A CDN cache hit ratio is hits / (hits + misses) at the edge (sometimes layered: POP vs regional shield). Product people treat 95% as a grade. Engineers treat it as a function of cache key, TTL, and variance. If every URL is unique (?session=), the ratio is near zero and origin dies. If you cache Set-Cookie personalized HTML as public, you leak users. The metric is only useful with a defined key.
Keys are the product
Default keys often include the full query string. A tracking param (utm_source) shards the cache. Normalize: ignore known params, sort the rest, or drop query on static assets. Vary: Accept-Encoding is correct; Vary: Cookie on a page that sets a session cookie is "do not cache." Separate static (long TTL, immutable hashed names) from HTML (short TTL or cache at the app with surrogate keys).
bad key: /app.js?cb=Math.random()
good key: /app.7f3a2c.js Cache-Control: public, max-age=31536000, immutableCookies on static hosts: don't. Move assets to a cookieless domain or strip cookies at the CDN. Authorization headers usually force miss unless you built a private cache design (signed URLs, not public POPs).
TTLs, stale, and origin shield
Short TTL + high QPS still helps if revalidation is cheap (ETag). stale-while-revalidate keeps the POP serving while one request goes to origin. An origin shield (regional cache) collapses thundering herds. Purging: wildcard purge of /* after every deploy is how you DDoS yourself; hashed filenames avoid purge.
Hit ratio by status and path prefix beats a global average. APIs at /v1/ might correctly be 10%; /static/ should be 99%. A blended 70% hides a broken static config.
Measurement
Use the CDN's analytics, not curl once. Compare byte hit ratio versus request hit ratio (large video vs tiny CSS). A miss that 304s is not an origin render.
Read RFC 9111 for Cache-Control and Vary. Then dump the top 20 miss URLs. If they differ only by junk query params, you do not have a capacity problem. You have a key problem, and the hit ratio is telling the truth.