A small product team ships a tool with decent documentation and a support inbox that receives the same twelve questions on rotation. The docs answer all twelve. People do not find them, because finding them means knowing which of the four guides the answer lives in, and customers do not know that — they know what they are trying to do.
A grounded chat assistant on the docs site closes that gap without anyone writing new content. Here is how a team would set it up.
Public access comes first
Every share and embed method delivers a visitor to the same chat experience, so the document's access level decides whether they can chat at all. For a public docs site that means setting Access Level to Public on the document editor's Settings tab.
Docutrain is blunt about this if you forget. The Embed & Share tab shows a This document is not public banner naming the current level and spelling out the consequence — anonymous visitors arriving through embeds or share links will hit a login wall, a passcode prompt, or an access restriction. Public access is itself a paid-plan feature; on plans without it, the Public and Passcode tiles carry a lock icon and selecting one opens an upgrade prompt.
Getting it onto the site
There are three shapes, in two places.
Button and Inline live in the document editor under Embed & Share. A Button embed is a call-to-action button that opens the chat in a fullscreen slide-in panel; you set Button Color, CTA Button Text (default "Chat with AI Assistant") and an optional Show Document Title with its Title Position, watch the live preview, then Copy Code. Inline puts the chat directly into your page layout inside a frame your web team can size — a streamlined view without the footer or document switcher, which suits a dedicated help page.
The floating chat bubble is different: it is a saved embed, built in the dashboard's Embeds tab, which appears for owner admins on plans that include the embed widget (Business and higher). Create Embed asks for a Name — only you see it — and what to show. Choose Chat bubble and pick the Document it answers from; that target cannot be changed later, so a bubble for a different document means a different embed. Its editor sets hex colours for the Button, Header and Header text, a Screen position of Bottom left or Bottom right, the Header text (default "AI Assistant"), and an optional Greeting shown before the visitor's first question.
The reason saved embeds are worth the extra step is in the tab's own description: "Each embed keeps its own settings, and editing one updates every page that uses it — no re-pasting." Changes save as a draft and reach your site only when you Publish, propagating within about a minute. The statuses tell you where you stand — Draft, Published, Unpublished changes, Inactive — and copying the code is disabled while an embed is still a draft, because a draft's key resolves to nothing. One warning to take seriously: deleting an embed destroys its key permanently, so any snippet already pasted on a website stops working for good. Embedding on your website covers the rest.
Making it look like yours, and saying what it is
Branding lives in Owner Settings → Branding, which requires a plan with branding features. Two settings carry most of the weight: Group logo, shown in chat headers and on your landing and collection pages, and Accent color, a hex value used for buttons and highlights across branded views including embedded widgets. Set the accent to your product's colour and the widget stops reading as a bolted-on third-party thing. See branding and landing pages.
Then decide what visitors are told before they type. Two settings on the document's Messages tab do this, and they are not the same.
Show explainer intro displays a short About this assistant notice: answers come only from the material it was trained on, and it will say when it does not know rather than guess. It is off by default, shown once per browser session, and Explainer text replaces the generic wording. One detail matters here: the explainer never appears when the chat is embedded inside another website, because the host page provides its own framing. Write the equivalent sentence into your page copy or the bubble's Greeting instead.
Show disclaimer modal is a gate rather than a note — readers must accept before they can chat, with I agree and I decline buttons. For ordinary product documentation that is usually overkill and adds a click between a confused customer and their answer; for anything with contractual or safety weight it is the right tool. If both are on, the explainer comes first.
The controls that shape what people ask
Two tools turn other people's questions into signposts. Recent questions, under UI Options → Engagement, shows readers a live list of what others have asked — off by default, and it needs at least two suitable questions before anything appears. Duplicates are collapsed and questions the assistant could not answer are filtered out. Clicking an entry asks it; if the document has been retrained since the saved answer, a fresh one is generated rather than a stale replay. An optional Country flags toggle sits beside it.
Pinned Questions is the curated version. Administrators with the highest access level see a plus button beside each recent question, opening a Question Actions dialog with Pin Question and Delete Question. Pinning promotes a question so every reader sees it; deleting removes it from the recent list without touching the underlying conversation. Pinning your twelve rotation questions is ten minutes of work and does more for first-contact resolution than any amount of prompt tuning.
Escalation needs a route out. Contact form, also under UI Options → Engagement, adds a Contact button that opens a short form — your Intro text, the reader's email, and a message of up to 5,000 characters. Messages go to the addresses in Recipient email, entered as pills where the first is the To and every later one a CC, with the sender set as reply-to. Each address must confirm it is willing to receive that mail: a Send link control emails a confirmation, and until the To address verifies, the Contact tool does not appear to readers at all. Nobody is handed a form whose messages would go nowhere.
What you learn once it is live
The direct signal is thumbs-down feedback. Selecting it under an answer opens a What went wrong? panel where readers pick one or more reasons — Factually inaccurate, Too short, Not helpful, Too slow, Off topic, Confusing, Missing references — and those land in the Conversations viewer as tags. A red dot in the conversation list is your highest-value review queue, and Missing references on a support assistant usually means a doc page exists but was never cited.
Docutrain also applies automatic classifiers you do not configure. Every conversation is checked for profanity and junk content; flagged ones are marked banned, kept out of the public recent-questions list, and excluded from the group-wide admin feed, though you can audit them in the per-document view where they carry a BANNED badge. Banning is fully automatic — there is no manual ban button — and the filter can occasionally catch a short legitimate reply, such as a bare number answering a follow-up. Banning deletes nothing and the reader's own experience is unaffected apart from the question not being promoted publicly. Conversations and moderation covers the viewer and the CSV export.
For the aggregate picture, the Intelligence tab is where documentation gaps surface by name. It unlocks at 5 recorded questions. Keyword density counts the literal phrases customers type, which is often not the vocabulary your docs use. Once your organization has around 200 questions, the Topics panel adds a Content gaps list ranked by a gap score — higher means answers on that topic more often found no source, were re-asked, or were thumbed down, weighted by how often people asked. That list is a writing backlog in priority order. Turn on Exclude owner & admin traffic first so your own testing is not shaping it.
One operational limit to plan around: each organization has a daily chat allowance across all of its assistants and readers, 300 questions per rolling 24 hours by default and independent of plan tier. When it runs out, readers see an amber banner reading "This assistant has reached its daily question limit." with the time more questions become available, and the input is disabled until capacity rolls back in. It is a rolling window, not a midnight reset. If a public docs widget is going to see real traffic, ask the Docutrain team to raise the allowance before launch rather than after.