TroubleshootingBeta

Most widget problems are one of a handful of cases, and each has a clear sign. Find the symptom below. You can also describe it to the setup assistant: it reads the widget's actual settings before it suggests anything.

The chat does not load on my website

This is almost always the domain allowlist, and it is the one problem the test widget cannot show you, because the test widget does not run on your site. Open Install first: pages where the widget could not load are listed there with the reason, reported by the loader on your site.

What Install reports	Cause and fix
domain not on the allowed-domains list	Click Allow domain, or add it under Security. It applies immediately. See Allowed domains.
page without a domain	a file:// page or sandboxed iframe. Test over http://localhost, see Local testing.
no version has been published	publish the widget, see Publish and versions.
the page could not reach DeutschlandGPT	your site's Content Security Policy blocks the widget, see Content Security Policy.

If nothing is listed and no chat button appears at all, first check whether the widget is paused (budget cap, no credit or spending limit, see the next section): a paused widget shows no button. Otherwise the snippet itself did not run. Check that it is on the page (view the page source), that it was pasted unchanged, and that no optimisation plugin rewrote it. The browser console shows [DeutschlandGPT Widget] messages when the loader runs into a problem.

The chat opens but shows an error or does not answer

This is never the domain allowlist. The allowlist decides where the chat may appear, not whether it answers, and the test widget in the builder is not affected by it at all. Do not add domains to fix it.

Check these instead:

• Credit and budget. The widget pauses, and an open chat answers "This assistant is currently disabled", when the workspace has no credit, the workspace's monthly spending limit is reached, or this widget's budget cap is used up. Install says which. Fix the cause, then publish again: a paused widget does not switch itself back on.
• Model. A fixed model that has been withdrawn cannot answer. Choose another under Brain.
• Knowledge. With knowledge retrieval on but nothing indexed, the bot has nothing to answer from. See Knowledge sources.
• Limits. "Too many requests" and "This conversation has reached its maximum length" come from the limits.

Publishing is refused

The message names the reason and links to the fix. The full list is in When publishing is refused.

A website cannot be read

The setup assistant names the reason when a site or page cannot be taken in:

• Needs JavaScript. The page ships an empty shell and loads its text in the browser afterwards. The crawler runs no JavaScript, so it cannot read such pages. Common for portals and search masks. Upload the content as documents instead.
• Certificate error. The server does not send its intermediate certificate. Browsers quietly repair this, automated clients cannot. Your web team needs to serve the full chain. The same gap probably affects other services that fetch from your site.
• Address does not exist. Check the spelling of the domain.
• Skipped because of robots.txt. See robots.txt and domain verification.

The bot does not know something that is on my website

Ask the setup assistant to test whether the fact can be found in the knowledge base. The usual causes are a page that was never selected, a page that changed after it was indexed, or content inside a PDF. Re-reading pages that did not change cannot fix a wrong answer; see Keep it current.