dummyforms.com

Lovable what actually works on this platform

Lovable apps run as Cloudflare Workers, and the platform rewrites Cache-Control after your code returns. The usual caching advice does not apply — these are the fixes that do.

• Served from the platform edge with no-cache on the HTML — the expected fingerprint, and not something a header change can fix.
• A crawler reads about 46 characters of text from your HTML — thin enough that search engines and AI assistants see close to a blank page.
• Time to first byte measured 292ms on this run — one sample, so treat it as a hypothesis and re-run six times before acting.

1. Measure first, six runs, take the median. A cold worker isolate inflates the first request on almost every Lovable site — often 4-5x the median. One request, or an average that includes run 1, invents a problem that is not there.
2. Do not try to fix this with Cache-Control headers. Lovable runs apps as dynamically-loaded Cloudflare Workers. The hosting layer rewrites Cache-Control after your worker returns, and the Cache API throws. Any header-level fix is discarded — the win has to come from inside the application.
3. Find out which stack the project is. TanStack Start (server-rendered, has src/routes/) and Vite React SPA (client-rendered, has src/App.tsx) have different problems and different fixes. This cannot be told from the outside — the response headers are byte-identical — so check the repo before changing anything.
4. Give crawlers something to read. Put each route's real content in the initial HTML, route-appropriate rather than a copy of the homepage. Emitting it into the shared shell instead makes every other route pay for content only one page needs, which shows up as every route converging on one identical time.
5. If TanStack Start: cache the render in two tiers. An in-isolate LRU alone hits about 1 request in 6, because traffic spreads across isolates that do not share memory. Put a shared store (KV, or a table keyed on URL) behind it and the hit rate goes to roughly 7-8 in 8. Emit an x-app-cache: HIT-MEMORY|HIT-SHARED|MISS header while you measure — a HIT that is still slow proves the cost is cold start, not the cache.
6. If Vite SPA: the bundle is the cost, not the server. TTFB is usually fine; what feels slow is downloading, parsing and executing JavaScript before anything appears. Split routes with lazy imports, prefetch the chunk on hover, and measure the JS transferred before first paint rather than TTFB.
7. Never cache an authenticated route. Serving one visitor a page rendered for another is far worse than a slow site. Keep /admin, /profile and anything carrying credentials off the cache, and verify a logged-out fetch renders the signed-out view before you enable public caching.
8. Press Publish. Lovable syncs code to git automatically but does NOT deploy. Production keeps serving the previous build until someone presses Publish — this is the most common reason a real fix looks like it did nothing.

The full instruction set, and a machine-readable copy of this report, live at dummyforms.com.md — paste that URL into your AI builder's chat and it can work straight from these measurements.