AI Chatbots on Your Website: What Actually Works

Every visitor who can't find an answer does one of two things: they email you and wait, or they leave. Most leave. That's the entire case for putting AI chat on a website — not novelty, but closing the gap between a question and an answer.

It works. It also gets oversold, badly, and the numbers in most vendor pitches don't survive contact with production. Here's the honest version.

The friction problem it actually solves

Think about what a customer has to do to get a simple answer on a typical site. Scan the nav. Guess which page. Scroll. Give up. Find the contact form. Type the question. Wait a day. Get an answer that generates a follow-up question. Wait another day.

That's four steps and two days for something the business could have answered in one sentence. Every one of those steps loses people — and the ones you lose are disproportionately the ones who were closest to buying, because pre-purchase questions are buying signals.

A well-built assistant collapses that to: ask, get an answer, carry on. Under two seconds instead of under two days. That's the mechanism behind every other benefit on the list — faster answers, fewer dead ends, more people still on the page when they decide.

What the data actually says

Here's where you should be sceptical, including of us.

Vendors typically claim 40–60% deflection — the share of conversations a human never has to touch. But customer-measured results tell a different story: the median for AI self-service sits closer to 22%, with a range from 8% to 45%. Enterprise implementations with serious investment reach a median around 41% for tier-one queries, with the top quartile near 59%.

So the technology works — just at roughly half the rate the pitch deck implies, unless you invest like an enterprise.

More useful than any headline number is the fact that deflection varies enormously by question type:

Read that table before you buy anything. If your inbound is mostly "where is my order" and "do you ship to Alberta," a chatbot will carry a lot of it. If it's mostly complex troubleshooting or upset customers, expect far less and plan for humans.

One distinction worth holding onto: deflection is not resolution. Deflection counts conversations no human touched. Resolution counts problems actually solved. A conversation where the customer gave up and left also counts as deflected — which is why deflection alone is a dangerous metric to optimise.

On the commercial side, the widely-cited figures — roughly 23% higher conversion, up to 25% of abandoned carts recovered versus about 10% by email — come overwhelmingly from vendor-published studies. Directionally plausible, methodologically unverifiable. Treat them as a reason to test, not a forecast.

Where retention actually comes from

The retention argument is real but usually explained badly. It isn't that people love chatting with bots. It's that:

• Answers arrive while intent is still alive. A question answered in 5 seconds keeps someone in the session. The same answer tomorrow reaches someone who already bought elsewhere.
• It works outside business hours. For a business selling nationally across time zones, a meaningful share of traffic arrives when nobody's at a desk.
• It removes the "ask a human" tax. Plenty of customers won't email over a small question. They'll just guess, or leave. Chat lowers that threshold to near zero.

What we built for Danphe Stores

We built an assistant into the Danphe Stores app, a Nepali grocery business shipping across Canada. It sits on every screen — storefront, product pages, cart — so a customer wondering about an ingredient, a substitution, or where their order is can ask without abandoning what they're doing.

That last part is the design decision that matters most. The assistant is in the flow, not a detour. A shopper with a full cart and one hesitation gets it resolved without leaving the cart — which is precisely the moment where a lost answer becomes a lost order.

When you should not build one

We'd rather say this than sell you something that annoys your customers:

• Your content is thin or wrong. An assistant grounded in a bad knowledge base confidently produces bad answers. Fix the source material first — you'll get some of the benefit from that alone.
• Your volume is tiny. If you get five questions a week, answer them yourself. Warmly.
• Your questions are almost all complex or emotional. See the table. You'll frustrate people at the exact moment they need care.
• You want it to replace support entirely. It won't, and designing for that produces the trapped-in-a-loop experience everyone hates.

What separates a good one from an infuriating one

• Grounded in your actual content — products, policies, order data — not a general model guessing about your business. This is the single biggest quality factor.
• An obvious escape hatch to a human, visible from the first message. Hiding it is the fastest way to turn a support tool into a complaint generator.
• It says "I don't know." A model that admits a gap and hands off beats one that invents a shipping policy.
• It can see order state where relevant. "Where is my order" is only answerable if the assistant can look it up.
• Measured on resolution, not deflection — plus a satisfaction signal, so you can tell the difference between solved and abandoned.

The honest summary

An AI assistant is one of the highest-return additions to a content-heavy or transactional site — when your questions are routine, your content is solid, and you measure the right thing. It is not a support team replacement, and the deflection rate you'll be quoted is roughly double what you should plan around.

If you want one built on your actual content and wired to your real order data, that's the work we do. See how we approach AI-powered applications, or get in touch and we'll tell you honestly whether your question mix justifies it.