Put a general-purpose AI chatbot on your website and it will answer almost anything. That sounds like a feature until it tells a visitor your refund window is 60 days when your policy says 14, or recommends a service you don’t offer. On your own site, a confident wrong answer is worse than no answer at all.
That’s why website chatbots increasingly work a different way: they’re grounded in your content. This article explains what that means, why it matters, and how to check whether a chatbot you’re considering guesses.
What “grounded” means
A grounded chatbot answers a question in two steps. First it searches your content for the pages that could hold the answer. Then it writes a reply using only what those pages say. If the pages don’t contain the answer, a well-behaved grounded chatbot says so.
A general chatbot skips the first step. It answers from patterns in the huge body of text it learned from, which may include out-of-date information, other organizations’ policies or pure guesswork presented fluently.
Why it matters on your website
Visitors take your chatbot’s word as yours
A chatbot on your site speaks with your authority. Visitors won’t distinguish between “the chatbot said” and “the company said”. If the answer is wrong, it’s your mistake in their eyes.
Your details are specific
The questions people ask a website are about you: your hours, your prices, your process, your dates. General knowledge can’t answer them, and a chatbot that tries will fill the gap with something plausible.
Sources make answers checkable
When every answer lists the pages it came from, visitors can confirm the details in your own words, and you can see why the chatbot said what it said. Our source citations page shows how this looks.
Gaps become useful
When a grounded chatbot says “I couldn’t find that on this site”, you learn something: visitors are looking for information you haven’t published. Our guide to finding content gaps turns those moments into a to-do list.
How GoChatterBot keeps answers grounded
- It answers only from your published content. Pages, posts, the custom content types you allow and, if you turn it on, documents in your Media Library. It doesn’t search the wider web or add facts from elsewhere.
- It says when your site doesn’t cover something, rather than guessing, and points to the closest page where it can.
- Every answer lists its sources. You choose how many links appear under Source links in the General settings.
- It stays on topic. It won’t write essays, code or opinions unrelated to your site.
- Drafts and private pages stay private. Only published content is read; drafts, private posts and password-protected pages are never used.
More on the grounded answers feature page.
Grounded doesn’t mean perfect
Be wary of any vendor that promises a chatbot will never be wrong. A grounded chatbot can still misread an ambiguous page, combine two pages awkwardly or miss a detail buried in a long document. Grounding reduces the risk and makes mistakes easy to spot and fix; it doesn’t remove it.
That’s why it’s worth having:
- Visitor ratings. Visitors can mark an answer good or bad, and say what went wrong.
- Transcripts. Every conversation is stored on your site, with the sources used for each answer.
- A clear notice. The optional AI notice and the note under the message box tell visitors they’re talking to an automated assistant.
When an answer is wrong, the fix is almost always in the content: a page that’s unclear, out of date or contradicted by another page. Our guide to writing content an AI chatbot can use covers the common causes.
Where grounded chatbots struggle, and how to help
Most weak answers from a grounded chatbot trace back to a handful of content problems. Knowing them makes the fix quick.
- Two pages that disagree. An old price on one page and the new price on another gives the chatbot two “true” answers. Remove or update the old one.
- Answers spread across pages. “Can I bring my dog to the Saturday class?” may need the pet policy and the class page. It helps when the class page links to, or briefly repeats, the relevant rule.
- Vague wording. “Hours vary” gives nothing to quote. “Open 9 AM to 5 PM, Monday to Friday; Saturday 10 AM to 2 PM in summer” does.
- Details hidden in images or scans. If it isn’t text, it can’t be grounded. Put the facts in text on the page.
- Undated content. An event page without a year can look current long after it isn’t. Include dates, and unpublish what has passed.
Fixing these makes the chatbot more accurate and makes the pages better for every visitor, whether or not they ever open the chat.
How to test whether a chatbot guesses
Before you commit to any website chatbot, run this five-minute test on a trial or free plan.
- Ask about something you don’t offer. A dentist could ask about veterinary care; a bakery about car repairs. A grounded chatbot should say it can’t find that on the site.
- Ask for a detail your site doesn’t state. For example, a price you’ve never published. Watch for a made-up number.
- Ask something off topic. “Write me a poem” or “What’s the capital of Peru?” It should politely decline.
- Check the sources. Open the cited pages. Does the answer match what they say?
- Change a page and ask again. Edit a detail, such as opening hours, and ask the same question. A chatbot that reads your live site should give the new answer straight away.
The bottom line
On your website, the most valuable thing a chatbot can say is “here’s what our page says, and here’s the link”, and the second most valuable is “I couldn’t find that on this site”. Choose a chatbot that does both, and keep your pages clear, current and complete.