A Custom ChatGPT Chatbot for Your Website, Built Around Your Business

Three things people mean by “ChatGPT on my website”

When an owner asks us for a ChatGPT chatbot, the request usually points at one of three different products, and only one of them can actually live on a business website.

The first is ChatGPT itself, the consumer app at chatgpt.com. It is a destination rather than a component, and there is no supported way to place the ChatGPT app on your homepage and have it speak for your company. The second is a custom GPT, the configurable assistants people build inside ChatGPT and share by link or through the GPT store. Custom GPTs are handy for internal work, but they run inside ChatGPT: a visitor has to leave your site to use one, and the conversation stays on OpenAI’s platform instead of arriving in your inbox. The third is OpenAI’s developer API, which gives software access to the same family of models behind ChatGPT. A custom ChatGPT chatbot for a website is built on that third option.

Building on the API also settles who is in charge. Your business writes the instructions, decides what the model may read and what it may do, and keeps the records. That control is what separates a novelty from something your front desk can depend on.

Teaching a general-purpose model about one business

Straight out of the box, a GPT model knows a great deal about the world and nothing reliable about you. Ask it whether your company offers weekend appointments and it will produce a confident, plausible sentence with no connection to your real schedule. Everything a custom chatbot says about your business therefore has to come from material we supply, arranged so the model can find the right piece at the right moment. We narrow a general model in four layers.

Instructions

A written brief the model receives on every turn: whom it represents, whom it is talking to, what it is for, how it should sound, and a list of things it must never do. We write the brief in plain language and keep it under version control, so each change is dated and can be rolled back.

Approved knowledge

Your service descriptions, policies, hours, coverage area, FAQs and intake requirements, rewritten where needed so every fact appears once and matches what staff would say on the phone. For each question the chatbot retrieves only the relevant passages rather than memorizing the whole site, which is why it reflects a policy update as soon as the knowledge base changes.

Actions

Specific things the chatbot may trigger: send a lead to your email or CRM, offer a booking link, check a ZIP code against your coverage list, or flag a conversation for a person. Each action is a narrow, validated function on our server, never open access to your systems.

Boundaries

The topics it declines, the promises it cannot make, and exact wording for sensitive moments such as complaints, emergencies or a request for a quote it has no basis to give.

Fine-tuning, which retrains a model on your own examples, comes up in almost every consultation. For a business website it is rarely step one. It shapes style and format well but is a poor place to store facts that change, like pricing or staffing. Retrieval from approved knowledge handles facts; instructions handle tone. We reach for fine-tuning only when there is a measured problem it suits.

What the chatbot does while a visitor is on your page

Visitors never see the layers above. They see a chat window that behaves sensibly for the page they are reading. A walk-through for a hypothetical residential moving company shows how the pieces fit together:

1. An opening that fits the page. On the long-distance moving page, the greeting offers help with interstate moves instead of a bland “How can I help?”, because the widget sends the page address along with the first message.
2. A factual answer. The visitor asks whether the crew packs fragile items. The reply comes from the approved packing entry in two or three sentences, with a link to the page that covers it in full.
3. A question it must not answer alone. The visitor wants a price for moving a three-bedroom house from Houston to Denver. Its instructions forbid inventing estimates, so the chatbot explains that pricing depends on inventory and distance and offers to pass the details to an estimator.
4. Lead capture inside the conversation. It asks for the move date, origin and destination ZIP codes, home size and a phone number or email, one or two items at a time rather than as a wall of fields.
5. A clear next step. The details land in the sales inbox and CRM as a structured lead with a short summary, and the visitor is told who will follow up and how.

Notice what never happened: no guessed number, no request for information the company didn’t need, and no visitor left wondering what comes next. Those outcomes were designed in advance; they did not emerge on their own.

Stopping your chatbot from becoming free ChatGPT for strangers

A model capable enough to handle your customers is also capable of writing essays, fixing code and debating politics. Put it on a public page without limits and some visitors will use it for exactly that, on your usage bill and under your logo. Others will poke at it, typing “ignore your previous instructions” to see what spills out. We plan for both groups.

No guardrail is perfect, so transcript review is part of the plan. The first weeks after launch are when genuine questions start arriving, and reading them is how remaining gaps get closed.

Your data, OpenAI’s role, and who reads the transcripts

A chatbot on your site collects information from your customers, so its data path deserves as much attention as its answers.

• The API key never reaches the browser. Quick tutorials often call OpenAI straight from JavaScript on the page, which exposes the key to anyone who opens developer tools. Our chat window talks only to a server endpoint we control, and that endpoint holds the key.
• The account belongs to you. We set up the OpenAI API account under your company’s ownership and billing, or work inside one you already have. If you ever part ways with us, the account, key and usage history stay with you.
• Model training. OpenAI’s published API terms have stated that business data sent through the API is not used to train its models by default. Policies evolve, so we review the current terms with you before launch instead of quoting them from memory.
• Retention on our side. You decide how long transcripts are kept and who may read them. We configure the chatbot never to request card numbers, Social Security numbers or passwords, and to stop a visitor who starts typing them.
• Disclosure. The chat window states plainly that visitors are talking to an AI assistant, and your privacy notice is updated to describe what the chatbot collects and why.

When the conversation needs a person

A chatbot that insists on finishing every conversation itself will eventually mishandle an important one. We agree on handoff triggers before the build, usually some mix of these:

• The visitor asks for a human, in whatever words they choose.
• The topic falls in a category you reserve for staff, such as billing disputes or anything involving safety.
• No approved information exists for the question and the chatbot has already said so once.
• The lead is valuable by your own definition, for example a commercial account or a large project.
• The visitor’s tone turns sharply negative.

What your staff receives matters as much as when. Forwarding a raw forty-message transcript wastes the time the handoff was supposed to save. We put a brief summary on top: who the visitor is, what they want, what the chatbot already told them, and what is still open. The full transcript follows for reference, delivered wherever your team already works, whether that is a shared inbox, the CRM or a text alert to whoever is on call.

During business hours a handoff can open a live chat if someone is available to pick it up. After hours, the chatbot says honestly when a person will reply and gathers what that person will need, rather than implying someone is standing by.

Lead capture follows the same principle: ask for the minimum, at the point it becomes useful. Demanding an email address before offering any help drives people away. Someone who has just received a genuinely useful answer is usually far more willing to say how to reach them.

How EVOTECH builds it, from first call to live chatbot

1. Free consultation. On a no-cost phone or video call we review your site, where inquiries arrive today and what you want the chatbot to accomplish. Sometimes the honest answer is a better contact form, and we will tell you so.
2. Question inventory. You share a sample of real questions from email, calls and forms. We sort them into three piles: answer directly, collect details and route, or reserve for humans.
3. Written, fixed-scope quote. The quote lists the pages the chatbot appears on, the knowledge it draws from, the actions it can take, the integrations and the review period after launch.
4. Knowledge cleanup. We draft the approved knowledge base from your site and documents, flag contradictions (two different cancellation policies on two pages is common) and ask you to sign off.
5. Build and test. Instructions, retrieval, actions and the chat window are tested against your real questions plus a batch of deliberately awkward ones.
6. Staging review. Your team tries to trip it up on a private link before any visitor sees it.
7. Launch and early review. We read transcripts closely in the weeks after launch, fill gaps in the knowledge, tighten instructions and adjust handoff rules.

From the first call onward you deal with the same US-based team, reached by phone and video wherever your business operates. The company behind it, EVOTECH IT LLC, has operated for over 20 years and holds a 5.0 Google rating.

Plugin, custom GPT or custom build: an honest recommendation

Plenty of WordPress and Shopify plugins promise ChatGPT on your site, and several are perfectly decent. They tend to fit when your questions are simple, your content is already tidy, and leads only need to reach an email inbox.

They become harder to live with once you need any of the following: answers confined to approved content, with a real “I don’t know” instead of a guess; handoff rules that differ by service or time of day; leads written into a CRM with the right fields filled; control over where the key is stored and what gets logged; or reporting beyond a count of conversations. Many plugins ask you to paste your OpenAI key into a settings screen and run a generic prompt you can only partly edit.

A custom build takes more setup than installing a plugin, and if the plugin route is enough for you, we will say so. When it isn’t, everything we build belongs to you: the instructions, the knowledge base, the code and the conversation records. For agents that go beyond answering questions, see our AI agents page and our custom AI website agent page; for the engineering behind each model call, read about a GPT-powered website chatbot.

Related services

Frequently asked questions