SA tiles & bathroom retailer · Tiles, bathroom & flooring retailTwo tests that only worked on a cold cache

• Both experiments targeted product pages by reading the SKU out of the buy-box form inside a Convert location condition.
• QA saw the variants on a fresh load and then never again, a classic 'works once' report that usually gets blamed on the variant code.
• The Convert snippet loads asynchronously in the head, while the buy-box form sits roughly 2,000 lines into a 425 KB document.

The experiments were not failing, they were never starting. If activation waited for the form instead of racing it, both tests would bucket every eligible visitor.

• Traced the timing: with Convert's JavaScript cached for five minutes, the location condition executed before the form was in the DOM, evaluated false, and the experience never activated for that pageview.
• Moved activation into a shared, status-gated Global JS function: wait for the buy-box form, then the page title, run the product-title regexes, set a window flag and push the activation event.
• Selector test: two-option control (single item vs full set) above add-to-cart that rewrites the native form's product ID and swaps the headline price, with a stylesheet keyed on an html class to suppress conflicting promos.
• Panel test: page targeting by title patterns that accept fixtures and reject tiles, flooring and fitting phrases.
• Re-initialise on bfcache restore so the Back button gets a clean state.

Measured on add-to-cart, set chosen, atc from selector. Test results belong to the client and are shared in conversation, with their permission, rather than published here.