What “grounded” means
A grounded chatbot answers only from a fixed set of sources it has been given, instead of from what its AI model learned in general. Apex Bot's sources are the pages it reads from your website and any text you add yourself. If the answer isn't in those sources, Apex doesn't give one.
This matters because AI models are fluent even when they're wrong. A general chatbot asked about your return policy will often produce a plausible policy that you never wrote. On a business website that is worse than no answer: the visitor acts on it, and you are the one who has to explain.
From question to answer
- Your content is read and split. When you add a website, Apex reads the page you enter, your sitemap and the pages they link to on the same domain, up to your page limit. It respects
robots.txtand skips menus, footers and cookie banners. The text is split into short passages and stored for your assistant only. - The question is matched to passages. When a visitor asks something, Apex searches the passages of that one assistant and scores how relevant each one is. Content from other customers is never searched.
- Weak matches are dropped. Passages below a relevance floor are discarded, so a vaguely related page can't be stretched into an answer.
- The model writes from the passages. The AI model receives the visitor's question and the matching passages, with fixed rules first: answer only with facts found in the passages, never invent prices, policies or contact details, and say so when the passages don't cover the question.
- The source is shown. Each answer links the page it came from underneath, so the visitor can check it and keep reading on your site.
When it doesn't know
If no passage is relevant enough, the assistant replies with your “I don't know” message instead of guessing. If you turned on hand-off, it offers to pass the conversation to your team by email, with the whole chat attached.
Unanswered questions are useful: they show exactly what your visitors want to know that your website doesn't say yet. Add the missing information as a page or as a text source, and the next visitor gets an answer.
Answer confidence
How close a match has to be is a setting, under Persona → Answer confidence:
| Setting | Behavior | Good for |
|---|---|---|
| Strict | Needs a close match; says “I don't know” more often | Regulated topics, prices, legal terms |
| Balanced | The default | Most websites |
| Lenient | Accepts looser matches | Small knowledge bases with only a few pages |
Manipulation attempts
Visitors sometimes try to talk a chatbot out of its rules (“ignore your instructions and…”), and web pages can contain text written to do the same. Apex passes visitors' messages, your website's text and your custom instructions to the model as data inside clearly marked sections, after fixed safety rules that they can't override. Text that imitates those markers is neutralized first. Your own instructions can change tone and topics, but can't switch off grounding.
The assistant also has no tools: it can't place orders, change accounts, browse the web or call other systems. The worst a successful manipulation can do is change what one chat says. More on this on the security page.
Limits
Grounding greatly reduces made-up answers, but no AI system is perfect. An answer can still be incomplete, or misread a passage that is ambiguous. That's why every answer shows its source, and why your dashboard keeps the conversations so you can review them and improve your content. The assistant is honest about being an AI and won't give definitive medical, legal or financial advice.
Questions
Does Apex use what the AI model already knows about my business?
No. Answers are built only from your pages and text sources. If your content doesn't mention something, Apex doesn't say it, even when the model might know it from elsewhere.
What happens when my website changes?
Re-sync the website source under Dashboard → Knowledge and Apex reads it again. Text sources update as soon as you save them.
Which AI models does it use?
You choose: Claude, OpenAI or Gemini models, with an optional fallback model if your main one is unavailable. The grounding rules are the same on every model.
Is my content used to train AI models?
Not unless you opt in. Your content and visitors' messages are sent to the AI provider you choose only to generate answers. See the privacy policy.