Type a question into most AI chatbot websites and you get a confident, well-formatted answer. Ask a follow-up, though. A real follow-up, with a pronoun and a missing noun. The thing falls apart. It forgets what you asked, parrots the homepage copy back at you, or invents a feature that doesn’t exist. Four seconds of magic, then twenty seconds of backtracking.
That gap between the demo and the product is where most of the money gets lost. There are thousands of AI chatbot websites now, from $19-a-month plugins bolted onto WordPress to enterprise platforms billing nearly a dollar per conversation. In a marketing video they all look identical. Once a real customer types “it’s still not working,” they behave very differently.
What separates a working chatbot from a demo
Three things, in my experience: retrieval quality, scope discipline, and what happens at the handoff to a human.
Retrieval is the boring part nobody markets. A model is only as good as the chunks you feed it. If your help centre has nine articles about refunds that contradict each other, the bot will blend all nine into one confidently wrong policy. The teams that get this right spend more time on content hygiene than on prompt engineering, and they re-index whenever prices or policies change.
Scope discipline means the bot knows what it can’t answer. A support bot that says “I don’t have that, here’s a human” beats one that guesses every time. Hallucinated answers cost far more than a slightly higher escalation rate.
The handoff matters most and gets tested least. A conversation that dies mid-sentence because the bot couldn’t route to an agent is worse than no chatbot at all.
The four kinds of AI chatbot websites you’ll run into
Support deflection bots
These sit on a help centre or product page and try to answer tier-one questions: where’s my order, how do I reset my password, why was I charged twice. Companies like LivePerson have pushed this category toward routing conversations by intent rather than by menu tree, and the difference shows in how LivePerson AI is making customer conversations smarter. The good ones hand off with full transcript context so the customer never repeats themselves.
Document and knowledge bots
Upload the PDFs, get a chatbot. Legal teams, HR departments and SaaS vendors love these. They live or die on citations. If the bot can’t point to the exact clause it’s quoting, nobody trusts it. Expect per-seat pricing and a lot of tedious work cleaning up tables that got mangled during ingestion.
Sales and lead-qualification bots
B2B sites use these to ask three qualifying questions before booking a call. The failure mode is aggressive: a bot that pops up after eight seconds and demands an email address gets closed. Used well, they’re closer to a smart form than to a conversation.
Personal assistant bots
Different category, same label. These run on your own accounts and do chores. I spent a week with one and found Google’s Gemini Spark genuinely useful for scheduling and email triage, mostly because it worked in the background instead of waiting to be prompted.
What these things actually cost
The sticker price is misleading. Most AI chatbot websites now bill on outcomes or tokens, and both scale faster than buyers expect.
Intercom’s Fin charges roughly $0.99 per resolution. That sounds cheap until you’re resolving 20,000 tickets a month, at which point you’re staring at a five-figure line item that has to be justified against the agents it replaced. Token-based platforms flip the problem: a long conversation with a large document attached can cost more than a short one, so a chatty bot is an expensive bot. Then there’s the seat fee for the dashboard, an onboarding fee, and usually an extra charge per language.
Model choice moves the number more than vendors admit. Running a small model for intent classification and reserving the big one for final answers can cut inference costs by 60% or more with no noticeable drop in quality.
Build or buy?
There’s no universal answer, but the decision usually comes down to where your knowledge lives and how strange your workflows are.
- Buy a platform if your content sits in a standard help centre, you want something live this week, and you have no engineers to spare.
- Build in-house if the bot needs to touch internal systems like order databases, ticketing or billing, or if per-conversation pricing will exceed a developer’s salary inside a year.
- Go hybrid if you’re like most teams: a vendor for retrieval and the widget, your own code for the actions the bot triggers.
- Don’t build if you can’t commit someone to re-indexing content every month. An abandoned chatbot is a liability with a chat bubble on it.
The questions nobody asks until something breaks
Two risks deserve attention before launch, not after.
The first is access. A chatbot wired into email, calendars or internal drives is a privileged account with a friendly face. Earlier this year Microsoft disrupted an AI-assisted platform that had compromised around 12,000 accounts, and the pattern is familiar: broad OAuth scopes granted during a trial that nobody ever revokes. Audit permissions before you connect anything that matters.
The second is content leakage. Publishers have been arguing with AI vendors about how much text a chatbot reproduces, and the figures are contested on both sides. Microsoft, for one, claimed almost nobody was pulling NYT articles through its chatbot. If you’re building on licensed content, log what yours returns.
There’s a third, quieter problem: the web is filling with bot-generated pages designed to game search. Communities are already pushing back, as you can see in the argument over whether Reddit can fend off a new wave of AI SEO spam. If you run a content site, a chatbot with publish rights will damage your domain faster than it saves you time.
How to test a chatbot website in ten minutes
Skip the demo. Open the widget on the vendor’s own site, or ask for a sandbox with your documents loaded, then run these:
- Ask a question whose answer appears in two places with slightly different wording. Watch which one it picks.
- Ask something wildly out of scope, like the weather in Lisbon, and see whether it declines or invents.
- Ask a follow-up that relies on a pronoun. This breaks more bots than any other test.
- Request a human, then check whether the transcript follows you.
- Throw in a typo and a mixed-language sentence. Real customers write badly.
Fail two or more and keep shopping. Vendors rarely fix retrieval problems after the contract is signed.
Where this is heading over the next year
The interesting shift isn’t smarter conversation. It’s action. Chatbots are turning into agents that issue refunds, reschedule deliveries and file tickets without asking permission for every step. That moves the hard problems from language to permissions, audit logs and rollback. Expect “what can it do without me?” to replace “how well does it answer?” as the buying question.
For now, the practical advice stays unglamorous. Fix your documentation. Narrow the bot’s scope. Log every conversation for the first month and read fifty of them yourself. The teams who do that end up with a chatbot that quietly works while everyone else is still shipping a demo that falls apart on the second question.