Technology Stack
Next.js 15.5 (App Router), React 19.2, TypeScript, Tailwind CSS 3.4, GSAP, Lucide, react-fast-marquee, react-d3-speedometer, react-syntax-highlighter, next-sitemap, Geist Mono, Cloudflare Turnstile, Crisp, Termly, @next/bundle-analyzer — plus a production asset toolchain in Node: Replicate, sharp, d3-geo, topojson-client, world-atlas. Deployed on Vercel and on dedicated infrastructure.
Overview
Duration: 12/2025 - 8/2026
LightningProxies operates a 20M+ IP network across 250+ locations, selling residential, unlimited residential, ISP, datacenter, IPv6, and mobile proxies to teams running data collection, SERP monitoring, ad verification, and multi-account operations at scale.
The buyer is technical and always comparison-shopping, which sets two requirements at once. The platform has to be wide — a dedicated page for every product, price list, market, and integration — because this category is won on organic search for long-tail intent like "residential proxies japan" or "octobrowser proxy setup". And it has to be useful before purchase, which is why a suite of free network diagnostics runs on the same domain as the pricing pages: the tools bring qualified technical traffic in, and the product pages convert it.
Eight months of continuous delivery, with the majority of that time spent on the parts that compound — search surface, market coverage, and the attribution layer that decides whether a partner gets paid.
Platform Surface
- Products — six product lines, each on its own page, sharing one presentation layer driven by a typed per-product content model
- Pricing — a matching page per line, composed from a set of small, reusable pricing primitives so a plan change is a data edit
- Markets — 49 country pages, statically generated, each with local city coverage
- Integrations — dedicated pages for 12 anti-detect browsers, the highest-intent traffic in the category
- Use cases — SERP monitoring, web scraping, market research, ad verification, brand protection
- Free tools — proxy checker, IPv6 checker, WHOIS lookup, speed test with blacklist check, live network status
- Content — 28 in-depth technical articles, structured as a topic-cluster link graph
- Programs — affiliate, reseller, partners, brand assets, help centre, and full legal surface
Keeping product and pricing on separate URLs is deliberate: a buyer comparing three vendors wants the capability page and the price page open side by side, not hidden behind a tab in each other.
Affiliate Attribution That Can't Break the Page
Partner attribution is directly load-bearing on revenue: a visitor who arrives on a partner link has to still be recognisable as that partner's referral when they create an account minutes later, on a different subdomain. This is the most carefully engineered part of the platform.
It runs once at the application root, so no page has to opt in, and it does two things:
- Reports qualified visits to the tracking API — and only when a partner code is present, so ordinary organic traffic generates no request at all. The code is captured from the URL, persisted, and paired with a stable visitor ID
- Carries attribution across the domain boundary through a single capture-phase click listener that rewrites outbound auth links in place. Modified clicks — new tab, meta, ctrl, shift — are left alone so normal browsing behaviour is untouched
The demanding part is the Cloudflare Turnstile token the backend verifies. Tokens are single-use and short-lived, so a fresh one is minted per report, and three specific decisions are what make that reliable in production rather than intermittently broken:
- The challenge script is loaded on the platform's terms, not the framework's — the standard script helper forces attributes that Cloudflare rejects outright, which silently prevents any token from being issued
- The readiness callback is used but never trusted. It does not always fire in production, which would leave the widget permanently stalled; a guarded fallback renders it regardless, exactly once
- Every step is timeout-bounded, with a final attempt to recover a token that solved but never reached the callback
The whole chain ends in a fire-and-forget request that survives page unload. Attribution is commercially critical, and it is also invisible: if any part of it fails, the visitor sees a normally rendering page.
Free Tools on a Protected API Layer
The five diagnostics are the reason to bookmark the domain, and each is a thin client over a server route. Two deliberate shapes:
The checker endpoints validate before they forward — address and domain validation, a hard size cap on the speed test, and per-endpoint timeout budgets tuned to what each operation actually costs, from a fast DNS lookup to a multi-blacklist sweep. Upstream failures map to accurate status codes rather than being flattened into a generic error, and every response is CORS-enabled with a matching preflight handler. The full contract is documented with request and response examples.
The network status endpoint exists for a different reason: it holds a private API key server-side and serves live status to the browser without that credential ever reaching it. No caching, clean failure when unconfigured.
The speed test pairs a streamed download with a live gauge and the blacklist check in a single view, so a visitor gets throughput and reputation in one action.
Scaling Market Coverage
Forty-nine market pages each need three pieces of artwork, and seven had finished assets. Commissioning the remaining set was not viable on the timeline, and generic stock would have made the new markets look like a different site.
The answer was a production asset pipeline, and its value is in knowing exactly where automation stops:
- Landmark artwork is style-locked to the existing hand-made set through image-to-image generation against the finished reference assets, so every new market inherits the established palette, line treatment, and composition instead of drifting. Prompts are built from a hand-authored landmark table, not generic country names
- The market data card is not generated at all. It carries a real flag, country name, city, and live status — things generative models cannot render legibly — so it is produced deterministically and offline from authoritative geographic data, projected, drawn as vector, and rasterised. No API, no rate limit, no variance
- Post-processing is programmatic — background keying derived per-pixel so map artwork composites cleanly over the brand gradient instead of sitting in a white box
- Wiring is automated — the pipeline updates the data layer itself, idempotently, with dry-run and single-market flags
Every step discovers its own work by reading the data layer, so adding a market is adding a record. The pipeline documents its own output quality per asset type, which is what keeps it trustworthy for the next person who runs it.
Search Architecture
The platform is built to be found, with the machinery centralised rather than scattered across pages:
- A single article registry holds every post's title, description, publication and revision dates, keywords, and topic cluster. The cluster drives the internal link graph automatically, so each article points at the products its reader actually wants — the link structure is data, not hand-wiring
- One canonical URL helper resolves the site origin from the environment and version-stamps social images, which is how a cached link preview actually gets refreshed on platforms that never re-fetch
- Structured data as components — an Organization graph at the root, a service definition per product line, and breadcrumbs wherever a trail exists
- Static generation across every dynamic route, so all 49 market pages and all 12 integration pages are prerendered
- Automated sitemap and robots generation on every build, with private and utility routes excluded
Performance & Delivery
Four third-party vendors, four different deferral strategies, none of them blocking first paint:
- Live chat does not load until the visitor's first real interaction, with an idle fallback and every listener released after the first hit
- Ad tracking loads at lowest priority
- Video embeds are click-to-play facades over static thumbnails, so no player is downloaded until it is wanted
- Consent management ships with an explicit policy map so essential services are never deferred into breaking
Images are served as AVIF and WebP with an explicit quality ladder, static assets carry a one-year immutable cache policy, and bundle analysis is wired into the build. The platform ships on Vercel and, in parallel, to the client's own dedicated infrastructure through a scripted deployment path.
This is the largest surface in the portfolio, and the work that matters most is the work a visitor never sees: an attribution flow that fails open instead of failing the page, an API layer whose job is to hold credentials the browser must not have, and an asset pipeline that knows precisely which of its outputs automation has no business attempting.