By Nazreen Daud · Published 19 Oct 2026 · Updated 19 Oct 2026 · Originally shared on LinkedIn
The short answer: a website AI assistant gives better answers when you feed it fewer, current pages, not your whole website. Start with your services, prices, FAQs and how to book. Leave out old posts and anything you no longer offer. Then test it with ten real client questions, and ask them again every time your website changes.
The day a client's website told the machines the wrong name
Earlier this month, we ran a routine health check on a client's website.
To a human visitor, the site looked lovely. To a machine, it told a slightly different story.
- The short plain-text summary written specifically for AI tools (an
llms.txtfile) gave the business a name that actually belongs to a completely different organisation. - A seasonal campaign page from months earlier was still live, still asking people to act now.
- An SEO tool installed months earlier was quietly injecting a hidden block of 31 FAQs into the home page. Visitors couldn't see it. Machines could. Some of it was advice the client had never reviewed.
Nobody had been careless. Campaigns end, tools get forgotten, and nobody re-reads what only machines read.
Now picture a website AI assistant trained on all of that, answering a customer's question at 11pm. It wouldn't hesitate. It would be confidently wrong, in the client's own name.
We removed the hidden tool and the wrong name the same day. It's how I now explain AI assistants to clients: the assistant is only as good as what you let it read.
Why does feeding an AI assistant more pages make it worse?
I read a discussion on Reddit this week that put it perfectly. Someone loaded an entire website, 3,000 pages, into an AI assistant's knowledge base. Its answers got worse than when it only had 40 pages.
Why? Because 3,000 pages of website is not 3,000 pages of content. In the original poster's words, it was "nav and legal and 400 blog posts from 2019". A tight filter took those 3,000 pages down to roughly 250 worth keeping.
The replies were even better:
- One said the 2019 posts are the worst part, because they're "confidently wrong" about what the business does now.
- Another described an internal assistant that kept quoting the cookie consent banner to answer billing questions. (I laughed, then I winced.)
- A third said most assistants don't need a huge crawl at all. "They need about three good pages."
Think of it as briefing a new receptionist
Imagine your new front-desk hire starts on Monday. You wouldn't hand them the keys to the storeroom, point at nine years of filing and say, "Everything you need is in there."
You'd give them one tidy sheet: what we offer, what it costs, how people book, and what to say about refunds. Plus: "If you're not sure, ask me."
An AI assistant deserves the same courtesy. The difference is that it never says "I'm not sure". If an old page says your package costs less than it does today, it will quote that price with a smile.
Does it really matter if the assistant gets something wrong?
Yes. Here's a real case.
In 2024, a Canadian tribunal ordered Air Canada to compensate a customer after the airline's website chatbot said a bereavement fare could be claimed retroactively, after booking. The airline's own policy page said otherwise. Air Canada argued it couldn't be held responsible for what the chatbot said. The tribunal disagreed, and the airline was ordered to pay about C$812 (CBC News).
The line that stuck with me: it makes no difference whether the information comes from a static page or a chatbot. Your customer sees one business, not two.
For a small Singapore business, the cost is usually quieter: a wrong price, a lost booking, or a review that starts with "Your chatbot told me…"
What should you feed your website AI assistant?
My rule of thumb: feed it what you'd hand a new staff member on day one, and nothing you'd be embarrassed to see quoted back.
Start with your current pages
- Services, and what each one includes
- Prices, or honest price ranges
- FAQs, especially the ones about money and time
- How to book, and what happens after
- Opening hours, locations and how to reach a human
- Your cancellation, refund and rescheduling policy
If you've already written a one-page list of facts you must never get wrong (step two in my three steps for AI adoption), that list is gold here.
Leave out the noise
- Old blog posts, past campaigns and expired promotions
- Services you no longer offer
- Menus, footers and cookie banners
- Anything hidden from human visitors. If people can't see it, your assistant shouldn't be quoting it. (Google applies a similar rule to structured data: don't mark up content that isn't visible to readers.)
Keep a list of what you left out
A fair counterpoint from the same thread: cut too much and you can drop the one page that matters. Legal pages are noise for most questions, but not for "Can I cancel today?". One developer admitted an over-eager exclusion took them a month to spot. Their fix: filter hard, and log everything you exclude.
A hypothetical to make it concrete. Picture a physio clinic in Tampines that switches on an assistant trained on its whole site. A 2021 post still mentions an old introductory package. A parent asks, "How much for my son's first session?" The assistant quotes the 2021 package. Now the front desk has to explain a price the clinic never meant to offer. Leaving that post out would have prevented the whole conversation.
How do you test a website AI assistant before it goes live?
Write ten real questions your clients actually ask, in their words, and mix them up:
| Type of question | Example | What a good answer does |
|---|---|---|
| Common | "How much is a first consultation?" | Quotes today's price or range |
| Awkward | "Can I get a refund if I cancel?" | Uses your actual policy |
| Out of hours | "Can I book for Saturday morning?" | Gives the real booking route |
| Out of scope | "Is this pain serious?" | Hands over to a human |
| Local phrasing | "Got weekend slot or not?" | Understands it and answers properly |
| Stale trap | A service you dropped last year | Says you don't offer it |
Then comes the step that does the most work: ask those ten again every time you change your website or the assistant. In my experience, problems often come from a website change nobody noticed, and a small set of test questions is what catches them.
It's the same principle behind making sure someone actually owns your automations: tools don't stay right on their own.
Not sure what your website is quietly telling the machines? Book a call with me and we'll look at it together.
My view: if it doesn't fit on one page, it isn't ready
After building websites since 2013, and now running more than 11 AI agents alongside my own lean team, here's my honest opinion:
A website AI assistant isn't a search engine for your website. It's a new member of staff with a perfect memory and no sense of what's out of date.
So I brief mine like staff. The agent that helps draft my LinkedIn posts reads a single brand file before it writes anything. For my free Offer Generator, it gets a short list of verified features with a blunt instruction: don't add any beyond these. Short, current, checked. The less it has to guess, the less it gets wrong.
If you can't list what your assistant needs on one sheet of A4, you're not ready to switch it on. It's the cheapest preparation you'll ever do, and it echoes more is probably the problem: sometimes the upgrade is removing things.
Clear, current pages also help AI search describe you properly, as I covered in AI search visibility for small businesses.
A 30-minute checklist before you switch one on
- Start with your current pages: services, prices, FAQs and how to book.
- Leave out old posts and anything you no longer offer.
- Keep a list of what you left out, in case a question needs it later.
- Write ten real questions clients ask, including the awkward ones like refunds and cancellations.
- Ask those ten again every time you change your website or the assistant.
Want a second pair of eyes before you switch one on? Book a call with me.
FAQ
Should I connect my whole website to an AI assistant?
Usually not. Most websites are full of menus, legal text, old posts and past campaigns. Start with your current service, pricing, FAQ and booking pages, and add more only when a real question needs it.
How many pages does a website AI assistant need?
There's no magic number. Most small service businesses need far fewer than they think: the pages that explain what you offer, what it costs, how to book and your policies. If it fits on one page of notes, you're in good shape.
How do I test a website AI assistant?
Write ten real client questions, including awkward ones about refunds, cancellations and out-of-scope requests. Ask them before launch, and again after every website or assistant change. Check each answer against your current policies.
What if my AI assistant gives a customer the wrong answer?
Treat it as your business giving the wrong answer, because to the customer it is. Correct it with the customer, fix or remove the page it learned from, and add that question to your test list.
Do old blog posts hurt my AI assistant?
They can. Old posts often describe prices, packages or processes you've since changed. An assistant can't tell they're out of date, so it may repeat them confidently. Leave them out, or update them first.
If you added an AI assistant to your website tomorrow, what's the first question you'd test it with?
About the author: Nazreen Daud is a Malay web consultant and marketer in Singapore. He has built websites for businesses since he started freelancing in 2013, registered Roquepress in 2019 and founded Mulacan. Today he runs a lean team alongside more than 11 AI agents, and writes practical, hype-free notes on websites, SEO and AI for service businesses.