AI Chat Plugins: Grounded, Scripted, or Just a Writing Tool
Searching for an AI chatbot plugin and searching for a ChatGPT or Claude plugin returns two unrelated product categories, and the visitor-facing one has a failure mode the marketing never mentions: a confident answer about your refund policy that you never wrote.
Two different searches lead here, and they want opposite products. Someone looking for a WordPress plugin for ChatGPT usually wants help writing posts. Someone looking for an AI chatbot wants a widget that answers visitors. Treating them as one category is how sites end up with an editor toolbar they never use and a support widget nobody grounded.
| Category | Where it runs | Real job |
|---|---|---|
| Editor assistant | wp-admin, while you write | Drafting, rewriting, summarising, alt text |
| Scripted bot | Front end | Routing, common questions, capturing details |
| Grounded assistant | Front end | Answering from your own documented content |
| Live chat with AI assist | Front end plus an inbox | Deflection, with a human behind it |
Editor assistants
These are the plugins named after model providers, and a WordPress plugin for Claude and one for ChatGPT are usually the same product with a different provider behind it. They add a panel or block to the editor and call an API with a key you supply. Useful, low risk, and easy to evaluate: does it work inside the block editor properly, does it store your key securely rather than exposing it to the browser, and can you switch providers without losing your prompts?
Do not let one write your entire site unattended. Published pages carry your name, and the correction cost of a plausible-sounding error in an evergreen article is far higher than the time saved. Anything factual — prices, policies, technical steps — needs a human who knows the answer.
Scripted bots
The unfashionable category, and still the right one for many sites. A decision tree with buttons cannot hallucinate. If your goal is to route people to the right page, book a call or collect a query, a scripted flow does it deterministically and costs nothing per message.
The honest test is whether your top ten enquiries have fixed answers. If they do, a scripted bot answers them perfectly and a language model only adds risk. If the questions are open-ended and your documentation is genuinely good, retrieval becomes worth it. For anything that ends in "send us the details", a plain form is still better than either — see contact form plugins.
Grounded assistants
This is what most people mean by an AI chatbot plugin: your pages, products or documentation are indexed, and the model answers using retrieved passages rather than its own recollection.
Evaluate on the retrieval, not the model name:
- What gets indexed, and how re-indexing happens when you edit a page. A stale index answers with last quarter's pricing.
- Whether answers cite sources. A link back to the page lets the visitor verify, and lets you spot a bad answer.
- Refusal behaviour. What it says when nothing relevant is retrieved. "I do not have that information, here is how to contact us" is the correct answer and a surprising number of products cannot produce it.
- Handoff. A route to a human, with the transcript attached.
- Transcript review. You need to read what it said. Logs are the only way to find out that it has been confidently wrong for three weeks.
- Provider portability. Being able to point the plugin at a different provider protects you from one vendor's pricing or availability changes.
Cost, security and consent
These three decide more real deployments than answer quality.
Cost is usage-based on almost every provider, so an unmetered public widget is an open tap. Insist on per-session rate limits, a conversation length cap, and a provider-side spend limit. Check the plugin does not accept an arbitrary model or prompt from the browser.
Security. The API key belongs in server-side configuration, never in front-end JavaScript, and the endpoint that proxies to the provider must be authenticated or at least rate limited. Anything that lets an anonymous visitor drive your key is a liability — review it against the WordPress security checklist, and check which user roles can see transcripts, since those often contain personal details.
Consent and data protection. Chat transcripts are personal data, and they frequently contain more than the visitor intended to share. Establish where they are stored, how long for, whether they are used for model training, and whether the widget loads third-party scripts before the visitor has agreed — that last one is a cookie consent question and it needs answering before launch, not after.
Performance is the smaller concern but not nothing: chat widgets are heavy, third-party and usually loaded site-wide. Defer the widget until interaction, and check its effect as part of speeding up a WordPress site.
The decision
Write down your ten most common visitor questions. If the answers are fixed, build a scripted flow and spend the budget on documentation instead. If they are open-ended, choose a grounded assistant and judge it on refusal behaviour and citations rather than fluency. Then run it for a fortnight with transcripts reviewed weekly, and only then let it appear on the pages where money changes hands.
Frequently asked
- Plugins named after a model provider are usually editor tools that help you draft and rewrite content inside wp-admin. AI chatbot plugins are visitor-facing widgets that answer questions on the front end. They share an underlying API and almost nothing else.
- Ground it in your own content with retrieval, restrict it to answering from retrieved passages, and give it an explicit fallback for when nothing relevant is found. A bot allowed to answer from general model knowledge will eventually state a policy, price or delivery time that is not yours.
- You do, if the plugin calls a metered API with your key. Require rate limiting per session and per IP, a maximum conversation length, and a spend cap at the provider, before the widget goes on a public page.