Getting started
Install AnswerRails, connect it to an AI provider, give it something to answer from, and get your first real answer. About five minutes of work, in this order.
What you need first
- WordPress 6.2 or later, on PHP 8.0 or later. A standard single-site install; WordPress Multisite is not currently supported.
- An account with an AI provider. Mistral, OpenAI or Anthropic Claude, on every plan. A Mistral key is free to create with no card, which is the quickest way to see the plugin work.
- Something worth answering from. A handful of pages a visitor actually asks about beats a whole site dump.
1. Install and activate
Install the free plugin from Plugins → Add New by searching for AnswerRails, or upload the zip at Plugins → Add New → Upload Plugin. Then activate it from the Plugins list.
On activation you are asked whether to share diagnostic data — your site address, your WordPress and PHP versions, and activation events — with the licensing platform, Freemius. That is off unless you accept it, and it never includes visitor conversations.
An AnswerRails menu then appears in the admin sidebar, with Overview, Chat Logs, Chat Settings, Advanced Features and Knowledge Base. Nothing answers yet: the widget needs a provider key.
If you are starting on Pro, checkout completes in-admin from AnswerRails → Advanced Features and the Pro build installs in place, leaving your settings, knowledge base and chat logs untouched. If you were sent the Pro zip directly, upload it the same way and activate your licence key on the Account screen when prompted.
2. Connect an AI provider
AnswerRails is bring-your-own-key. You supply an API key, the calls go straight from your site to that provider, and the usage is billed to your own provider account. There is no AnswerRails server in between.
Go to AnswerRails → Chat Settings → AI Provider.
Choose your provider from the drop-down. The key and model fields for that provider appear below it.
Paste your API key and choose a model.
Pick from the list, or choose Custom model id… to enter any model id the provider supports — you are told when an id is not one we curate, and nothing is rejected on save.
Save, then run Test APIs on the Overview.
It makes a real call to your chat provider, to knowledge search and to the moderation service, and names in plain language which one failed if any did.
New installs default to Mistral, whose API key is free to create at console.mistral.ai. OpenAI keys come from platform.openai.com; Anthropic keys from platform.claude.com. Keys are stored in your WordPress database, sent only to the provider they belong to, never reach the visitor’s browser, and are always deleted on uninstall.
Training on your API traffic is on by default until you disable it. In the Mistral console open Admin → API → Privacy and disable the option to improve Mistral’s models. That is the API privacy page — the separate “Vibe” privacy page does not cover API traffic — and the opt-out applies only from the moment you set it. AnswerRails cannot check that setting for you. Mistral’s “Labs” models are not in our list because Mistral trains on their inputs regardless of the opt-out.
3. Give it something to answer from
The assistant answers only from your knowledge base, so this step decides what it can say. Everything is under AnswerRails → Knowledge Base. There is nothing to switch on first: Enable Knowledge Base, in that screen’s settings, is ticked from the start. It is there so you can turn grounding off — unticked, the assistant stops using your content and stops being told what pages your site even has, which leaves it nothing of yours to answer from.
The fastest start is Site Content: a multi-select picker with search and category filters that adds existing pages and posts in bulk, with rows already in the knowledge base badged so you cannot add them twice. Results are paged, and selecting everything that matches reports its own limit rather than presenting itself as the whole site. Keep the page open while a large batch runs; anything already added is kept if you leave, and re-running the same selection safely skips it. Items sync as they are created.
Say you run a small roofing business. The questions visitors bring are predictable: which areas you cover, whether you are open on a Saturday, how to get a quote, what the guarantee is on a new roof against a repair, and whether a job will need scaffolding. Add the pages that answer those — your services page, your areas page, your contact page, your guarantee terms — and the assistant answers from them. A whole-site dump adds pages nobody asks about and adds nothing to those five answers. The operator guide works the same example through each of the four ways in.
Three other routes are there when you need them: Upload Document for files (plain text is read into the local index; other formats are indexed by title), Structured Data (JSON) for FAQs, specs and price lists, and, on Pro, External URLs for a page beyond your WordPress site. The operator guide covers all four.
4. Ask it something
Under AnswerRails → Chat Settings → Basic Setup, check that Enable Chat Widget is on and set Display Location: home page only, all pages, a list of pages you choose, or shortcode only. Choosing the custom-pages option reveals the picker directly beneath the drop-down, with published pages and posts as separate groups; each group lists the 50 most recently published and says so when there are more.
Then open your site, open the widget, and ask something your pages answer. You should get an answer with a Sources block under it naming what it drew on. Ask something your pages do not cover, and it should say so rather than improvise — that refusal is the product working, not failing. When it says “I don’t know” explains why.
5. Then, when you are ready
- Better search. An OpenAI key enables optional vector search, which matches on meaning rather than keywords, whichever provider writes the reply. Turning on vector search.
- Consent, logging and retention. Nothing is logged until you switch it on. Consent, privacy and retention.
- Escalations to a person (Pro). Setting up escalations.
