Compare
Nine questions worth asking of any AI assistant before you put it on your site — with our answer, and what we found when we asked them of the most-installed alternatives.
We read the public documentation of the five most-installed WordPress AI chat plugins and hosted widgets on 20 August 2026. The right-hand column is what those pages say, summarised; where a page is silent on a question we treat that as not stated, not as no. We do not name them here, because a fair comparison needs more room than a table — if you are weighing a specific product, ask it these questions directly and read the answer against ours.
| The question | AnswerRails | Across the five we checked |
|---|---|---|
| Where does the AI run, and who is in the middle? | On your own provider account — OpenAI, Mistral or Anthropic — called directly from your server. No AnswerRails server is in the path and nothing is metered. Your key, your bill | Three let you bring your own key; two are vendor-hosted, with conversations passing through and billed by the vendor, one of them per conversation and one per message credit. |
| Does it answer only from my content, and what happens when the answer isn’t there? | Answers are grounded in your knowledge base and checked on the way out; an unsupported answer or an invented link is withheld and an honest refusal served instead. When it says “I don’t know” | All five retrieve from your content. One states it answers solely from your data and hands off to a human otherwise; the other four do not state what happens when the content doesn’t hold the answer. |
| Can the visitor see where an answer came from? | A Sources list under every grounded answer. | Two state that links are shown under replies; three do not state it. |
| Can I ask for consent before the chat starts, and prove it later? | An optional consent gate with server-side receipts recording the exact wording agreed to, and withdrawal as easy as giving. Consent, privacy and retention | Two offer a consent notice that locks the chat until accepted; one recommends a consent field; one shows a notice; one does not state it. None describes a stored consent receipt. |
| Where do transcripts live, and for how long? | Logging is off by default. When on, transcripts are in your WordPress database, anonymised by default, with identified logs purged after a retention period you set. Where your data goes | Three keep transcripts in your WordPress database; two keep them in the vendor’s cloud. One offers a configurable auto-delete period; four do not state one. |
| What happens when a visitor asks for their data? | Registered with WordPress’s own Export and Erase Personal Data tools, plus a panel to find, export, anonymise or delete by email, session or IP, with legal hold and an audit trail. | None states integration with WordPress’s Export/Erase tools. Two describe deletion on request to the vendor; three do not state a process. |
| Is what visitors type checked before it reaches the model? | Every message is checked by your provider’s moderation service first, on every plan, with no off-switch. | Two offer banned-word lists and an optional provider moderation check; three do not state input moderation. |
| Is it accessible? | Built to WCAG 2.2 AA and tested on real assistive technology; the detail, including what we know is imperfect, is in our accessibility statement. | One states that it supports WCAG, without a conformance level; four do not state accessibility. |
| Is the compliance layer free? | Every privacy and compliance control is free — and we don’t move them behind payment. Free vs Pro | The two hosted products meter the AI itself. Of the three self-hosted, two put part of their privacy controls behind a paid tier, and one puts grounding in your own content behind one. None documents the full set at any price. |
Where they are broader
Several of the five are broader products: live human chat in the same widget, site-wide content generation and automation, external vector databases, and long provider lists. AnswerRails is built to be one thing — a customer-service assistant that can be trusted in public — and that focus is deliberate. It is not a ceiling: providers, models and search options are added as they earn their place, and the model catalogue is kept current as providers release and retire models. What it can’t do is the honest list of today’s limits.
