Introduction
Here’s the honest reality: you can’t actually embed ChatGPT in your website – ChatGPT is a product from OpenAI that is used at chatgpt.com. What you can do is embed something powered by the same underlying AI model on your website, and the result you get depends on which method you pick and how much work you do to control what it says.
“Embed ChatGPT in website” is one of the most-searched AI questions of 2026, and almost every guide that answers it is missing the point. This guide explains what’s actually possible, the three real methods (and their trade-offs), the customer-facing risks nobody warns you about, and the techniques that make the result safe to put in front of real customers.
Disclosure: we make Resolve247, an AI support chatbot that builds on top of OpenAI’s models (focusing it on your own content) for reliable use in customer facing situations. We’ve included it as one of several options in Method 2, and we’ll tell you when one of the alternatives is the better fit. If you’re not specifically interested in the relation to ChatGPT, you may want to start with our broader companion guide on how to add an AI chatbot to your website.
Quick Answer: Can You Embed ChatGPT in Your Website?
You can’t embed “ChatGPT” itself in your website, but you can embed something powered by the same underlying model using one of three methods: share a Custom GPT link, install an AI chatbot widget, or build a custom integration via the OpenAI API. Each has very different trade-offs for real customer-facing use.
Part 1: What You’re Actually Asking For
Can You Actually Embed ChatGPT in a Website? (The Honest Answer)
No, not in the sense that most people mean.
It helps to separate two things that get conflated. ChatGPT is the consumer product at chatgpt.com made by OpenAI. The OpenAI models (GPT-5.5, GPT-5.4, and so on) are the underlying neural networks that power ChatGPT, and they’re available separately through the OpenAI API.
ChatGPT-the-product is more than just the model. It has its own undisclosed system prompt, its own memory, and built-in tools (Browse, Code Interpreter, image generation, Canvas, voice). It was designed for individual users having open-ended conversations, not for sitting on your support page answering questions about your refund policy.
The OpenAI API gives you the raw model without any of that consumer-product wrapping. It’s the same “engine” used by ChatGPT; that you can then build on top of.
When people say “embed ChatGPT in my website”, they usually mean one of three different things, all of which produce different results:
- Sharing a Custom GPT link: building a Custom GPT inside chatgpt.com and putting a link to it on your site. It’s not really “embedded”; the link opens chatgpt.com in a new tab.
- Installing an AI chatbot widget: a purpose-built platform that uses the same model but focuses it to only use your content and renders a chat widget on your site.
- Building a custom integration via the OpenAI API: you write the code yourself, you control the prompt, you build the chat UI.
Each gives a different result for your customer, with very different effort, cost, and risk. We’ll cover all three below. But first, you need to understand something important about what ChatGPT knows out of the box, because that’s the thing that determines which method you should actually pick.
What Does ChatGPT Actually Know? (The Critical Detail Most Guides Skip)
Out of the box, an OpenAI model knows what it learned from the public internet up to its training cut-off date. That includes a lot:
- It knows about your competitors.
- It knows general industry conventions and common practices.
- It might know about your business, if your public content existed before the cut-off and hasn’t changed since.
It does not know:
- Anything that happened or changed after the training cut-off.
- Your private information, internal policies, current pricing, inventory, customer accounts, today’s stock.
- The specific way your business does things, unless that’s well-documented publicly.
The hidden danger is what happens when you ask it something it doesn’t know. Because the model is trained to be helpful, it often invents a plausible-sounding answer rather than admitting ignorance. The industry term is “hallucination”. A recent benchmark from the Vectara Hallucination Leaderboard puts baseline hallucination rates for ungrounded large language models in the low-to-mid single digits even on factual summarisation tasks, and customer-support questions are harder than summarisation.
In practice, that looks like:
- “What’s your refund policy?” → it invents one based on industry norms.
- “Which is better, [your product] or [competitor]?” → it compares them, sometimes favourably for the competitor it knows more about.
- “Can I get a discount?” → it might cheerfully offer one you don’t actually give.
- “Is [feature] supported?” → it guesses, sometimes hedging with “probably” or “I believe”.
An AI that confidently invents answers is worse than no AI at all, because customers trust what the chatbot says. And what your chatbot says, your business has to stand by. Four well-documented cases make that point concrete.
Air Canada (February 2024). Air Canada’s website chatbot told Jake Moffatt, a grieving customer, that he could apply for a bereavement fare retroactively after the flight. He did. Air Canada then said the policy didn’t exist and refused the refund. The British Columbia Civil Resolution Tribunal ruled in Moffatt v. Air Canada that Air Canada was responsible for what its chatbot said and ordered it to honour the invented policy plus pay damages. The airline’s defence, that the chatbot was a “separate legal entity”, was explicitly rejected.
DPD (January 2024). A customer prompted DPD’s chatbot until it swore at him and wrote a poem criticising DPD’s customer service. The exchange went viral on Twitter and was picked up by the BBC and the Guardian. DPD disabled the bot the same day.
Chevrolet of Watsonville (December 2023). A car dealership embedded a ChatGPT-powered chatbot on its website. A user prompted it until it “agreed” to sell a 2024 Chevy Tahoe for $1 and called it “a legally binding offer – no takesies backsies”. The exchange went viral, the dealership pulled the bot, and although the deal was never enforced, the lesson is still real: an off-the-shelf ChatGPT integration with no guardrails can be made to agree to almost anything.
New York City “MyCity” chatbot (March 2024). An investigation by The Markup found NYC’s small-business chatbot was telling business owners to do things that were illegal under city law, including that landlords could refuse tenants on housing vouchers, and that bosses could take a cut of workers’ tips. The bot was confident and plausible, and it was wrong.
The pattern is the same in every case. The model wasn’t grounded in the right source material, no guardrails caught the bad answers, and the business carried the cost. Whatever your AI says on your website, your business will be expected to stand by. The three methods below differ in exactly how much control you have over what it actually says.
What’s Actually Inside ChatGPT (and What an AI Chatbot Widget Adds)
The previous section showed what goes wrong when ChatGPT answers about your business. This section shows why, by breaking ChatGPT down into its components, and then shows what each of the three methods changes about those components.
What’s inside ChatGPT (the consumer product at chatgpt.com):
- The underlying language model (GPT-5.5, GPT-5.4, and so on): OpenAI’s neural network. The part that does the language understanding and generation.
- General language capabilities: reasoning, writing, summarising, translation, code, maths.
- Pre-trained knowledge from the public internet: everything in the training data up to the cut-off. Knows your competitors. Doesn’t know your business specifically unless it’s public. This is the component that causes hallucinations about your business.
- OpenAI’s hidden system prompt: undisclosed instructions OpenAI applies on every conversation (persona, format, safety rules).
- Custom instructions and memory: optional, set per-user inside the user’s own ChatGPT account.
- The chatgpt.com interface and built-in tools: the chat UI on OpenAI’s domain, plus tools like web search, Code Interpreter, image generation, Canvas, and voice.
What an AI chatbot widget (Method 2) keeps, swaps, adds, and removes:
- ✅ Keeps: the underlying model and its general language capabilities.
- 🔄 Swaps: OpenAI’s hidden system prompt → your system prompt (e.g. “Only answer using the retrieved context. Refuse off-topic questions. Use this tone. Never recommend competitors.”).
- ➕ Adds: RAG (Retrieval-Augmented Generation). Your knowledge base is indexed, and the most relevant pieces are pulled into the model’s context per query. This is the headline addition. It’s how the chatbot answers about your business instead of just the public internet, and the reason it can answer accurately about content that didn’t exist at the model’s training cut-off.
- ➕ Adds: Guardrails. Input filtering for jailbreak patterns, output filtering for competitor mentions or off-brand topics, refusal patterns for off-topic questions.
- ➕ Adds: A branded chat widget embedded on your site (not chatgpt.com).
- ➕ Adds: Business features. Conversation analytics, lead capture, helpdesk integration, human handover, per-user context (account information, order history).
What a bespoke API build (Method 3) gives you:
- ✅ Keeps: the underlying model, the language capabilities, OpenAI’s API-level safety guardrails.
- ⚠️ Everything else is on you to build. If you want RAG, you build it. If you want guardrails, you build them. If you want a widget UI, you build it. The API is the engine; you build the rest of the car around it. Method 2 platforms have already done this work for you.
The bottom line: “embedding ChatGPT” probably isn’t the solution you’re looking for. The next section explains the three techniques that Method 2 platforms apply for you, and that you’d have to apply yourself in Method 3, to keep the AI on-message.
How to Control What ChatGPT Says (The Three Techniques)
Controlling what an LLM says isn’t one thing. It’s three layers, and good systems use all of them together.
Technique 1: System Prompts (Prompt Engineering)
A system prompt is a block of instructions you give the model that the user can’t see. It runs ahead of every conversation and shapes how the model behaves: tone, persona, refusal topics, response format, safety boundaries.
A drastically simplified system prompt for a support chatbot looks something like this:
You are the support assistant for [Company]. Answer questions using only the context provided below. If the retrieved context doesn’t contain the answer, say so honestly and offer to connect the user with a human. Never invent prices, policies, or features. Never recommend or compare with competitor products. Match this tone: friendly, direct, no marketing language. Refuse questions outside [Company]’s product and support remit by politely declining and suggesting the user search or contact us.
What it can do: set tone, persona, refusal topics, response format. What it can’t do: make the model know things it doesn’t already know. And system prompts can be jailbroken with enough adversarial pressure (see DPD).
Technique 2: In-Context Grounding
In-context grounding means putting the relevant information into the conversation itself, so the model is answering from facts you supplied rather than from its training data.
The simplest version: paste your refund policy directly into the system prompt. The model now has the right answer in front of it when someone asks.
The limit is the context window. Modern models can hold tens of thousands of tokens of context, but pasting your entire knowledge base into every conversation is slow, expensive, and brittle. In-context grounding works for small, stable sets of facts. It doesn’t scale to a real knowledge base.
Technique 3: Retrieval-Augmented Generation (RAG)
RAG is the grown-up version of grounding. A retrieval system indexes your knowledge base (help docs, website pages, FAQs, uploaded files), and on every query it pulls the most relevant pieces and feeds them to the model as context. Automatically. Per query.
That means the model answers from your content, not from what it half-remembers from the public internet. RAG can ground answers in a knowledge base of any size, cite sources, and update instantly when your content changes.
This is what serious AI chatbot widgets do under the hood, what Botpress, Chatbase, CustomGPT, Pickaxe and Resolve247 are all built around, and what Method 2 below productises. For a deeper framework on evaluating tools in this category, see our guide on how to choose an AI customer support chatbot.
The limit: RAG is only as good as the source material. Bad docs in, bad answers out.
Defence in Depth
Good systems combine all three, and more. A clear system prompt sets the rules. RAG retrieves the right facts on every query. Guardrails sit on top, input filters catch jailbreak patterns, output filters catch off-brand content, and refusal patterns (“I can only help with questions about our products”) keep the bot on topic.
No single technique is enough. The bots in the example stories above probably had some kind of system prompt but what they’re missing was a reliable combination, made by specialists.
Part 2: The 3 Real Methods to Embed ChatGPT in Your Website
There are three genuinely different ways to embed ChatGPT in your website (or, more accurately, to embed something powered by the same model). Each method below follows the same structure: what it actually does, setup, what the customer experiences, effort and cost, control over output, when it’s the right pick, and when it isn’t.
Method 1: Share a Custom GPT Link
What it actually does: you build a Custom GPT inside chatgpt.com using OpenAI’s GPT builder. You give it custom instructions and (optionally) upload some knowledge files. You then share its URL. This is the most literal interpretation of “add ChatGPT to your website”: you add a link to ChatGPT. There is no real embed. Clicking the link opens your Custom GPT on chatgpt.com in a new tab.
Critical caveats most guides skip:
- The GPT runs on chatgpt.com, not on your site. Your customer leaves your site to use it.
- Your customer needs a (free) ChatGPT account to use it. No account, no chatbot.
- You can’t style the page to match your brand. The interface is OpenAI’s.
- It’s still the open-ended ChatGPT experience, so the GPT can talk about anything, including your competitors, not just your business.
- You can’t track usage analytics from your end.
Setup:
- Sign in to chatgpt.com with a paid ChatGPT account (Custom GPTs are a paid-tier feature).
- Click Explore GPTs → Create.
- Use the builder to write a name, description, and custom instructions.
- (Optional) Upload reference files for grounding.
- Set the GPT to Public or Anyone with a link.
- Copy the share URL and add it to your site as a link or button.
Total time: about 30 minutes.
Effort and cost: very low effort, no ongoing cost to you (the customer pays via their own ChatGPT account or hits the free-tier limit).
Control over output: medium. Instructions and uploaded files give some grounding, but the GPT is still the open-ended ChatGPT experience underneath.
Right pick when: you want a free novelty AI assistant on a low-stakes page (a marketing demo, a personal site, a portfolio), and your visitors are technical enough to already have ChatGPT accounts.
Wrong pick when: most actual business cases. Anything customer-facing where the customer leaving your site, brand inconsistency, or invented answers would matter. Which is most of them.
Method 2: Use an AI Chatbot Widget
What it actually does: a purpose-built platform that you point at your knowledge base (help docs, website pages, FAQs, uploaded files). The platform learns your content, retrieves the most relevant pieces per query, and feeds them to the model as grounding context. Then it renders a chat widget on your site, styled to match your brand.
This is what Part 1’s Technique 3, RAG, looks like productised. The grounding work is done for you. The widget uses the same underlying model family as ChatGPT (GPT-5.5, GPT-5.4 and so on), but it’s properly grounded in your content rather than the public internet.
What the customer experiences: an embedded chat widget styled to match your brand. This is what most people are actually picturing when they search for a ChatGPT chatbot for their website: answers come from your content, the widget can cite sources, hand off to a human, capture leads, and keep the customer on your site throughout the conversation.
The main options (alphabetical):
| Tool | Best for | Pricing (entry) |
|---|---|---|
| Botpress | Developer teams who want a visual flow builder plus AI. Steeper learning curve, more powerful for branching logic. | Free tier; paid from $189/mo |
| Chatbase | Quick deployment, simple setup. Strong on URL crawling and file ingestion; lighter on integrations. | From $40/mo |
| CustomGPT | Larger knowledge bases and citations-first answers. Strong source-attribution UI. | From $99/mo |
| Pickaxe | No-code builders and indie hackers who want to ship AI tools quickly without a developer. | From $37/mo plus AI credits |
| Resolve247 | SMBs who want to add easily AI to their existing workflow, with an anti-hallucination guarantee. Disclosure: this is our product. | From $35/mo (see pricing) |
Pricing accurate at time of writing, check each vendor’s site for current rates.
Setup: connect your knowledge sources (URL crawl, file upload, direct integrations like Notion or your helpdesk), install the widget script on your site (or integrate into your existing chat widget), customise the system prompt and branding, and publish. Typically one to three hours start-to-live for a small site.
Effort and cost: low-to-medium effort, depending on how clean your knowledge base is. Monthly subscriptions in the $30-$500 range for SMB use. Maintenance is minimal, the platform handles model updates, retrieval, and scaling. You re-index when your content changes, often automatically.
Control over output: high. You control the source material, the system prompt, and (depending on the platform) guardrails, refusal topics, and tone. Higher than Method 1, and more predictable than rolling your own Method 3, because the platform vendors have already spent years tuning the retrieval and guardrail layers.
Right pick when: your goal is customer-facing AI that answers accurately about your specific business: support questions, FAQs, product questions, sales objections, anywhere accuracy and brand alignment matter.
Wrong pick when: you need a fully-scripted conversation flow with pre-defined answers to choose from (these solutions can do this, but there are cheaper options available if you want a rule-based chatflow), or you only want a low-stakes novelty AI (Method 1 is cheaper).
A genuine caveat on RAG: it isn’t magic. The model still needs good source material. If your help docs are sparse, contradictory, or out of date, the chatbot’s answers will be sparse, contradictory, or out of date. You can always give the AI custom instructions to try to override content retrieved by RAG, but for the best reliability please ensure your documentation is correct.
Method 3: Build Your Own Using the OpenAI API
What it actually does: this is the do-it-yourself path to integrate ChatGPT into your website at the deepest level. You write the backend yourself. You control the system prompt, you decide whether and how to add grounding, you build the frontend chat UI, you handle billing alerts, abuse detection, and everything in between.
Architecture: frontend chat UI → your backend endpoint → OpenAI API → response streamed back to the user.
There’s much more to a chatbot than the widget you can see. There’s the contextual retrieval layer, source scraping, chunking & ingestion, session management, history management, rate limiting, abuse detection, billing alerts, observability and all kinds of other parts that come with a production grade AI system.
Effort and cost: high.
- Setup: one to four weeks of developer time to get to a passable v1.
- Maintenance: continuous. Model updates, prompt drift, security, abuse handling, billing alerts.
- Cost: OpenAI API token costs scale with usage. They can spike unexpectedly under abuse, virality, or runaway prompt loops. Set a hard usage cap on your OpenAI account, set per-user rate limits in your backend, and log every call. Always.
Control over output: total, if you put the work in. Which is the catch. The Method 2 platforms have been doing the work for years; you’re starting from scratch.
Right pick when: you have engineering capacity and a genuinely bespoke conversational workflow no off-the-shelf tool supports. Taking bookings against a custom system, deeply custom logic on your own data, internal-tool integrations, or an experience that demands a chat UX no widget vendor can deliver.
Wrong pick when: your goal is “answer customer support questions accurately”. You’d be rebuilding Method 2 from scratch, and badly for the first few months. The platforms exist for a reason.
Function calling is worth mentioning: the OpenAI API supports the model calling functions you define, which is how you give the model the ability to do things (look up an order, refund a payment) instead of just answering. Powerful, and well documented in the OpenAI API docs.
A 30-Second Decision Guide
- Free novelty AI on a marketing or personal site, and you don’t mind sending visitors to chatgpt.com? → Method 1 (Custom GPT share link).
- Customer-facing chatbot that answers accurately about your business (support, FAQs, product questions)? → Method 2 (AI chatbot widget).
- Engineering capacity and a bespoke conversational workflow no off-the-shelf tool supports? → Method 3 (build via the OpenAI API).
If you’re still not sure which side of the support/non-support line you sit on, our guide on whether an AI chatbot is right for you walks through the situations where AI helps and where it doesn’t.
Side-by-Side Comparison
| Method 1: Custom GPT link | Method 2: AI chatbot widget | Method 3: OpenAI API build | |
|---|---|---|---|
| Setup effort | 30 min | 1-3 hours | 1-4 weeks |
| Maintenance | None | Minimal | Continuous |
| Cost (entry) | Free | $30-$100/mo | API tokens + dev time |
| Control over output | Medium | High | Total (if you build it) |
| Grounded in your content | Partial | Yes (RAG) | Only if you build it |
| Customer stays on your site | No | Yes | Yes |
| Time to live | Same day | Same day | Weeks |
| Best for | Novelty demos | Customer-facing support | Bespoke workflows |
Frequently Asked Questions
Can I embed ChatGPT directly in my website?
No, not exactly. ChatGPT-the-product lives at chatgpt.com and can only be linked to, not embedded. What you can embed is something powered by the same underlying OpenAI model, such as an AI chatbot widget (the most common option), or your own integration built on the OpenAI API.
Will customers need a ChatGPT account to use the chatbot on my site?
Only with Method 1 (sharing a Custom GPT link). Method 2 (an AI chatbot widget) and Method 3 (an API build) run on your infrastructure and don’t require the customer to have a ChatGPT account.
What’s the difference between ChatGPT and the OpenAI API?
ChatGPT is the consumer product at chatgpt.com, the chat interface plus OpenAI’s hidden system prompt, and built-in tools like Browse and Code Interpreter. The OpenAI API gives you raw access to the underlying models (GPT-5.5, GPT-5.4, and so on) without the consumer-product wrapping. Same engine, but different interface for different purposes.
Can ChatGPT answer questions about my specific business?
Only if your business information was on the public internet before the model’s training cut-off, and hasn’t changed since. For anything current (your live pricing, your refund policy, today’s stock), the model has to be grounded in your content using RAG (Retrieval-Augmented Generation), which is what AI chatbot widgets do. Without grounding, ChatGPT will often invent plausible-sounding answers, with real consequences (see the Air Canada case).
What is RAG and why does it matter?
RAG stands for Retrieval-Augmented Generation. A retrieval system indexes your knowledge base, and on every query it pulls the most relevant pieces and feeds them to the model as context. The model then answers from your content rather than from its training data. RAG is the reason a good AI chatbot widget can answer accurately about your current pricing or policies, and the reason raw ChatGPT can’t.
How much does an embedded ChatGPT chatbot cost per month?
Method 1: free to you, customer pays via their ChatGPT account. Method 2: $30-$500/month typical SMB range. Method 3: OpenAI API tokens (a few dollars to a few thousand per month, depending on volume) plus continuing developer time.
Is there a ChatGPT WordPress plugin?
Several. They generally fall into two camps: Method 2 platforms with a dedicated WordPress install (the safer option), and thin Method 3 wrappers that don’t include grounding or guardrails out of the box. Either way, you’re getting ChatGPT for WordPress via a third-party tool, not from OpenAI directly. For a customer-support use case, see our WordPress chatbot integration and the broader walkthrough on adding an AI chatbot to a WordPress site.
How do I embed ChatGPT in WordPress, Shopify, Wix, or Squarespace?
For Method 1, you add the share link as a button on any platform. For Method 2, every major widget vendor installs via a single line of JavaScript pasted into your theme’s header, same on WordPress, Shopify, Wix, and Squarespace. Method 3 requires a backend, which is harder on hosted platforms (Wix, Squarespace) than on WordPress or Shopify, where you have more code access. Our broader companion guide on how to add an AI chatbot to your website covers the platform-by-platform specifics.
What is a ChatGPT iframe or ChatGPT embed code?
There’s no official iframe or embed code from OpenAI for ChatGPT-the-product. People search for these terms hoping for a copy-paste solution, but what they actually want is one of the three methods in this article, almost always Method 2 (an AI chatbot widget), which is installed via a copy-paste JavaScript snippet.
Conclusion
When you set out to embed ChatGPT in your website, you’re not really embedding ChatGPT. You’re embedding something powered by the same model, and which method you pick determines what your customer actually experiences.
The honest recommendations:
- For demos and marketing novelty: Method 1 (Custom GPT share link). Free, fast, fine for low-stakes use.
- For accurate customer-facing chat (support, FAQs, product questions): Method 2 (a RAG-based AI chatbot widget). It’s what most readers actually need.
- For bespoke conversational workflows: Method 3 (build on the OpenAI API). Powerful, expensive, slow.
If your goal is customer support specifically (accurate answers about your business, in your tone, without the risks Air Canada and DPD learned the hard way), that’s the use case we built Resolve247 for. Compare it to the other AI chatbot widget options above and pick the one that fits your situation.
If you’re not committed to ChatGPT specifically and want to evaluate the wider AI chatbot landscape, our companion guide on how to add an AI chatbot to your website walks through eight different options, including the ones that don’t use OpenAI at all.