Importing your existing pages

Context is the part that takes longest, and most of what belongs in it is already written — on your services page, your pricing page, your FAQ. Import reads those pages and drops their text in for you to trim.

How to use it

On the bot's card in the portal, put a public page whose text you want the agent to know — pricing, FAQ, services; it does not have to be the homepage — and press Read page(s). One address reads that page and the pages it links to on the same site (robots.txt obeyed, a dozen pages at most). Give several addresses, comma-separated, to read exactly those and follow no links.

The text is appended to whatever is already in Context — nothing you wrote by hand is overwritten. Read it, delete what a visitor would never ask about, save the draft, Preview, then Publish. Nothing changes for visitors until you Publish.

If Persona or Guardrails is empty, the checked options can draft a starting point from the imported pages. They never overwrite text already in those fields, and you should review them before publishing.

It is a copy, not a live feed

This is the part worth understanding. Import runs once, at setup, into a box you can read and trim. Your agent has no ability to fetch a web page while it is answering, which is what makes it safe to put in front of strangers: it cannot be talked into fetching something mid-conversation.

So import is a one-time copy. Edit a page next month and Context does not change; import it again, or edit the box directly. Imported knowledge remains text you can see and change on that screen.

Trim it

Context is sent to the model on every message, so its size is a standing cost rather than a one-off. Your plan includes 20,000 to 100,000 characters per bot (the 30-day trial matches Starter's 40,000), and saving more is refused rather than quietly cut short — a truncated box would leave your agent confidently unaware of whatever happened to be at the bottom.

Import already drops navigation, footers and scripts. What is left to cut is usually testimonials, awards, and paragraphs about your values. Keep the things people ask about: what you do, prices, hours, areas, what you do not do.

What it will not read

• PDFs, images and downloads. Only HTML pages. Other file types may simply be reported as unreadable; paste their text yourself.
• Pages behind a login. It fetches them the way an anonymous visitor would.
• Private or internal addresses. Anything on your own network is refused.
• Redirects. Give the address the page actually lives at.

If an address cannot be read, the reply names it and gives the safest available reason. An internal hostname may look like any other unreachable address because the service does not reveal private network details.

A caution about summarising

It is tempting to run the imported text through something that shortens it. We tested that carefully and do not do it by default. On a real page it removed 84% of the text — and turned a list of seven named examples into "7 worked examples", which reads perfectly well and is useless to someone asking whether you cover their case.

Shortening drops specifics first, and specifics are the entire reason the agent is useful. Trim by hand, deleting whole paragraphs you do not need rather than compressing the ones you do.