How it works
The details, for the curious. SiteGuide turns your pages into a small search package, and the visitor's browser does all the work.
Building the guide
SiteGuide reads your site from its sitemap, at most two pages a second, and follows your robots.txt. It can also read a folder of built HTML, Markdown, text or PDF files.
Each page is split into sections at its headings, and each section keeps its link, so an answer can take the visitor straight to the right paragraph. Menus, banners and footers that repeat on every page are removed.
Searching
Two kinds of search run together and their results are combined:
- Keyword search finds the sections that use the visitor's words, and fixes typos against the words your site actually uses.
- Meaning search uses a small language model to find sections that mean the same thing in different words.
The guide shows the best match from each page, quoted exactly as written, with a link to open that section.
Adapting to the device
| Level | When | Download |
|---|---|---|
| Keyword search | Always | The search data (about 1–2 MB) |
| Meaning search | 2 GB memory or more, no data saver | A ~23 MB model, once, then cached |
| Browser's built-in model | Browsers that ship one | None |
If a level ever crashes a tab, the guide skips it next time. The search data loads only when the guide is opened or the browser is idle, so it never slows down your page.
Training a model on your site
For better results on everyday wording, SiteGuide can fine-tune its small model on your own content. It writes practice questions for every section, trains on them, and tests the result on questions it never saw. Training runs on your own machine by default, and a model trained on confidential documents is marked so it can't be published in a public guide.
For AI assistants
Every build also produces llms.txt, a Markdown copy of every page, and a small MCP server with search_site and get_page tools, so AI assistants can read and quote your site accurately.
How well it works
On a 150-page documentation site, with 50 questions written by hand before testing, the right page came first for 86% of everyday questions and was in the top three for 98%. Questions in words the site never uses are harder: the right page came first for about half. When nothing matches, the assistant says so instead of guessing.
These are small tests on one site. For your site, we measure on your own questions.
Get started
Enter your website's address on the home page. We read up to 25 public pages and give you a preview link of your own site with the site assistant added, plus a working MCP server address. Previews last 7 days.
To go live, you add one line of code to your site. We can do the rest for you.