Free
$0per month
- 1
- Bots
- 25
- Documents
- 100
- Messages / mo
- Gemini 3 Flash
- “Powered by Grounded” removed (not included)
- Content Gaps and gap drafts (not included)
- Email escalation (not included)
- MCP server (not included)
- Reranking (not included)
Support deflection for B2B SaaS
Upload your documentation and get an assistant that answers customers on your own site, cites the exact section it used, and tells you what your docs are missing.
Free plan, no card. One script tag to install.
Hi — I am Grounded, running on Grounded's own documentation. Ask how it works, what it costs, or how to install it.
Answers come from Grounded Docs’s documentation, with the section cited.
This chat is Grounded, running on our own docs. Ask it anything.
These numbers come from the bot above and move as people ask it questions.
Formats it reads, sites it installs on, clients it connects to. Not customer logos.
How it works
Read
Upload files, paste text or point at a page. Markdown, plain text, PDF and DOCX are parsed on our servers; a URL is fetched and stripped of its navigation.
Split
Each document becomes sections of a few hundred words, cut at the headings the author wrote. A citation points at one of these, so the split is what makes a citation precise.
Index
Every section gets a vector embedding for meaning and a full-text index for exact terms. Support docs are full of error codes and setting names that a purely semantic search blurs.
Answer
A question runs against both indexes, the results are fused, and the top passages reach the model. It writes from those, cites them, and says so when they do not contain the answer.
The loop
Every other step is yours to approve. Nothing reaches your visitors that you have not read.
Admit
A confidence signal on every answer, a visitor’s Helpful or Not helpful under it, and a handoff link to a person when the docs run out. Two rates on your overview: what the model claimed, and what visitors confirmed.
Gap
Unanswered questions and answers visitors rejected are grouped by meaning and counted, with the pages they were asked on. A to-do list for your docs, written by your customers.
Draft
One click drafts the missing section from the nearest passages. Anything it cannot find is a placeholder you fill in, and publishing is blocked until you do. You read it; then the bot cites it.
Test
A test set of questions with the facts a correct answer must mention — and some the bot must refuse. Run it on demand, on this model or another, and read the regressions first.
What you get
Markdown, plain text, PDF, DOCX and public URLs go through the same parse, split and embed steps. Each file shows its own state while it indexes, and a failure says why and offers a retry.
Citations are first-class. A chip under the answer opens the exact passage the answer was built from. Not the document, not a search result, the text.
Choose Add URL in the bot’s Documents tab and enter a public page. Grounded fetches it on its own servers, strips the navigation and indexes the article text.
Every question the bot could not answer — and every answer a visitor rated not helpful — is grouped with its siblings and counted. Draft the fix, mark it addressed, export the rest as CSV.
One click drafts the missing section from your own documentation. What it cannot find becomes a placeholder, and Publish stays off until you fill it in. Nothing is applied on its own.
Is there an on-premise option?
Grounded runs on one server from a Docker image next to its database. Self-hosting is documented under Plans, limits and your data. [fill in: whether a support contract covers it]
Questions with the facts a correct answer must mention, and some the bot must refuse. Run them on demand, compare two models on your own docs, and see regressions first.
The launcher lives in a closed shadow root and the chat in an iframe from our origin, so nothing of ours runs in your page. Under 5 KB, and it fetches nothing until the first click.
<script
src="https://app.grounded.dev/widget.js"
data-bot="pk_your_bot"
async
></script>The same knowledge base is exposed over the Model Context Protocol. Any MCP client can search it and gets the same cited answers.
tool search_knowledge("rate limits")
→ passages, each with its section and source
Where it fits
A search box that answers instead of listing pages.
Put the widget on your documentation site. Visitors ask in their own words and get the section, not ten links.
The ticket that was already answered on page 14.
Embed it on your help page. Questions the docs cover never reach the queue; the ones they do not become Content Gaps.
Open the chat from your own Help button.
Two lines of JavaScript open the widget with a question pre-filled, from any screen in your app.
Runbooks, policies and onboarding, searchable by meaning.
Point a bot at your internal docs and let the team ask from Claude Desktop or Cursor over MCP.
Pricing
Every limit here is the one the product enforces. The same table drives the gates in the app.
$0per month
$39per month
Sign up free, upgrade from Billing.
$129per month
Sign up free, upgrade from Billing.
Questions
Create a bot, upload the docs you already have, and ask it the question your support queue got this morning.