Semantic links and the work of keeping automation visible

Internal linking is one of the most interesting parts of this project. Finding relevant connections across a whole website is repetitive work, particularly when an exact phrase is not already present in the right paragraph. I have been combining semantic matching with language models to make the process more flexible and to manage the resulting changes.

Matching the intent behind content

The matching system gained a fourth vector layer for semantic intent alongside the existing signals. The aim is to find relationships that are relevant to what a reader is looking for, not just occurrences of a prewritten keyword. Bulk tools can create or edit links around matched content, and a related-posts system uses semantic relationships to help surface other material.

The linking workflow can make small changes to text to fit an anchor naturally. I chose to write changes into the post itself rather than relying on a display-time content filter: I was concerned that a filter bug could damage the displayed content. The system tracks changes and saves recovery information so that work can be undone. This still needs observation and review; semantic similarity is a useful signal, not proof that every link is appropriate.

Seeing what the system is doing

A link-juice dashboard, readiness views and setup guidance provide more ways to inspect configuration and activity. Related-content automation gained run history and batch-failure diagnostics. The Qdrant dashboard and vector health work expose index configuration and operational checks, helping identify problems with the machinery that matching relies on.

These checks have limits. A healthy connection or populated index does not prove the quality of every generated link, the accuracy of an article or an improvement in search performance. I have not yet measured whether the newer linking approach changes site performance.

Access and reliability

The MCP overhaul was accompanied by security remediation. The changes enforce token scopes together with WordPress user capabilities, restrict administrative tools, remove token authentication in query parameters and bind vector credentials to an approved endpoint. The work also strengthened media-import handling and made failure states more visible. These are concrete controls around the integration, not a claim that the system is free of security risks.

The video builder was removed during this period. Its CPU use on a server without a GPU slowed the other sites running there, so I extracted the idea into a separate project rather than retaining it in this plugin.