Can You Use ChatGPT for Customer Service on Your Website?

If you already use ChatGPT every day, the idea writes itself. It drafts your emails and explains your own product back to you, so why not let it answer the customers who ask the same ten questions every week? Plenty of small businesses try exactly that: they build a custom GPT, upload the FAQ and the returns policy, and then look for the button that puts it on the website. There isn't one.

This guide is for the owner or support lead in that position. It explains what ChatGPT can and cannot do for customer service in 2026, why a custom GPT is the wrong shape for a support channel, what building your own assistant on the OpenAI API involves, and what a purpose-built AI support agent adds on top of the model.

The OpenAI details below come from OpenAI's own help center and API documentation, read at the end of September and the start of October 2026. OpenAI changes its products often, so check the linked pages before you plan around a date.

Can you put ChatGPT on your website?

No, not ChatGPT itself and not a custom GPT. OpenAI's help center article on GPTs in ChatGPT answers the question directly: GPTs are designed to work in ChatGPT, they are not a way to embed ChatGPT in an external website or application, and anyone who wants an assistant inside their own product should use the API.

That leaves three real routes for answering customers, plus one that helps your team rather than your customers:

1. Link out to a custom GPT. You build a GPT, share its link from your site, and visitors chat with it on chatgpt.com. It is the quickest route and, as the next sections show, the weakest support channel.
2. Build on the OpenAI API. A developer uses OpenAI's models through the API to build a chat window, a search index over your content and everything around them. You get full control and a software project to maintain.
3. Use an AI support agent. An AI support agent is a hosted product that runs a large language model over your own knowledge base, sits on your website as a chat widget, takes actions such as order lookups, and hands conversations to your team when it cannot help. Some, including PepoChat, are built on OpenAI's models, so the writing feels like ChatGPT while the plumbing a support channel needs is already in place. Our explainer on what an AI support agent is compares it with a classic chatbot and a copilot.
4. Use ChatGPT behind the scenes. Your team keeps using ChatGPT to draft replies and articles, while customers talk to whatever sits on the website.

Routes 2, 3 and 4 combine well. Route 1 mostly doesn't, for reasons that have little to do with how good the answers are.

What is ChatGPT good for in customer service?

ChatGPT is a strong writing and analysis assistant for the people who do support, even when it never talks to a customer. The safest and most useful jobs are the ones where a person reads the output before anyone else does.

The last column follows one rule: ChatGPT writes fluently whether or not it knows the facts, so the facts have to come from you. Our guide on why chatbots hallucinate explains the mechanism. For internal drafting, the fix is simply a person who checks.

Some of this work may not need ChatGPT at all. Some support inboxes include a rewrite button; PepoChat's team inbox has one called Enhance, which fixes grammar and clarity in a draft reply while keeping your facts and tone. The difference is that the draft never leaves the inbox.

The pattern-spotting row deserves a closer look, because it is where support teams get the most out of ChatGPT and where privacy goes wrong most often. An inbox export usually contains names, email addresses and full transcripts. PepoChat's CSV export, for example, includes visitor names, emails and transcripts for the date range you choose. Strip the personal columns before you paste anything into a chat window.

Be careful what you paste into ChatGPT

Whether your conversations can be used to train OpenAI's models depends on your plan. OpenAI's help center says that on Business, Enterprise and Edu plans, data is not used for training by default, while on consumer plans (Free, Go, Plus and Pro) it may be, depending on whether you have opted out. Data sent through the API is a separate case: OpenAI's data controls documentation says that since March 1, 2023, API data is not used to train models unless you opt in, and that abuse-monitoring logs are kept for up to 30 days by default.

The practical rule for a small team is simple. Do not paste customers' names, emails, order numbers or payment details into a personal ChatGPT account unless you have checked the data setting and your own privacy notice covers it. If ChatGPT is part of your daily support work, a business plan is the cleaner choice, and our GDPR guide for chatbots covers how to list AI providers in your privacy notice.

Can a custom GPT be your customer service chatbot?

Not a good one. A custom GPT is a version of ChatGPT configured with your own instructions, uploaded knowledge files and optional tools, and OpenAI's own documentation describes several properties that rule it out as a support channel for most businesses.

Your customers need a ChatGPT account. OpenAI's help center says GPTs are available to all ChatGPT users, but people must be signed in to start a conversation. Public GPT pages may be visible without signing in, and visitors are prompted to sign in before chatting. A shopper who wants to know whether you ship to Canada will not create an account with another company to find out.

You cannot read the conversations. The same article states that GPT builders cannot view the individual conversations people have with their GPTs. A support channel whose questions you never see cannot be improved, cannot be checked when a customer complains about an answer, and cannot be picked up by a colleague.

There is no human handoff. A GPT has no team inbox, no escalation status and no way to alert you when someone needs a person. The conversation ends wherever the model's answer ends. Our explainer on human handoff covers why this is the most important feature of any customer-facing bot.

It forgets everyone. GPTs do not use saved memory or previous conversations; each conversation starts fresh. That is fine for a writing tool and awkward for a returning customer who expects you to remember last week.

Live data means building Actions. A GPT can reach outside services through Actions, which call a REST API you describe with a schema and an authentication method: none, an API key or OAuth. That is real development work: connecting a GPT to your order system is a project, not a setting.

You can no longer start one on a personal plan, and GPTs are being retired. OpenAI's article on creating and editing GPTs, updated at the end of September 2026, says new GPT creation and publishing are not available on personal accounts, including Free, Go, Plus and Pro. Business, Enterprise and Edu workspaces can still create GPTs where their settings allow. The same page says OpenAI plans to retire custom GPTs in favour of Plugins, with retirement planned for December 11, 2026 in affected Enterprise workspaces and other plans expected to follow the same timeline.

A custom GPT's knowledge is also limited to files you upload by hand, up to 20 of them at up to 512 MB each. For a private tool your own team uses, these limits are reasonable. For the front door of your support, they add up to a channel your customers must sign up for, your team cannot see, and OpenAI is winding down.

How do you add a ChatGPT-style chatbot to your website with the OpenAI API?

You build it: the API gives you OpenAI's models and a few building blocks, and everything a visitor sees is your code. OpenAI's current route is the Responses API, the interface for sending a conversation to a model and getting its reply, used together with the Conversations API for storing threads.

The older Assistants API is gone. OpenAI's deprecations page lists its shutdown on August 26, 2026, a year after the deprecation was announced, with the Responses and Conversations APIs as its replacement. Tutorials written before the shutdown may still use it, so check the date on anything you follow.

For support answers, the key building block is OpenAI's file search tool, which lets a model retrieve passages from files you uploaded to a vector store, using semantic and keyword search. That covers the core of grounded answering. It does not cover:

• crawling your website, re-fetching changed pages and keeping the index current;
• the chat widget itself, how it loads on your pages and how it behaves on a phone;
• visitor sessions, rate limits and abuse protection on a public endpoint;
• the instructions and guardrails that keep the model on your content and make it say "I don't know";
• escalation, a team inbox, notifications and replies from named people;
• live lookups such as order status, with input validation and least-privilege credentials;
• analytics, exports and retention rules for the transcripts you now store.

Storage becomes your responsibility as well. By default the Responses API keeps application state for 30 days, and anything you log on your own servers follows whatever retention you set.

None of this is exotic engineering, and a capable developer can usually ship a prototype in days. The cost arrives in the months after: model and API changes like the Assistants shutdown, security fixes, and all the support tooling around the model. Our build vs buy breakdown prices each component, and the guide to AI chatbot security risks lists what a public bot that can call your systems has to defend against.

Building makes sense when the assistant is part of your product, needs deep access to your own systems, or must run under unusual data requirements. For a website that needs answers about shipping, plans and bookings, it is usually more work than the problem deserves.

ChatGPT, a custom GPT, the OpenAI API or an AI support agent?

The four options differ less in answer quality than in where they run, who can see the conversations and what happens when the answer is not in your content.

The table explains why so many teams end up using two columns at once: ChatGPT inside the team for drafting, and an AI support agent on the website for customers.

"Training ChatGPT on your website" is really retrieval

People ask how to train ChatGPT on their website, but support tools do not train a model on your pages. What they do is retrieval-augmented generation (RAG), a method in which the system searches your content for the passages most relevant to each question and the model writes its answer from those passages.

The difference matters for three practical reasons. First, updates are immediate: change a page or add a Q&A pair and the next answer uses it, because nothing has to be retrained. Second, the bot can say "I don't know": when the search finds nothing relevant, a well-built agent says so and offers a person instead of improvising. Third, your content sets the ceiling, so a clear FAQ written in plain sentences beats a pile of marketing pages.

Our plain-English RAG explainer walks through the indexing and answering steps, and the guide to training a chatbot on your website and PDFs shows the content preparation that makes the biggest difference to answer quality.

What does a customer-facing AI chatbot need that ChatGPT doesn't have?

A support channel needs five things a general assistant lacks: answers limited to your content, a person behind it, controlled access to live customer data, a record of every conversation, and your own brand on the surface.

Grounding with an exit. The agent should answer from your knowledge base and nothing else, and say plainly when the answer is not there. ChatGPT's great strength, a fluent answer on any topic, is a liability on a returns page.

A human handoff and a team inbox. When the visitor asks for a person, or the agent cannot help, the conversation should land with your team with the transcript attached and someone notified, during business hours and outside them.

Live lookups with verification. "Where is my order?" and "When does my plan renew?" need live data from Shopify, WooCommerce or Stripe. The agent should collect the order number, call a read-only API and show only the fields you allow. Anything personal should wait until the visitor has proved who they are. A verified visitor is one who has confirmed an email address with a one-time code.

Records and controls. You should be able to read every conversation, export it, delete it on a schedule, and decide which websites may load the widget.

Your brand, not someone else's. The chat should look and sound like your business, live on your own pages, and work without an account.

Where PepoChat fits, honestly

PepoChat is an AI support agent of the third kind, and it runs on OpenAI models: a fast model for replies and topic labels, and a vision-capable model for reading scanned PDFs and images. If you like how ChatGPT writes, the replies will feel familiar. What changes is everything around them.

The agent answers only from your knowledge base: your website (crawled, imported by sitemap or added link by link), uploaded documents, text snippets and Q&A pairs, pages synced from Notion or Google Drive, and published Help Center articles. It does not browse the internet or answer from general knowledge about your business. When retrieval finds nothing relevant, it says so and offers a person. When a visitor asks for a person, it escalates at once, your team is emailed, and the conversation waits in the team inbox with its transcript.

Installing it takes one script tag, the WordPress plugin, or no website at all: the hosted chat page gives your agent its own link for emails, QR codes and link-in-bio pages. That link is the honest replacement for "try our custom GPT": the same convenience, with no ChatGPT account for the customer and every conversation visible to your team. The script tag guide covers the website install on every common platform.

For live data, prebuilt actions look up Shopify order status, WooCommerce orders and Stripe subscriptions, and custom actions call any HTTPS API you point them at. Visitors can verify their email with a six-digit code, which unlocks personal features such as meeting memory and pre-filled bookings.

The limits are real. The widget is web only, with no WhatsApp, Messenger or Instagram today. Voice runs through your own Vapi account and does not share the chat agent's knowledge base. It is built to answer questions about your business from your content, not to be a general assistant for your visitors, which is the point but worth knowing. And PepoChat does not currently claim certifications such as SOC 2 or ISO 27001, so ask about current status if your procurement process needs one.

Every feature is on the free plan apart from new-lead emails to your team: 200 AI replies and 5 knowledge sources a month, with no credit card. Starter is $49 a month for 2,000 replies and Growth $149 a month for 10,000, as the pricing page lists.

What to do next

Keep ChatGPT for the drafting it does well, on a plan whose data settings you have checked. For the website, compare hosted options in our roundup of free AI chatbots for a website, and only take the build route if the assistant belongs inside your product. If you want to see an AI support agent answer from your own content, create a free PepoChat workspace, import your website as one knowledge source, and ask it the ten questions your customers ask most.