A plumbing company in Leeds bolted a chat widget onto its site in March. Six weeks later it had logged 412 conversations and booked 31 jobs it could trace back to the widget. It had also irritated about a dozen people who asked about emergency callouts and got a cheerful reply about office hours.
That split is the honest state of the ai chatbot for website market right now. The technology is genuinely good at a narrow set of jobs, mediocre at plenty of others, and almost always oversold by the companies selling it.
What the widget is actually doing
Strip away the branding and most website chatbots work the same way. A visitor types a question. The system searches your indexed content, meaning help docs, product pages, past support tickets, for the closest matches. A language model turns those matches into a sentence. The whole loop finishes in under two seconds.
Two details decide whether it feels clever or clueless. The first is what’s in the index. The second is what happens when the search comes back empty. A bot built on your real returns policy will quote it accurately. A bot built on your marketing homepage will invent something plausible and wrong, in confident prose, at 2am, to a customer who is about to spend £400.
Three jobs a website chatbot handles well
Answering the questions your FAQ buries
On a typical ecommerce site, somewhere between 40% and 60% of inbound chats are variations on four things: where is my order, what’s the returns window, do you ship to my country, and can I change my delivery address. Those answers already exist on your site, usually three clicks deep on a page nobody opens. A chatbot surfaces them in one sentence. That alone justifies the widget for a lot of small businesses.
Capturing leads at 11pm
A B2B software site I looked at recently got 28% of its demo requests between 8pm and 6am. Nobody was awake to answer. The chatbot’s job there isn’t to sell, it’s to ask three qualifying questions, capture an email, and tell the person when a human will follow up. Set that up properly and it pays for itself in a fortnight.
Steering people to the right page
“Do you do commercial contracts?” is a routing question, not a support ticket. A bot that replies with two sentences and a link to the relevant page stops a visitor from bouncing, and costs you nothing in staff time.
Where they fall apart
Most failures cluster in the same few places.
- Anything involving money or policy. Refunds, cancellations, discounts, warranty claims. If the model hasn’t got the exact rule in front of it, it will guess, and a guessed refund promise is a real cost.
- The second question. Bots handle the first query well because it’s almost always in the index. Ask a follow-up that requires remembering the first answer and things get shaky fast.
- The angry customer. Someone who has been charged twice does not want a friendly paragraph. They want a person, and they want them quickly.
- The confident wrong answer. Worse than a refusal, because it costs the customer time before they discover it.
Vendors quote containment rates of 70%, 80%, sometimes 90%. What they usually mean is that the bot replied to the message, not that it resolved the problem. A month of watching new AI tools that survived real use makes the pattern obvious: the ones that hold up are boring, narrow, and honest about their limits.
What you need to feed it
The knowledge base is the product. Everything else is packaging. Before you install anything, gather the twenty documents customers actually ask about, then write them in plain sentences rather than legal language. A bot can only be as clear as its source material.
Two settings deserve more attention than they usually get. First, the fallback: when confidence drops below a threshold, the bot should say “I don’t know, here’s a human” rather than improvise. Second, the escalation rules: trigger a handover on words like “cancel”, “refund”, “complaint”, and “speak to someone”, and don’t make the visitor ask twice. If you’re wiring the bot into other systems, the thinking behind architectural guardrails around AI agents applies directly here. Hard limits beat polite instructions every time.
Tone matters more than model choice. Someone at your company needs to read fifty transcripts a week for the first month and rewrite the system prompt when the bot starts sounding like a chatbot instead of a person. The breakdown of Claude AI tools worth using is a reasonable starting point if you’re weighing up which model to put behind the widget, though honestly the difference between the top three is smaller than the difference a good prompt makes.
What it costs
Three rough tiers, at the time of writing. A no-code widget from a mid-market vendor runs £30 to £80 a month with a message cap. That suits a site doing a few hundred chats monthly. Per-resolution pricing, where you pay per conversation the bot actually closes without a human, sits around £0.70 to £1.00 per resolution on the established platforms, which is fair value once you’re past a thousand chats. Custom builds start at a few thousand pounds and only make sense when the bot has to talk to your order management system.
Watch the message cap. Overage fees are where small businesses get stung, and a busy month can triple the bill. The broader question of which subscriptions earn their spot is worth revisiting regularly, and this AI tools list worth keeping is a decent gut check on what to pay for and what to drop.
The numbers to watch after launch
Ignore total conversation count. It tells you nothing. Track four things instead: containment rate (chats closed without a human), escalation rate, the percentage of chats where someone clicks “talk to a person”, and thumbs-down feedback as a share of total chats. A healthy bot on a support-heavy site lands somewhere around 50-65% containment with an escalation rate under 20%. If thumbs-down creeps above 8%, something in your knowledge base is wrong and the bot is guessing.
Also read the transcripts. Not all of them, but twenty a week. The phrases people use in the widget will tell you what your site’s navigation has been hiding for years.
A two-week test that tells you the truth
Before the widget goes live, sit down with a laptop and try to break it. Write out thirty questions and log every answer.
- Five questions whose answers are clearly on your site.
- Five questions whose answers are not on your site at all.
- Five with typos, slang, or a mix of two questions.
- Five about money: refunds, pricing, cancellation, discounts, warranties.
- Five hostile ones, phrased the way a furious customer would phrase them.
- Five in another language, if you sell abroad.
Score each answer as correct, wrong, or refused. Anything in the “wrong” bucket is a knowledge base gap or a prompt problem, and both are fixable in an afternoon. Anything in the refused bucket on a question you know the answer to means your indexing is broken.
Then run the same thirty questions past two colleagues who haven’t seen the bot before, without telling them what to expect. If they can trip it in ten minutes, so can a customer. Launch when the wrong bucket is empty, not when the demo looks impressive.