Performance When AI Features Are Added

Do not load the model on every page

A marketing site should not download a chat runtime in the head. If a feature exists, its script loads on the route that needs it, after the content.

Most of our AI use never becomes frontend JavaScript at all.

Precompute what you can

Summaries, meta descriptions, and alternate headings that survive review can be baked into the static build. That is faster than asking a model on each visit.

Request-time calls are reserved for input the build could not know.

Measure the feature, not the demo

A local demo on a fast connection hides the cost. We look at the page without the feature first, then with it, on a modest phone profile.

If the feature pushes the experience backward, it waits.