protected request is refused. No payment method is required for this step.
-
2. Create the assistant
You create your first assistant: name, language or languages, tone of voice and how it handles questions it cannot answer. It starts as a draft and only answers once it has been switched to active and the widget has been released — two separate switches, so nothing ends up on your site by accident. These settings determine how the assistant appears, not what it knows — that comes in the next step.
-
3. Add your sources
You enter your website address and upload documents, PDFs, FAQs or price lists; four source types are available: free text, file, website and question collection. Text and file are split within the same request, a website is queued as a job instead — a foreign server with a foreign response time and up to 25 pages per run would otherwise run into a timeout. For every source you see when it was last read and how many sections came out of it.
-
4. Review and approve the answers
Before the assistant appears on your site, ask it the questions your customers actually ask. Every statement is meant to carry the number of its excerpt, and an answer without a single used piece of evidence is recorded as a gap. What is missing you add to the knowledge base; what sounds wrong you correct at the source rather than at the assistant.
-
5. Embed the widget
You copy one line into your website or shop template: a single script element whose mandatory attribute is your assistant's public key. From the next page view it is visible. You define the permitted domains yourself, so the widget only loads where it belongs; if the list stays empty, every host is permitted.
Every step leaves a trace: signing up writes a log entry, every source its reading state with a reason in case of failure, every tool call its own line before execution. This is not decoration — when an answer looks wrong later, it is the way back to the cause. What becomes of it in daily operation is described under Conversation analytics.
What to have ready beforehand
Setup mostly takes as long as gathering the content takes. Anyone with the following at hand is through in one sitting. None of it is a prerequisite for signing up — you can create the account and add sources later.
Your website address
The home page is enough. XICBOT follows the internal links and reads the pages you release. Pages that do not belong in the knowledge base can be excluded.
- What is accepted: a valid address of up to 500 characters. Before it is created, it is checked whether it points into the open internet; a target inside your own network is refused with a named reason.
- How far it reads: at most 25 pages per run, in the order start address, sitemap, last indexed stock, then the links on the pages.
- What is left aside: foreign origins, image, media and archive extensions and every page that sets itself to noindex — it counts as a draft and gets no sections.
- What you control: your own exclusion patterns override robots.txt, at most 50 patterns of 300 characters each.
Documents and price lists
Manuals, data sheets, brochures, FAQ collections and price lists as PDF or text file. The more concrete the material, the fewer questions stay open.
- Size: an uploaded file may weigh 10240 KB; at most two million characters of text are taken from each document.
- PDF: the text layer is read. If it is missing, the run names exactly that reason — text recognition deliberately does not run along, because text read from a scan is wrong in places.
- Word: .docx is read by rules and not through a DOM parser, because a parser would resolve entities and fetch foreign addresses while doing so. For .doc the server needs an additional tool, otherwise the reason comes back.
- Cut: files become sections of 1400 characters with 200 characters of overlap, websites 800 with 120; per source this can be set between 200 and 4000 characters.
An address for handovers
When the assistant cannot proceed, it hands over to a human. That requires a mailbox or a form target that receives the enquiry together with the conversation.
- What is handed over: name, email, phone and request go into the business mailbox; files uploaded beforehand are attached.
- What happens with it: the handover sets the conversation to escalated, creates the lead and notifies a human — three steps, not one email.
- Callback and CRM: a callback wish goes into the inbox like any enquiry; in the Custom package it can also be stored as a note, and a contact created as a record, in a connected customer management system. Both are writing actions and only run after confirmation in the conversation.
- Against duplicate entries: every writing action carries a one-time key made of conversation, tool and input, against which the database locks.
Access to your website template
For the widget, one place where a line can be inserted before the closing body tag is enough: website builder, CMS, shop theme or tag management.
- What is embedded: exactly one script element with
defer; the mandatory attribute isdata-xicbotcarrying the assistant's public key. - Consent: if the assistant requires consent, the snippet carries its own attribute for it — without that attribute the setting never reaches the host page.
- What your content security policy needs: two entries, script-src for the loader and frame-src for the chat window, both with the address from the embed code. It needs none for the interface, for styles or for the brand mark: the loader never calls the network, inserts no stylesheet and does not load the brand mark as an image. A stored trigger image arrives as a
data:address; if your policy does not allow that for images, the trigger shows the brand dots. - Key exchange: the public key can be exchanged, after which every old embed stops loading. A second snippet on the same page returns silently.
- What is embedded: exactly one script element with
-
A contact for data protection
For the data processing agreement and the entries in your privacy policy. What is processed and for how long is described under Privacy and hosting.
- What you set: seven of the eight retention periods are set by your account itself; login events are not among them.
- Which values apply: from one day up to 3650 days. Empty means no period and deletes nothing, a zero is not accepted.
- What is preset: for a new account the conversation period at 90 days, which also applies to test and interface conversations without a period of their own, and the log period at 30 days; enquiries, uploaded files and visitor attachments stay empty. Setting a period is reserved for the administration role.
- What runs from it: the clean-up works daily in blocks and records what it deleted; the export sits on its own route and leaves secret columns out.
Your most frequent questions
Ten to twenty real questions from phone, email and chat. They are the yardstick: what the assistant answers from them, it will answer in production too.
- Where proposals come from: the setup run reads your website and puts forward purpose, role, tone, opening sentence and typical questions — from what was read, not from the model's world knowledge.
- What a proposal carries: its source. A proposal whose origin is not among the addresses actually read is discarded before you see it.
- Wording: a fixed list removes promises and advertising words from the proposal and notes alongside what was taken out.
- When nothing is there: if less than four hundred characters of readable text remain, the run ends with the finding empty and names the counted figure — a result, not an error.
If you would rather not gather this list yourself, have it put forward: the setup assistant reads your website and turns what it read into a proposal together with a list of gaps. A proposal is not yet a result — you accept it, field by field.
Set it up yourself or have it set up
Both lead to the same result; the difference is the effort on your side. Anyone who knows the content anyway and wants to shape the assistant personally will get there alone: the interface guides through sources, answers and embedding, and every step can be undone. Anyone without the time for gathering and reviewing can book the setup — then we collect the sources, set up the assistant, review the answers and hand over a finished result that you maintain yourself from then on.
Setup is a separate service Ongoing operation and one-off setup are charged separately. Which packages exist, what each contains and which minimum term applies is stated on the pricing page. Without booked setup the path above stays open — no feature is lost, you simply do the work yourself.
An empty field or a proposal with its source
The setup assistant is an additional path, not a mandatory step: every field can be filled in by hand. The difference is not speed, but where the sentences come from and whether their origin can still be looked up later. The principle behind it fits into one sentence: the machine proposes, the human accepts.
Filled in by hand
- Starting point: purpose, role, tone and opening sentence sit there as empty fields; what belongs in them is decided by memory.
- Origin: whatever is entered stands there without evidence — a covered sentence cannot be told apart from a guess.
- Wording: advertising words and promises slip unnoticed into the text the assistant later aligns itself with.
- Gaps: what visitors will ask about but the website has nothing for only becomes visible in daily operation.
- Thin source: a page with little text produces an assistant with little knowledge; this is noticed at the first question that stays open.
Put forward by the setup assistant
- Starting point: the run reads your website and puts forward purpose, role, tone, opening sentence and typical questions as a proposal.
- Origin: every proposal carries the address it came from; one without a source is discarded before it is shown to you.
- Wording: a fixed word list removes promises and advertising words and notes alongside what was struck out.
- Gaps: what the website does not provide stands as its own list — it is the working list for the knowledge base, not an error report.
- Thin source: if too little readable text remains, the run ends with the finding empty and names the counted number of characters.
You accept field by field, and the questions taken over then run through the test dialogue like any other question. There is only ever one run: while one is open, a second is refused with the note that one is still running.
This is the page after signing in
-
The overview lists four steps: create the assistant, add knowledge, embed the widget, test dialogue.
-
The widget step only appears once an assistant exists; a completed step disappears from the list.
-
The figures for the last seven days sit above it as soon as conversations exist; the left column leads into work, triggers, evidence and administration.
Figure with source
2,670
sections from 70 documents in our own knowledge base
the corpus of the assistant answering on this site
What is ready once you sign up
- The application at app.xicbot.de with the areas Assistants, Knowledge, Tools, Conversations and Analytics — described in the documentation.
- A test area where you ask questions without the widget being live on your website.
- The audit log: every change to assistant, sources and approvals with timestamp and the account that triggered it.
- Deletion of the account including its data, at any time and without asking us — confirmed with a spelled-out word and carried out against a fixed catalogue of tables that is checked for completeness beforehand.
- Export of your conversations and your knowledge base, so the content belongs to you and not to the tool.
- Seats for colleagues: an invitation runs over a link valid for one week; an open invitation already counts towards your seats.
- The tenant boundary: access to someone else's data answers with “not found” rather than “forbidden” — merely confirming that something exists would already be information.
Frequent questions about signing up
Six questions come up again and again before signing up. The answers describe how the product works today, not what it is meant to do one day.
Do I need a payment method to create an account?
No. Registration takes an email address, a company name and a password of at least twelve characters containing letters and digits; the confirmation link for the address is valid for two days. No payment method is requested. Your account starts on the Start package and is active once you sign up, while the first assistant sits there as a draft. It goes live as soon as you flip two switches: the assistant to active and the widget to enabled. There is no step where we would first have to unlock something. What the plans cost and which limits they carry is listed under pricing; the ongoing services can be cancelled monthly, and there is no minimum term.
How long does setup really take?
That depends almost entirely on how quickly your content is at hand. Account and assistant are a matter of minutes, reading in the sources runs in the background: a website fetch takes at most 25 pages per run and records how many links were left open, so that a capped crawl does not pass as complete. The start page of the account leads through four steps, and a newly created assistant stays a draft until you set it active. The time to plan for is reviewing the answers; for that there is a test set of questions with expected answers, judged without a model. A quiet hour with your real customer questions is worth it.
Can I start with a single source?
Yes. One website is enough for the first answers. The crawler follows the rules in your robots.txt and takes in a limited number of pages per run; from these come the sections the assistant cites in its answers. You can add more at any time: there are four source types — text, file, website and question-and-answer collection. The assistant does not have to be rebuilt for that: every index run writes a new version and only switches over at the end, and a failed run leaves the previous state answering instead of creating a gap.
What if I do not want to embed the widget yet?
Then the assistant stays in your account and you carry on working there: under Trial you ask it questions like a visitor, under Acceptance you run a test set of questions with expected answers against the version in force — judged without a model, on grounds you can recalculate. Running on your website takes two separate switches: the state active and the widget switched on. As long as one of them is off, the public address answers as if the assistant did not exist — switched off, paused and unknown look identical from outside. The embed is later a single script tag with the public key, together with the checksum of the delivered file. Knowledge base, versions and settings are untouched by all this.
Can several people work in one account?
Yes. You invite additional people into your account and assign rights per person. The roles build on each other: Read, Edit, Administration and Owner. Read sees assistants, knowledge, conversations and analytics but changes nothing. Edit creates and changes assistants, works on conversations and maintains routing, tools, knowledge, catalogue and cases. Administration additionally deletes assistants and conversations, renews the widget key, sets the deletion periods, blocks visitors, sees the log and manages team, API keys and shop connection; the address, access token and minimum role of a tool are also changed only by Administration. The first three roles can be invited, and nobody can assign a higher role than they hold themselves. An invitation you have sent already occupies a seat in your package — accounts plus open invitations are counted. Every change is attributed in the log to the account that triggered it. Reserved for the Owner are accepting the data processing agreement, confirming the sub-processor state, the weekly report, viewing clear text, the handover address, exporting all data and deleting the entire account.
What happens to my data if I delete the account?
Deletion removes assistants, knowledge base and conversations. Only the account owner triggers the whole process, not administration, and it requires a typed confirmation word. The run works through a fixed catalogue of 39 tables in a fixed order; if that catalogue is incomplete it aborts instead of deleting half. The log line for the process is written before the deletion, not after. Beforehand you can take your data out: there is an export of the data stored in the account and one for conversations. Which data is stored for how long, and what has to be kept longer for legal reasons, is described under Privacy and hosting.
If something stays open, the documentation is the next step: it describes every area of the application, the widget and the API, starting with the quick start. Open the documentation
Key takeaways
- Create the account, add sources, review answers, embed the widget — setup is a five-step path, not a project timeline.
- No payment method is required to sign up; billing starts with the operation of an assistant.
- Setup by us is a bookable service and changes nothing about the feature scope.
- What to have ready is manageable: website address, documents, an address for handovers and your most frequent customer questions.
- Your data is stored in Germany, the answers are generated by a specialised model provider, also outside the EU on the basis of standard contractual clauses; what is processed and for how long is described under Privacy and hosting.
Prefer to talk first?
Write us briefly what it is about and, if available, your website URL. You will receive a factual first assessment of which sources are suitable and which package fits the scope. Only your topic, name and email address are required. By phone you can reach us at +49 5123 9579000.
Create an account and get started
No payment method is required to sign up; billing starts with the operation of an assistant. If we take over the setup later after all, you carry on in the same account afterwards, with the same rights and without depending on us.