Finding what is actually slow

Lighthouse is a lab test: one simulated device, one throttled connection, a cold cache, and none of the consent banners and tag managers a real visitor loads. Google evaluates field data — Core Web Vitals from real Chrome sessions at the 75th percentile. A green lab score alongside a failing INP in the field is common, not a contradiction.

Often, yes. A profiling pass on two or three routes, plus whatever field data already exists, is usually enough to produce a ranked list. Sometimes the finding is that performance is not your problem — a checkout that loses people at the address form is losing them to the form. We will tell you that rather than sell you a sprint.

You rarely need to remove them. Most third-party weight is a loading problem: scripts fetched in the head that could be deferred, a tag container carrying tags for campaigns that ended, an A/B tool blocking first paint. We measure each script's cost in milliseconds on your own pages and hand that list to whoever owns the budget for them. The conversation goes differently when the cost has a number on it.

Budgets in CI fail the build when a bundle size or a query count crosses a threshold, and field monitoring alerts when a real-user metric moves. Both are wired to a person who can act — an alert routed to a channel nobody reads is the same as no alert, and that is how the last optimization sprint got undone.

No. Some ceilings are architectural: a request path that fans out to six services, a schema that cannot answer the reporting query without a full scan, a framework doing server work on every render. Tuning buys some room and then stops. When we hit that, we say so and show the profile that proves it, and the next conversation is about modernization, not more tuning.