Case studyCommerce engineering · GEO
Hi-Fi Storefront Overhaul
A technical, performance and AI-search overhaul of a live Shopify storefront, from an unversioned theme to a reviewed deploy pipeline and a purpose-built generative-search layer, with zero visual change for customers.
Client not named publicly. Everything below is accurate; only the attribution is withheld. Happy to walk through the work in detail, and introduce you to the client, on request.
- Shopify
- Liquid
- JSON-LD
- Git / GitHub
- Core Web Vitals
- pull requests merged
- 10pull requests merged
- modules preloaded
- 85 → 14modules preloaded
- real bugs caught pre-impact
- 2real bugs caught pre-impact
The storefront was live, profitable and completely unversioned — the way most small commerce sites actually exist. Any change was made directly against production with no history and no review. The first Hart & Heim deliverable was not a fix; it was the ability to make fixes safely.
Version control before anything else
The theme was pulled byte-for-byte into a private repository as its first commit and connected to the platform's integration, so merging to the main branch deploys and every change from that point forward has an author, a diff and a reviewer.
- Live theme captured exactly as it ran, with no cleanup pass that could hide a regression
- Branch → pull request → review → merge as the standing workflow
- Bidirectional sync respected: changes made in the visual editor come back as commits rather than being silently overwritten
Ten changes, none of them cosmetic
- Meta description fallback closing gaps across templates
- A case-sensitive tag bug that was silently excluding products from collections
- Breadcrumb and item-list structured data for search result context
- Hero image priority hints and lazy video loading for real-world load performance
- Module preloading traced to the actual dependency graph and cut from 85 to 14
- Legacy sections migrated to the current architecture, and organisation address data corrected
Built to be cited, not just crawled
Search increasingly means an answer engine reading a page and quoting it. That requires structured, unambiguous, machine-readable content — which is a different job from ranking.
- A site-specific llms.txt written on top of the platform-generated boilerplate, verified by byte-diff against the default
- FAQ schema added across the core product range, hitting and working within the platform's section limits
- A new buyer comparison guide published with article and FAQ structured data, linked from the footer
Verification as a deliverable
Every structured-data claim was checked by fetching the live page and parsing it, not by reading the template and assuming the output. That process paid for itself immediately.
- Two real bugs caught by parsing rather than eyeballing — one live for under an hour, one that never left pre-release testing
- Two findings from the original audit disproved the same way, avoiding fixes for problems that did not exist
- Preview environments torn down after each verification pass rather than left accumulating