SlashKeys
Game key price comparison where the hard part is knowing which prices to trust.
- Type
- Product
- Industry
- Gaming / price comparison
- Platform
- Web application
- Role
- Design, build, deploy
What made this hard.
A comparison engine for PC game keys: it aggregates offers from key sellers and official stores, normalizes seller and edition data, and routes buyers through tracked outbound links when affiliate approvals are in place.
Game key sellers change page structures, regional availability, price display rules and anti-bot behaviour constantly. A comparison site that shows a price it can no longer verify is worse than one that shows nothing.
So the real problem was never scraping. It was representing confidence honestly: distinguishing “this offer is unavailable” from “we have never successfully read this offer” from “we read it and the result failed validation”, and making monetisation work without letting affiliate incentives bend what a user sees.
What we built.
- Split one overloaded stock integer, which had quietly come to mean three different things, into an explicit three-state availability signal (available / unavailable / unknown), migrated across every write site with a disciplined expand-contract sequence.
- Moved offer suppression onto the signals that actually carry confidence, so a seller defending itself never reads as a genuine price change.
- Fixed a starvation bug in the URL-resolution pass: failed attempts wrote nothing back, so every scheduled run re-selected the same head rows forever. A per-offer attempt timestamp plus cohort rotation and typed per-class outcome reporting cleared it.
- Added per-seller circuit breakers with typed failure classification (blocked, no-HTML, unsupported, transient) so “their site is defending itself” and “our data is wrong” never get confused.
- Modelled the domain explicitly in the repository: a glossary and numbered architecture decision records fixing terms like displayable and purchasable offer scopes, reference offer, grace margin, and resolution attempt versus discovery cooldown.
- Kept affiliate metadata per seller (application status, network, URL template) with an outbound click service that falls back to the canonical seller URL until an approval actually lands.
What it runs on.
Chosen to be maintainable for years, not to win an argument on the internet.
- PHP 8
- Laravel
- Blade
- Alpine.js
- Vite
- Tailwind CSS
- MySQL
- Playwright
- Apache MPM event
Where it landed.
Feature tests cover each artisan command with container-bound fakes, and contract tests assert that the SQL scope and the in-memory scope agree, a pair that caught a real divergence where a product card had re-implemented the display rule inline instead of delegating to the shared scope.
Front-end performance was engineered rather than guessed at: long-lived immutable cache headers on hashed build assets, automated WebP twins served by content negotiation with the URL unchanged, self-hosted variable fonts in place of a render-blocking font import, LCP preloads, and an x-cloak fix that removed pre-boot layout shift. Apache moved from prefork with mod_php to MPM event so HTTP/2 could be enabled.
Need something in this territory? This is Custom web app development work. Start from all services or see pricing.
Related case studies.
Your project
Tell us what you're building. We reply within two business days.