The best AI model for a website chatbot: how to choose

There is no single best AI model for a website chatbot. For a support chat that answers from your own documents, a smaller, cheaper model from any major provider is usually enough, because finding the right paragraph is the job of the search, not the model. The choice that really matters is the provider: how it treats your visitors' messages, where it processes them, and what a conversation costs.

This guide compares the five providers most WordPress chat plugins support (OpenAI's ChatGPT models, Anthropic's Claude, Google's Gemini, Mistral and OpenRouter) on the points that matter for a business site. Provider facts were checked on their official pages in September 2026. They change often, so check again before you decide.

What matters when choosing an AI model for a website chatbot

Benchmarks and "smartest model" rankings measure hard reasoning, coding and long tasks. A support chat does something much simpler: it reads five short extracts from your returns policy and writes three clear sentences. For that, five things matter.

1. Staying on your documents. The model must follow the instruction "answer only from these extracts" and say when they do not cover the question. How well each model does this is something to test on your own questions, not to take from a ranking.
2. What happens to the messages. Is API data used for training? How long is it kept? Is there a data processing agreement?
3. Where it is processed. Inside the EU, in the United States, or wherever the provider has capacity.
4. Cost per conversation. It depends on the model size and how much text each question sends.
5. Your languages. Quality and tone vary between models and languages. Test in every language your customers write in.

There is a sixth, practical point: a document-based chatbot also needs a model that builds the search index (embeddings). Not every provider offers one, which can mean a second account.

Claude, ChatGPT, Gemini, Mistral and OpenRouter compared

This table summarises the API terms (not the consumer chat apps, which have different rules) as published on 30 September 2026.

OpenAI

According to OpenAI's data controls page, data sent to the API is not used to train its models unless you opt in, and abuse monitoring logs are kept for up to 30 days. EU data residency exists, but it has to be set up on a new project, needs approval from OpenAI's sales team, uses a separate address (eu.api.openai.com) and costs more on newer models. Before counting on it, check that your chat tool can send requests to that address.

Anthropic (Claude)

Anthropic states that inputs and outputs from its commercial products, including the API, are not used for training by default, and are deleted from its backend within 30 days. Its data residency page offers two inference settings, global and US only; no EU option is listed. Anthropic does not offer its own embeddings model, so a Claude chatbot needs a second provider to build the index (Anthropic's documentation points to Voyage AI).

Google (Gemini API)

The Gemini API terms separate paid and unpaid services. On the paid tier Google does not use prompts and responses to improve its products. On the free tier it does, and human reviewers may read them, but for users in the EEA, Switzerland and the UK the paid-tier rules apply to everything. Two more lines matter for a public chat: you may use only paid services when you make an app available to users in the EEA, Switzerland or the UK, and the API must not be used in a site directed at, or likely to be used by, people under 18. The free tier is fine for testing; a live chat for European visitors needs billing turned on.

Mistral

Mistral is a French company. Its help centre says data is hosted in the EU by default, unless you choose its US endpoint. On the free plan of its developer platform, Mistral says it may use inputs and outputs for training, with the right to opt out; look for the training setting in your account whatever plan you use.

OpenRouter

OpenRouter is not a model maker: it gives you one account and one key for many models from many providers. It does not store your prompts and answers unless you opt in, but each request is handled by an underlying provider with its own policy. In your privacy settings you can exclude providers that train on data. In-region routing that keeps data in the EU is available on Business and Enterprise plans. It is useful for trying several models; for a live chat, pin the model and read that provider's terms.

Model size and cost per conversation

Every provider sells a range of sizes. As of September 2026, for example: Claude Haiku, Sonnet and Opus; Gemini Flash-Lite, Flash and Pro; Mistral's Ministral, Small, Medium and Large; and OpenAI's smaller and larger GPT models. Larger models write better on hard questions and cost more per word. For a support chat that answers from your documents, start with a small or mid-size model and move up only if tests show a problem.

AI providers charge per token (a token is roughly three quarters of a word), with separate prices for text sent in and text written out. Each chat question sends your instructions, a few extracts from your documents and the question, and gets a short answer back. To estimate a cost:

1. Guess the size of one question: for example 3,000 tokens in (instructions plus five extracts) and 300 tokens out.
2. Open the provider's pricing page and find the price per million tokens for the model you want.
3. Multiply: (3,000 × input price + 300 × output price) ÷ 1,000,000.

With a hypothetical price of $1 per million tokens in and $5 per million out, that is $0.003 + $0.0015, less than half a cent per question. Real prices vary a lot between models, so do the sum with today's figures. Whatever you choose, set a daily ceiling on answers in your chat tool, so a busy day or an abusive visitor cannot surprise you.

Which provider fits which case

• Messages must be processed in the EU: Mistral does this by default. OpenAI's EU data residency and OpenRouter's EU routing are options if you qualify and your tool supports them.
• You want the simplest setup: a provider that both answers and builds the index with one key (OpenAI, Gemini or Mistral) means one account and one bill.
• You want to try before paying: Gemini's free tier lets you test the whole setup; switch billing on before European visitors use it.
• You already use Claude and like how it writes: it can answer your chat too; plan for a second key for the index.
• You want to compare many models: OpenRouter gives you one key for all of them.
• Your site is aimed at teenagers or children: read each provider's age rules first; Gemini's API terms exclude such sites.

Whatever you pick, accept the provider's data processing agreement and name it in your privacy notice. Our guide on AI chatbots, GDPR and the AI Act explains what else to write.

How to test models on your own documents

The only test that counts is your own questions on your own documents. Write down ten real questions from your inbox, including two that your documents do not answer, and ask them to each model you are considering. Check three things: is the answer right, does it stay on your documents, and does it say "I don't know" when it should?

In Answer Desk, our WordPress chat plugin, you can do this without rebuilding anything:

1. On Answer Desk > AI provider, in the Who answers card, click a provider, leave Model on The recommended one or pick another, and click Save changes.
2. Make sure its key is saved in the API keys card.
3. On Answer Desk > Start here, type each question in the Try it box and click Ask. You see the answer and the extracts it used.
4. Switch provider and repeat.

Each test costs what a real question costs. All five providers are included in the free version. The guides Connect an AI provider and Fine tuning and running costs explain the keys, models and spending limits, and Build the index covers the index provider. For how the chat finds the right extract in the first place, see AI chatbots trained on your own documents.

Summary

For a website chatbot that answers from your documents, the best AI model is usually a small or mid-size one from a provider whose data terms you are comfortable with. Choose Mistral if EU processing matters most and you want it by default; OpenAI, Gemini or Mistral if you want one key for everything; Claude if you prefer its writing and accept a second key; OpenRouter if you want to compare many models. Then test with your own questions, set a daily ceiling, and check the terms again once a year, because they change.