Troubleshooting

The widget does not appear

1. View source on the live page and confirm the script tag is actually there. Caching layers, CDN rules and theme editors that silently drop tags are the most common cause by a distance.
2. Check the project id in the tag matches the one in your dashboard. A wrong id serves an empty script rather than an error, so nothing appears and nothing complains.
3. Open the browser console. A content security policy on your site can block the script from loading; if so you will see it named there, and you will need to allow our origin as a script and frame source.
4. Check for an element sitting on top of it. The launcher pins itself to a corner at a very high stack order, but a cookie banner or sticky bar in the same corner can still cover it. Switching the launcher to the other corner is the quick test.

The bot keeps giving its fallback message

The fallback is returned for four distinct reasons, and your project’s analytics label every one of them. They look identical to a visitor but need opposite fixes, so read the label before changing anything.

• Below threshold — nothing in your content was close enough to the question, so no model call was made. Either the content genuinely is not there (write it, then re-crawl), or your similarity threshold is set too high.
• Model declined — the model was given real, relevant content and still refused. Usually the question needs a specific figure or name that your pages talk around but never state. Adding a pasted text source that says it plainly fixes this faster than anything else.
• Generation error — the model provider failed mid-answer. Nothing on your side to fix; if it persists, tell us.
• Blocked heuristic — the message looked like an attempt to talk the bot out of its instructions and was refused before it reached the model.

The unanswered-questions list in your analytics is the single most useful screen in the product. It is a ranked list of what your visitors wanted and your site did not say.

A page will not index

• It is too short. Under 20 words of extracted text and it is refused, because it would produce nothing retrievable. Contact pages, galleries and link hubs land here.
• The content is not in the HTML. Pages are re-fetched through a real browser when the first pass finds nothing, but content behind a login, a click, or a lazy-loaded tab will still be missed.
• The fetch failed. A redirect chain, a 403 from a bot filter, or a firewall rule. Check the URL opens in a private window with no session.
• You are out of credits. The source shows as over limit rather than failed. Free some allowance or upgrade, then retry indexing.
• It was skipped as unchanged. Re-crawling a page whose text has not changed is a no-op by design. If you expected a change, confirm it is live on the URL we fetch.

The chat returns a 403

Your allowed domains list does not include the origin the request came from. Add the exact hostname — and remember that example.com and www.example.com are two entries. Staging and preview domains need adding too, or the widget will work in production and fail everywhere you test it.

The answer cites the wrong page

Citations are the pages the answer was actually built from, not a guess after the fact. A surprising citation usually means two of your pages say similar things and the wrong one is more strongly worded. Narrowing or deleting the weaker source is the fix; competing near-duplicates are the most common cause of vague answers generally.

Still stuck

Email support@tekltalk.com with your project name, the page URL you are testing on, and the exact question you asked. Those three things resolve most reports in one reply.