Comparison
Field datavsPercentile latency (p50, p95, p99)
Field data
the dashboard shows the metric at the 75th percentile across everyone who actually loaded the page last month.
Measurements collected from real visits on real devices and networks. It is the only source that tells you what your users experience, and it is what Core Web Vitals are assessed against. It is also slow to react and hard to attribute — a regression appears days later as a shifted distribution, with no profile attached to explain it, which is why lab and field data are used together rather than as alternatives.
Full entry →Percentile latency (p50, p95, p99)
the average looked fine and one user in twenty was waiting eleven seconds, because output length varies enormously.
Latency distributions, not averages: p95 is what a twentieth of your users feel. LLM latency is high-variance (output length varies), so percentiles and timeouts matter more here than in typical services.
Full entry →