Aggregator of GPU prices across # GPU-AG · VRAMWATCH A market board for gaming and AI silicon. Live street prices pulled from real retailer and reseller search pages, tracked over time, with a chart and direct buy links for every part — RTX 5090 to H200. ## How it works For the full walkthrough — the sweep, the filter gate, the store, and how the crawler surface is generated — see **[docs/HOW-IT-WORKS.md](docs/HOW-IT-WORKS.md)**. - **Live sources (fetched server-side, verified working):** Newegg, Central Computer, Wiredzone (Supermicro-authorized — carries H100/A100/L40S), PC Server & Parts, TechMikeNY and Server Part Deals (refurb specialists, read through Shopify's structured product JSON — the secondary-market signal for V100/A6000-class cards), and the **Supermicro store**, the manufacturer selling direct, which is what lets the board price whole GPU systems. - **Official APIs (free keys, optional):** eBay [Browse API](https://developer.ebay.com) for the used market (A100s, 4090s, SXM modules, gray-market parts) and the [Best Buy Products API](https://developer.bestbuy.com) for consumer cards. - **Link-only sources:** Micro Center, B&H and Amazon bot-block server requests, so they get deep search links instead of fetched prices. Micro Center 403s a server on every path including `robots.txt`; the only ways around that are UA spoofing or proxy evasion, so it stays link-only. (Also probed and blocked: Walmart, CDW, Insight, Connection, Server Supply, Provantage, IT Creations, Server Orbit, antonline; SabrePC renders client-side.) - Every refresh accumulates into `data/history.json` as `{lo, hi, n, at}` per (GPU, source, day), so `lo` is the day's **true** low across every sweep — **charts grow real history as the tracker runs**. Matched listings with URLs land in `data/listings.json` and power the "where to buy" cards. Listings are filtered per part by title regex (`match` / `exclude`) and a price sanity band (see `lib/gpus.ts`) so waterblocks, server barebones and GH200 superchips don't pollute the series. Retail series track new-condition prices; used/refurb belongs to eBay (EOL parts like the RTX 4090 are exempt via `usedOk`). ## Run it ```sh pnpm install pnpm dev # http://localhost:3000 ``` ## Production ```sh pnpm build && pnpm start ``` The site keeps itself fresh: the UI polls `/api/tick` every 10 seconds, and the server re-sweeps all sources in the background whenever the stored data is older than the sweep interval (default 10 minutes — tune with `SWEEP_INTERVAL_MS`, floored at 60s to stay polite to the sources). No cron required; `pnpm refresh` still exists for warming data ahead of a deploy. Note: the store is process-local JSON (`data/*.json`) — run a single instance, or swap `lib/store.ts` for a shared database before scaling out. ### API keys (optional but recommended) ```sh # .env.local EBAY_CLIENT_ID=... # developer.ebay.com EBAY_CLIENT_SECRET=... BESTBUY_API_KEY=... # developer.bestbuy.com ``` ### Canonical origin Every canonical tag, `sitemap.xml` entry and absolute OG image URL is built from one value: ```sh NEXT_PUBLIC_SITE_URL=https://vram.swarms.world ``` It defaults to `https://vram.swarms.world` (and picks up `VERCEL_PROJECT_PRODUCTION_URL` on Vercel), so set it only when deploying somewhere else — a wrong origin here silently poisons canonicals. ## SEO Everything a crawler consumes is generated, never hand-maintained: - `lib/site.ts` — canonical origin, brand, boilerplate copy, keyword set. - `lib/seo.ts` — JSON-LD builders. Home ships `CollectionPage` + `ItemList` (all tracked parts, with live prices) + `FAQPage`; every GPU page ships a `Product` with an `AggregateOffer` built from that part's real listings (per-seller `Offer`s, condition tagged from the source) plus a `BreadcrumbList`. `Organization` + `WebSite` sit in the root layout. - Per-GPU `generateMetadata` puts the live price in the title and meta description (`GeForce RTX 5090 Price — $4,400 (Live, Central Computer)`), so the SERP snippet is current rather than boilerplate. - `app/opengraph-image.tsx` and `app/gpu/[slug]/opengraph-image.tsx` render 1200×630 social cards with `next/og`; the per-GPU card shows the current best price and 24h move and revalidates every 10 minutes. - `app/icon.svg` (the green box), `app/apple-icon.tsx`, `app/manifest.ts`, `app/robots.ts`, `app/sitemap.ts` (hourly, all 23 parts), `app/not-found.tsx`. - The FAQ copy on the home page and its `FAQPage` markup are rendered from the same `FAQ` array — Google requires the two to match, so they can't drift. ## Architecture ``` lib/gpus.ts tracked parts: query, match/exclude regex, price band lib/adapters/shopify.ts one factory covering any Shopify storefront lib/sources.ts source registry (mode: scrape | api | link, chart color) lib/adapters/*.ts one fetcher per source — the seam to add more lib/refresh.ts sweep: fetch → filter → snapshot lib/store.ts data/*.json persistence (swap for a DB here) lib/quotes.ts best price, 24h/7d/30d deltas, spark series lib/site.ts canonical origin + brand/boilerplate copy lib/seo.ts JSON-LD builders (Product, ItemList, FAQ, breadcrumbs) app/ dashboard, /gpu/[slug], /sources, /api/refresh app/*-image.tsx generated OG cards, favicon, apple icon app/{robots,sitemap,manifest}.ts crawler surface ``` To add a source: write an adapter returning `{title, price, url}[]`, register it in `lib/adapters/index.ts` and `lib/sources.ts`, and add its id to the parts that carry it. Source: https://github.com/kyegomez/vram-watch