Keep the model work off the page load
The common mistake is a plugin that makes a model call while a visitor waits. It is slow, it is expensive, and it fails when the provider is busy.
Generate and store, then serve what is stored. Live calls only where a visitor is genuinely asking a question.
What works well
- Support answering from your published content and documentation
- Content operations — briefs, drafts, internal linking suggestions
- Form and enquiry handling — classification and routing
- Metadata generation for pages missing titles and descriptions
- Internal search that understands meaning rather than matching words
Internal search is the quiet win. Most WordPress site search is poor, and visitors who cannot find something simply leave.
Architecture that stays fast
- AI processing in a separate service, not in PHP on request
- Results stored in post meta or a custom table
- Widgets loaded lazily, after the main content
- Aggressive caching, invalidated on content change
- No blocking external call in the critical path
The REST API is the connection point
WordPress exposes posts, pages, media, taxonomies and custom types over REST with proper authentication. That is enough for almost any integration.
Use application passwords or a dedicated key rather than a user's own credentials, and scope the account to what it needs.
Plugin or custom?
Off-the-shelf plugins are fine for simple content assistance. Anything that touches your business process, your data or your customers is worth building properly.
Also check what any AI plugin sends where. Several send page content to services under terms you may not have read.