AI Assistant for Website: How to Stop It Making Things Up

An AI assistant for your website can be trusted only if it answers from your own approved content and says so when it does not know. Invented answers come from missing sources and missing limits. This article shows how to set both, test them and keep them under control.

Why assistants invent answers in the first place

A language model is built to produce fluent text, not to check facts. When a visitor asks something the model has no source for, it still tends to write a confident reply. That is the invented answer you are worried about.

On a business site the damage is concrete. A made-up delivery time, a refund rule that does not exist or a wrong opening hour becomes a promise in your name. Visitors do not distinguish between the assistant and the company.

So the goal is not a smarter assistant. It is a more restricted one, which only speaks when it has a source and stays quiet otherwise.

Start with the questions your visitors actually ask

Before choosing any tool, collect real questions. Look at your contact form messages, support emails, phone notes and the search box on your site. Twenty to thirty genuine questions tell you more than any brainstorm.

Group them by topic: delivery, pricing logic, returns, appointments, technical details. Mark which ones have a clear written answer today. The ones without an answer are your first problem, and no assistant can fix that for you.

A common mistake is launching the assistant to cover everything. A narrow scope that works beats a broad one that guesses.

Ground every answer in content you approve

The reliable approach is to let the assistant search your own pages and documents first, then write a reply based only on what it found. This is often called retrieval: the assistant looks up passages before it answers.

Choose the source set deliberately. Current service pages, the FAQ, policy texts and product sheets belong in. Old campaign pages, drafts and outdated price lists stay out. If an old page remains in the source set, the assistant will happily quote it.

Assign one person to own the sources. When a policy changes, that person updates the page and the assistant follows. Without an owner, the assistant slowly drifts away from reality.

Write the rules it must follow

The instructions you give the assistant matter as much as the content. Tell it plainly to answer only from the supplied sources, to admit when the answer is not there and to never guess figures, dates or legal terms.

Define what happens at the edge. A good fallback sounds like this: it says it does not have that information and offers your contact details or a callback form. A bad fallback is silence or a vague, polite non-answer.

Also set the tone and the limits of its role. It should not give medical, legal or financial advice, and it should not promise outcomes. If your sector has advertising rules, ask your professional body what an assistant may say on your behalf.

• Answer only from approved sources.
• Say clearly when the answer is not available.
• Never state prices, dates or guarantees unless they appear in a source.
• Offer a human contact at every dead end.
• Stay inside the topics you defined.

Which setup fits your site?

There are three common routes, and the right one depends on how often your content changes and how costly a wrong answer is. A small brochure site has different needs from a shop with changing stock.

The comparison below is about risk of invented answers and effort, not about brand names.

Test it like a hostile visitor

Before launch, try to break it. Ask questions you know are not covered. Ask for a discount that does not exist. Ask it to confirm a wrong claim. A trustworthy assistant declines or redirects each time.

Run the same test with misspellings, short fragments and questions in a different language, since real visitors write that way. Note every answer that is wrong or too confident, then fix the source or the rule behind it.

Keep a small test list and repeat it after every content update. It takes minutes and catches regressions that nobody would notice by chance.

Keep a person in the loop after launch

Read the conversation logs weekly at first. Look for questions that got a weak answer or a fallback. Each one is either a missing page on your site or a rule that needs tightening.

Give visitors a visible way to reach a person, and make sure someone actually replies. An assistant that sends people to an unmonitored inbox creates more frustration than no assistant at all.

Also handle personal data with care. Tell visitors the chat is automated, avoid collecting more than you need and check your privacy notice covers the chat. For legal wording, consult your advisor.

Your next step: a small, grounded pilot

Pick one topic, such as delivery or appointments, and write down the thirty most common questions. Make sure each has an accurate, current answer on your site.

Then set up the assistant on that topic only, with the rules above and a clear route to a person. Test it with unanswerable questions and review the logs for two weeks.

Expand only when the pilot stays honest. An AI assistant for your website earns trust by what it refuses to say as much as by what it answers.