If you may keep only one CDN metric, keep this one. Hit ratio predicts user latency and origin cost at the same time, which makes it the rare number both engineering and finance should watch.
Why it moves everything
Every hit is served from the edge: fast for the user, free for your origin. Every miss is a trip home: slower, and billed as origin egress and compute. Moving hit ratio from 85% to 95% cuts misses by two thirds, which is exactly how it feels on both the latency chart and the cloud bill. It is also the metric most likely to be misread as fine, because dashboards report it blended across content types, where one well-cached asset class can hide a badly leaking one.
Why yours is lower than you think
Fragmented cache keys, query strings that vary needlessly, cookies that bust caching by accident, short TTLs left over from launch caution, and long-tail catalogs that no single POP sees often enough. Most of these are configuration debts, not architecture problems. The configuration-debt framing matters for morale as much as accuracy: nobody needs a migration or a budget line, just a week of unglamorous header hygiene that has probably been postponed since launch.
One number for intuition: at 100 TB monthly and typical origin egress rates, each percentage point of hit ratio moves roughly a terabyte of traffic between free edge and billed origin. Ten points of improvement, entirely typical for a first serious tuning pass, reprices ten terabytes a month without touching a contract. The improvement also compounds with growth: the tuned configuration keeps paying at every future volume, which makes cache hygiene one of the few infrastructure investments whose return rises as the business scales rather than eroding.
Raising it
Normalize cache keys, lengthen TTLs where correctness genuinely allows, enable origin shielding, and cache the almost-cacheable with short freshness plus revalidation. Unglamorous work with a measurable payback period, usually weeks. Origin shielding deserves its own sentence: routing all misses through one designated edge layer means your origin sees each object roughly once per freshness window instead of once per POP, which multiplies effective cachability for long-tail catalogs.
In practice
Pull hit ratio split by content type, not blended. Fix the worst class first: normalize its cache key, strip the query parameters that do not change the response, and double its TTL if correctness allows. Re-measure in a fortnight. Repeat quarterly. Teams that adopt this loop typically find the first pass alone repays the year’s attention, visible simultaneously in the latency chart and the origin bill.
A cache audit is part of every assessment, because it is the cheapest performance and cost win in the building.