How a static rebuild cut SitShape’s crawler compute from 103 Worker calls to zero

The first version rendered pages on request inside a Cloudflare Worker. People never noticed. Crawlers did: they walk every shape, every sofa and every seller page, and each visit woke the Worker and burned compute. For a marketplace that wants to be found by search engines and AI assistants, that bill grows with exactly the traffic you want more of.

Crawler sample · 45 seconds2026-10-01

pages

risk: Shape, sofa and seller pages, rendered on request

before

problem: 103 Worker invocations from crawlers in 45 seconds

cost curve

problem: More crawling meant more compute, linearly

the fix

note: Pre-render all 564 pages at deploy time, in 3 languages

after

fine: 0 Worker invocations in the same 45-second sample

Custom sofas are bought by eye and often arrive wrong, and small workshops can’t pay for photo shoots or marketplace commissions. So SitShape draws every sofa from a list of modules: no photography needed. Buyers start from 155 shapes in 8 families, fit one to their actual floor, and only then ask makers for a real price.

1. Step 1 (buyer): Design a shape, 155 shapes, 8 families.
2. Step 2 (automated): Shape Finder, fits the real room. Rejected here: Paid ranking (refused: fit ranks first).
3. Step 3 (buyer): Ask chosen makers, anonymous until then.
4. Step 4 (automated): Seller portal, magic link, no password.
5. Step 5 (maker): Maker quotes, direct to the buyer.
6. Step 6 (live): Static storefront, 564 pages, 3 languages.

1. agent://researcher

   Sized the market of independent sofa makers and how they win and lose work today.

   Never allowed to: set the price model.

2. human://vegardHUMAN GATE

   Chose the model: a flat listing fee by territory, zero commission, and a trial that ends at the first lead.

3. agent://architect

   Designed the module language that turns a list of pieces into a drawing, so no product photography is ever needed.

4. agent://builder

   Built the configurator, the Shape Finder, the vendor portal and a static build that renders all 564 pages in 3 languages.

5. agent://tester

   Runs 559 automated tests on every change, before anything ships.

   Never allowed to: mark a failing check as passed.

6. human://userHUMAN GATE

   A buyer decides which makers see the request. Until then, nobody gets their details.

7. agent://watcher

   Runs the daily housekeeping for trials and leads; the admin console shows funnel, demand, sellers and requests.

   Never allowed to: share a buyer’s details without consent.

The storefront is static files served from Cloudflare’s edge, so a crawler reading every page costs nothing in compute. The Worker only runs for the parts that need a server: buyer requests, the seller portal, payments. Stripe runs on Stripe’s own hosted pages, which keeps tracking scripts off the site and the no-cookie promise intact. One database, one daily cron, one email binding.

• The code and the buildThe configurator, the static build and the portal, in your repository.
• Your Cloudflare accountWorkers, static assets, database and email, under your name.
• Your Stripe accountPlans, trials and the customer portal, all billed to you directly.
• The product languageThe module system that draws your products, documented.
• The test suiteEvery pricing, consent and storefront rule, tested.
• The admin consoleFunnel, demand, sellers, requests and emails in one place.