AI chatbots for websites in 2026: what to build, what to avoid

What visitors actually want from a website chatbot

Five9's 2026 Consumer CX Report, published on 14 October 2026 and based on a survey of 3,000 consumers in the US, UK and Germany, found that trust in AI more than doubles when a human option is visible: from 26% when no person can be reached to 55% when one can. 71% said it is very or extremely important to know when they are talking to an AI agent, and 92% give AI less than 10 minutes before asking for a person. The most common reason for giving up was that the AI repeatedly misunderstood them (46%), and about a third said they had to repeat themselves after being handed over.

Five9 chief executive Amit Mathradas put it this way: “consumers aren't asking companies to choose between AI and people; they're asking for both, done right.” Gartner's survey of 3,566 customers, published in August, points the same way: 87% said companies using generative AI for service must offer a way to reach a human agent, and Gartner senior director analyst Eric Keller advises that “Service leaders should not use GenAI as a mandatory first step for every issue.”

What to build: a chatbot that answers from your own content

The best website chatbots in 2026 are narrow. They answer from your own pages, FAQs and documents, they link to the page the answer came from, and they say so when the answer is not there. This is often called grounding, or retrieval-augmented generation. Deskee and GoAdserver both make it a headline feature, and GoAdserver states that its chat “does not invent prices, features or customers” and hands questions it cannot answer to your team.

Start with the questions you already get every week: opening hours, delivery areas, service areas, what a project costs, how booking works. Write clear answers to them on your website first. A chatbot can only be as accurate as the pages it reads, and ConvoAI's own launch guide from 11 October gives the right advice for testing: where an answer is wrong, “fix the information rather than the wording”.

What to build: a visible, fast handoff to a person

The handoff is the feature most small sites get wrong. A good chatbot shows a clear way to reach a person from the start, decides which topics always go to a human (complaints, refunds, billing, anything legal or medical), and passes the whole conversation along so the visitor does not repeat themselves. If nobody is online, it should collect a name, contact details and the question, and promise a realistic reply time.

This is becoming law as well as good practice. In September 2026 California's governor signed AB 1609, which requires businesses with more than $500 million in annual revenue to make a good-faith effort to connect customers to a human within 15 minutes of a request, starting next year. A San Francisco Standard investigation on 15 October described customers stuck in circular chats with an AI support agent over billing errors. Assemblymember Rick Zbur, who wrote the law, told the paper that if the AI cannot resolve the issue, companies “would actually have to transfer you to a human being that would continue providing service through the chat function”. Most small businesses are far below that revenue line, but their visitors have the same expectation.

What to build: tell visitors it is AI

Under Article 50 of the EU AI Act, which applies from 2 August 2026, AI systems that interact directly with people must be designed so those people are informed they are interacting with an AI system, unless that is obvious. The information must be given clearly at the latest at the first interaction. For a website chatbot, the simple answer is a short label in the chat window, such as "AI assistant", and a first message that says what it can and cannot help with. We will cover the wider EU AI Act transparency rules for websites in a separate post later this month.

Disclosure is not just about compliance. Five9's finding that 71% of people want to know when they are talking to AI suggests that hiding it costs trust you then have to win back.

What to avoid: letting the bot make promises

A chatbot on your website speaks for your business. On 12 May 2026 the Higher Regional Court of Hamm in Germany ruled against a cosmetic medicine company whose website chatbot told visitors that its two doctors were specialists in plastic and aesthetic surgery, a qualification they did not hold. The court held that the chatbot's false statements were the company's own commercial statements, rejected the argument that the chatbot was an independent third party, and said that even training it only on correct data would not change that. The consumer group that brought the case notes that the question was foreseeable, so the company should have accounted for it when setting the bot up.

The lesson for any business is to keep the chatbot away from claims it cannot back up. Do not let it quote prices, discounts, refund terms, delivery dates or qualifications unless they come word for word from a page you control. Recent public complaints, such as a delivery customer who said a support bot promised a full refund and then refused one, show how fast a wrong promise turns into a dispute.

What to avoid: actions without limits and costs without a cap

Many 2026 chatbots can do things as well as talk: book appointments, change orders, cancel subscriptions. Every action is a new way to get something wrong. Start with read-only answers and lead capture, add one action at a time, require confirmation before anything that costs the customer money, and log every action. Chatbots that read pages and accept free text also need protection against hidden instructions, a risk we will cover in our post on prompt injection later this month.

Watch the pricing model too. Tidio's October 2026 comparison of ten AI support agents lists Intercom Fin at $0.99 per outcome, HubSpot's Customer Agent at $0.50 per resolution in credits, Zendesk at $1.50 to $2.00 per resolution beyond the plan allowance, and Tidio's own Lyro from $32.50 a month for 50 AI conversations. Per-resolution pricing is fair for busy support teams, but a small business site usually needs a predictable monthly plan with a hard limit on usage.

An eight-point checklist for a website AI chatbot

• Write clear answers to your top 20 visitor questions on your website before you add a chatbot.
• Ground the chatbot only in your own pages and documents, and have it link to the source of each answer.
• Make it say clearly when it does not know, and offer to take the visitor's details instead of guessing.
• Label it as an AI assistant in the chat window and in its first message.
• Show a way to reach a person from the start, and pass the full conversation to staff on handoff.
• Block promises on prices, refunds, delivery dates, qualifications and legal or medical topics unless they quote a page you control.
• Add actions such as booking or cancelling one at a time, with confirmation and logging, and set a monthly usage cap.
• Read the transcripts every week, fix the website content behind wrong answers, and test awkward questions before each change.

Where CoreCodify fits

A website chatbot is now cheap to switch on and expensive to get wrong. The ones that work in 2026 are narrow, honest about being AI, grounded in content you control and quick to hand a visitor to a person.

At CoreCodify we design and build websites, apps and AI features, from the content structure a chatbot reads to custom assistants connected to your booking, CRM or store. If you want a website chatbot that answers accurately and knows when to step aside, hire me for custom work at usamify.com.