A Page data tool lets the AI see what is on the page the customer has open — their order number, amounts, status or time left to pay. The chat widget on your site reads the elements you choose and passes their text to the AI when the customer asks. No API is needed.
Use it for pages our server cannot open. An order page that only loads for the logged-in customer looks empty to a webparser or a webscraper, because they visit it as a stranger. The chat widget runs on the customer's own page, so it sees what they see.
When to use Page data
- The information is on a page that only loads for the customer — their order, booking or account page — and there is no API the AI could call.
- The customer writes in the web chat on your site. Page data is not available in Telegram, WhatsApp, Viber or email conversations.
- The chat widget is installed on those pages.
If the page is public, a Webparser or Webscraper tool is simpler. If you have an API, an API integration is the most reliable option.
Step 1 — Add a Page data tool
Go to Projects → your project → Integrations. Create an integration for your site — its base URL is your site's address, and it needs no authentication. Then click + Add tool and choose Page data.
- Function name — the name the AI calls, for example
read_order_page. - Description (shown to AI) — what the tool returns and when to use it. Say that it takes no input and that the AI should use it before asking the customer for an order number.
Step 2 — Choose the pages and the fields
Pages — one pattern per line. A pattern that starts with / is matched against the page path; one that starts with https:// against the full address. * stands for one part of the path and ** for anything: /orders/** matches /orders/ORD-12345. If your addresses start with a language, as in /en/orders/…, use /*/orders/**.
Fields to read — each field has a Label (what the AI sees the text as), a Key (filled in for you) and a CSS selector for the element. To find a selector, right-click the element on your page and choose Inspect. Prefer a stable class or element name to a long generated path. You can add up to 20 fields.
One field that covers the whole order block is a good start. The AI reads label-and-value text well, and a broad field keeps working when the page changes state — paid, expired, completed.
Step 3 — Test on a copy of the page
Our server cannot open these pages, so you test on a copy. Open the page in your browser the way the customer sees it, right-click it and choose Inspect, then right-click the <html> element and choose Copy → Copy outerHTML. Paste it into Paste the page's HTML and click Read fields. The pasted HTML stays in your browser.
Each field shows how many elements matched and the exact text the AI will get. You can also enter the Page URL to check that it matches your patterns.
Below the tester, Result-handling guidance tells the AI how to read the result — for example: "This is the customer's own order page. If they ask where to pay, point them to the payment details on their order page." Then click Save.
Payment details are always removed
Crypto wallet addresses, card numbers and IBANs are replaced with a placeholder such as [crypto address hidden] before anything leaves the customer's browser, and again on our server. The AI never sees or repeats a payment address: if the customer asks where to send money, it points them to the details on their own order page.
This is deliberate. Malicious browser extensions are known to swap payment addresses on a page, and the AI must never confirm one.
What you see in the conversation
In the conversation's info panel, the Web session section shows Page data: the latest text read from the customer's page — the same text the AI gets.
How it behaves
- The page is read once it has loaded, and again right before each message the customer sends, so the AI answers from what is on their screen.
- Collapsed sections, such as a closed "Details" panel, are read too. Form fields the customer types into are never read.
- Limits: 20 fields, 2,000 characters per field, 8,000 characters in total. Each read is kept for 24 hours.
- If the customer is not on a matching page in the browser they are chatting from, the AI is told there is no page data, and asks for the details instead.
- If a matching page shows none of your fields — usually after a site redesign — the project's Audit log (in the security settings) records "Integration tool returned nothing", so you know to update the selectors.
Which integration to choose
- Data is in the page source → Webparser.
- Public page that fills in with JavaScript → Webscraper.
- Page that only loads for the logged-in customer → Page data.
- You have an API → API integration.