Picture a temp covering your front desk for a week. You hand them a binder with your prices, policies and opening hours, and one instruction: if it is not in the binder, do not make it up, take a message. A good temp follows that rule. A bad one guesses, sounds confident and books a job at last year's price.
Grounded AI answers are the software version of the good temp. The assistant looks things up in the content you gave it, answers from what it finds and admits when the binder is silent. That one design choice is most of the difference between an assistant you can leave running overnight and one you have to babysit.
It is not magic, and it does not fix bad source material. But once you see how it works, it becomes obvious which parts of accuracy sit with the software and which sit with you.
What an ungrounded assistant does
General-purpose chat tools are trained on enormous amounts of text. Ask one about a boiler service and it will produce a fluent, reasonable-sounding reply based on boiler services in general. It has no idea what you charge, which towns you cover or that you stopped doing Saturday visits in March.
The dangerous part is not that it lacks the facts. It is that it fills the gap with plausible ones. People call this hallucination, which makes it sound rare and strange. In practice it is very ordinary: a price that is roughly typical for the industry, a guarantee period that sounds normal, a refund policy that most businesses have. It reads well. It is just not yours.
How one question gets answered
Here is what happens, in plain steps, when a visitor asks a grounded assistant a question. The technical name for this pattern is retrieval augmented generation, and the glossary is the place to look up terms like it if you need them for a meeting.
- The visitor types: "Do you charge extra for weekend call-outs?"
- The assistant searches your content: crawled pages, uploaded documents, Q&A pairs and your business description.
- It pulls out the passages most likely to be relevant, for example a line on your pricing page about weekend rates.
- It writes a reply using those passages, in the tone and persona you set.
- If the search turns up nothing relevant, it does not improvise. It tells the visitor it does not have that information and offers to take their details.
- That unanswered question is recorded as a knowledge gap so you can see it later.
Step five is the one that matters most. The rest is what makes answers sound natural. Step five is what keeps them honest.
The difference in a real exchange
Here is the same question answered by an assistant whose content does not mention weekend rates.
Visitor: Do you charge extra for weekend call-outs? Assistant: I do not have details on weekend pricing, so I would rather not guess. If you leave your number and postcode, the team can confirm the price before anything is booked.
Compare that with an ungrounded reply, which might say weekend call-outs "typically carry a 50% surcharge". That figure came from nowhere in your business. The visitor now expects it, and your team has to either honour it or have an awkward conversation.
The grounded reply is less impressive in the moment. It is the one you want.
What grounding does not fix
Grounding is not a guarantee of accuracy, and anyone telling you otherwise is overselling. It shifts the source of errors from the software's imagination to your own content, which is better, because your content is something you can read and correct.
The common failure modes:
- Stale pages. The assistant faithfully repeats the price list from 2024 that is still on a forgotten page.
- Contradictions. Your FAQ says 30-day returns and your terms say 14. The assistant may pick either.
- Vague source material. "Prices on request" gives it nothing to work with, so it will decline often.
- Near misses. A question about "gutter cleaning" retrieves your "gutter repair" page and the reply blurs the two.
Every one of those is a content problem, which is good news. It means the fix is editing a page or adding a Q&A pair, not hoping a vendor's model improves.
What the owner controls
You have more control over accuracy than most people realise. The levers are simple:
- What goes in. Crawl the pages that are current, and exclude archived posts, old promotions and anything you would not want quoted. Our guide to keeping your assistant's knowledge up to date has a routine for this.
- Which source wins. If a document and a web page disagree, fix one of them. Do not leave the assistant to arbitrate.
- Direct answers for sensitive topics. Write Q&A pairs for prices, guarantees, cancellations and anything with legal weight, so the wording is exactly yours.
- What the description forbids. State plainly what the business does not do and what the assistant must not promise.
- The gap list. Unanswered questions show you what visitors want that your content lacks. Turning those into pages is covered in unanswered questions into website pages.
A quick way to test it yourself
Ask your assistant ten questions. Five should be things clearly covered by your content. Five should be reasonable questions your content does not answer, such as a service you do not offer or a policy you have never written down.
For the first five, check each answer against the source page. For the second five, the right result is a clear "I don't have that" plus an offer to take details. If any of those five produce a confident, specific answer, find out where it came from before you go live. You can run this kind of test on a sample assistant at /try before building your own.
Where to start
Pick the three topics where a wrong answer would cost you most, usually prices, availability and cancellations. Read your own content on each and fix anything outdated or contradictory. Then write a Q&A pair for each so the assistant has one clear source.
After that, check the unanswered questions list every week for a month. If the assistant keeps saying it does not know something, the fix is in your content, and our guide on what to do when it says I don't know walks through it.