Skip to main content
Privacy is built into how the guide handles every turn. Sensitive values are tokenized before they reach the model, telemetry is scrubbed before it leaves your app, and consent follows the posture you set for your workspace. This page explains each layer so you can position it confidently with your own privacy team.

Redaction at a glance

Before any text reaches the model or the guardrail checks, Puntego scans it and replaces well-defined sensitive values with stable, content-addressed placeholder tokens. Matching tokens map back to the same original value within a turn, so the model can reason about “the same email” or “that card” without ever seeing the real data.
Auto-detection covers the structured, high-confidence patterns above. It does not infer names, postal addresses, or free-form text, because those have no reliable signature to match on. If a workflow collects that kind of information, scope it into a dedicated field that your own systems handle, rather than relying on the guide to recognize it in conversation.

How the redaction round-trip works

Redaction is reversible by design. The original values never leave your app boundary, yet the visitor still gets a complete, natural reply. Because tokens are content-addressed, the same input always produces the same token, which keeps replies coherent while the underlying value stays private to your environment.

Page vision

The guide reads the page structure first. When that is not enough to pick out a control, it can send a screenshot of the visitor’s current view to the AI provider so it can see which control to point at. Page vision is on by default on every plan. To turn it off for your workspace, go to Dashboard → Settings → Privacy → Page vision. The guide then relies on the page structure alone. Before a screenshot leaves the visitor’s browser, Puntego hides:
  • Every editable form control (inputs, text areas, selects, and editable regions), whatever its label or type.
  • Page text that matches the patterns in Redaction at a glance.
  • Everything marked with data-puntego-ignore, as described in the next section.
  • The Puntego widget itself and every iframe on the page.
Only the visible viewport is captured, and nothing is captured when the visitor’s browser asks to save data. If Puntego cannot finish checking the page for private content, it sends no screenshot at all.

Form fields in the page structure

The page structure the guide reads names phone, email, and password fields by their label only, for example “Security contact phone”, so the guide can point at the right field. It never includes what the visitor typed, the field’s placeholder, or autofill data. If a label repeats the field’s value, the field is left unnamed. The guide can point at a password field, but it never types into, fills, or submits a password. The visitor always enters it themselves. Once a form holds a typed password, the guide doesn’t press any of that form’s buttons either: it points at the button and asks the visitor to press it. These fields carry no label at all:
  • Fields marked for payment-card or one-time-code autofill (autocomplete="cc-…" or autocomplete="one-time-code").
  • Phone, email, or password fields whose label asks for a card or bank number, a security code, a government ID, or a one-time or recovery code.
To keep a field out of the page structure entirely, mark it data-puntego-ignore.

Keep parts of your page private

Add data-puntego-ignore to any element that Puntego should never read or send. The attribute covers the element and everything inside it, including open shadow roots:
An ignored element is:
  • Left out of the page structure the guide reads, so it is never described, pointed at, clicked, or filled.
  • Left out of the text of any paragraph, button, or label around it, so its content is not repeated there either.
  • Hidden in every screenshot. It keeps its space in the layout but shows nothing.
  • Never sent as highlighted text. If a visitor’s text selection reaches into it, the selection is not sent.
  • Excluded from target capture in the guide editor.
Any value ignores the element, including an empty one, except data-puntego-ignore="false", which leaves it in scope. That helps when a template always renders the attribute. It cannot bring back an element inside an ignored parent.
data-gp-ignore="true", data-gp-private="true", and data-sensitive="true" are older names for the same attribute and behave identically. Use data-puntego-ignore in new markup.

Telemetry stays clean and isolated

Operational events are scrubbed of PII before they ever leave the app, so traces and analytics capture what happened without capturing who it happened to. Every event is workspace-isolated: one workspace’s telemetry is never visible to another. The result is the observability you need to run the guide in production, with none of the raw sensitive payloads. Set the consent posture for your workspace from Dashboard → Settings → Privacy. The guide follows the mode you configure, consistently, for every visitor. A few principles make this predictable:
  • GPC always wins. A Global Privacy Control opt-out is honored independently of location, regardless of the configured mode.
  • No geo gating. Consent is driven by the mode you choose for the workspace. A missing or undetected visitor country does not force the strictest mode, and never silently overrides your configuration.
Pair your consent mode with verified domains and a tight content policy so the guide only ever boots where you intend. See Security for the full trust model.

Next steps

Configure consent and privacy

Set the consent mode and launcher behavior for your workspace.

Review the security model

See how domains, tokens, and isolation protect every request.