diff --git a/.github/workflows/pr.yml b/.github/workflows/pr.yml
index 47a0a67..4fa2c1d 100644
--- a/.github/workflows/pr.yml
+++ b/.github/workflows/pr.yml
@@ -17,3 +17,9 @@ jobs:
       - run: bun install --frozen-lockfile
       - run: bun run check
       - run: bun test
+
+      # Social cards are assembled in the layout from post frontmatter, so the
+      # built HTML is the only place a broken one shows up. Neither biome nor
+      # tsc nor the build itself can tell that a cover is an unrenderable format.
+      - run: bun astro build --mode live
+      - run: bun scripts/check-og.ts
diff --git a/scripts/check-og.test.ts b/scripts/check-og.test.ts
new file mode 100644
index 0000000..e34a34b
--- /dev/null
+++ b/scripts/check-og.test.ts
@@ -0,0 +1,20 @@
+import { describe, expect, test } from "bun:test";
+import { dimensions } from "./check-og";
+
+describe("social card image headers", () => {
+	test("reads every allowed raster format", () => {
+		expect(dimensions("public/layout/icon.png", ".png")).toEqual({ width: 325, height: 300 });
+		expect(dimensions("public/blog/you-dont-need-it/sponge.jpg", ".jpg")).toEqual({ width: 480, height: 360 });
+		expect(dimensions("public/blog/to-wasm/duck.jpeg", ".jpeg")).toEqual({ width: 413, height: 255 });
+		expect(dimensions("public/blog/replacing-hls-dash/buffering.gif", ".gif")).toEqual({ width: 498, height: 280 });
+		expect(dimensions("public/blog/replacing-hls-dash/troll.webp", ".webp")).toEqual({ width: 217, height: 303 });
+	});
+
+	test("rejects non-image bytes with an image extension", () => {
+		expect(dimensions("package.json", ".png")).toBeUndefined();
+	});
+
+	test("rejects an image whose extension doesn't match its bytes", () => {
+		expect(dimensions("public/layout/icon.png", ".jpg")).toBeUndefined();
+	});
+});
diff --git a/scripts/check-og.ts b/scripts/check-og.ts
new file mode 100644
index 0000000..84baee7
--- /dev/null
+++ b/scripts/check-og.ts
@@ -0,0 +1,232 @@
+#!/usr/bin/env bun
+// Asserts every built page carries a social card that scrapers can actually render.
+//
+// Run in CI after `astro build`. It reads dist/, not src/, because the tags are
+// assembled in the layout from frontmatter and only the built HTML shows what a
+// scraper will really see.
+//
+// The bug that prompted this: a post shipped with `cover: ".../viewers-1000.svg"`.
+// Nothing complained -- the file existed, the path was right, the page rendered --
+// but no scraper renders SVG for a card, so the link unfurled bare. Biome and tsc
+// can't see a problem like that, and neither can a build, so a check that opens
+// the referenced image is the only thing that catches it.
+//
+// The rules below are what the major scrapers (X, Facebook, LinkedIn, Slack,
+// Discord, iMessage) agree on, so they're deliberately loose: a page fails only
+// when a card is genuinely broken, never on style.
+
+import { existsSync, readdirSync, readFileSync, statSync } from "node:fs";
+import { join } from "node:path";
+
+const DIST = "dist";
+
+// Every card needs these. og:image is the one that breaks loudly, but a missing
+// title or url unfurls just as badly.
+const REQUIRED = ["og:title", "og:description", "og:url", "og:image"];
+
+// Raster only. SVG is the trap: it's a perfectly good image everywhere else on
+// the site, and every scraper refuses it.
+const RASTER = new Set([".png", ".jpg", ".jpeg", ".gif", ".webp"]);
+
+// X's floor for summary_large_image, which is the card type the layout asks for.
+// Anything smaller silently downgrades to a thumbnail.
+const MIN_WIDTH = 300;
+const MIN_HEIGHT = 157;
+
+/** Every .html file under dist, recursively. */
+function pages(dir: string): string[] {
+	return readdirSync(dir, { withFileTypes: true }).flatMap((entry) => {
+		const path = join(dir, entry.name);
+		if (entry.isDirectory()) return pages(path);
+		return entry.name.endsWith(".html") ? [path] : [];
+	});
+}
+
+/** The content of <meta property|name="..."> tags, keyed by property. */
+function metas(html: string): Map<string, string> {
+	const found = new Map<string, string>();
+	const tag = /<meta\s+(?:property|name)="([^"]+)"\s+content="([^"]*)"/g;
+
+	for (const [, key, value] of html.matchAll(tag)) {
+		found.set(key, value);
+	}
+
+	return found;
+}
+
+/**
+ * The pixel dimensions of an image, or undefined when its bytes don't match the
+ * extension. Only the size matters here, so this validates headers rather than
+ * pulling in an image library for a check that runs on a handful of files.
+ */
+export function dimensions(path: string, ext: string): { width: number; height: number } | undefined {
+	const buf = readFileSync(path);
+
+	// PNG: a fixed IHDR chunk, width and height as big-endian u32 at byte 16.
+	if (
+		ext === ".png" &&
+		buf.length >= 24 &&
+		buf.subarray(0, 8).equals(Buffer.from([0x89, 0x50, 0x4e, 0x47, 0x0d, 0x0a, 0x1a, 0x0a])) &&
+		buf.toString("ascii", 12, 16) === "IHDR"
+	) {
+		return { width: buf.readUInt32BE(16), height: buf.readUInt32BE(20) };
+	}
+
+	// JPEG: walk the segment chain to a start-of-frame marker, which carries the
+	// dimensions. Segments are 0xFF, a marker byte, then a big-endian length.
+	if ((ext === ".jpg" || ext === ".jpeg") && buf[0] === 0xff && buf[1] === 0xd8) {
+		let at = 2;
+		while (at < buf.length) {
+			while (buf[at] === 0xff) at++;
+			const marker = buf[at++];
+			if (marker === undefined || marker === 0xd9 || marker === 0xda) break;
+			if (marker === 0x01 || (marker >= 0xd0 && marker <= 0xd7)) continue;
+			if (at + 2 > buf.length) break;
+
+			const length = buf.readUInt16BE(at);
+			if (length < 2 || at + length > buf.length) break;
+
+			// SOF0-SOF15, excluding the DHT/JPG/DAC markers interleaved among them.
+			if (marker >= 0xc0 && marker <= 0xcf && marker !== 0xc4 && marker !== 0xc8 && marker !== 0xcc) {
+				if (length < 7) break;
+				return { height: buf.readUInt16BE(at + 3), width: buf.readUInt16BE(at + 5) };
+			}
+
+			at += length;
+		}
+	}
+
+	// GIF: the logical screen width and height follow the GIF87a/GIF89a header.
+	if (ext === ".gif" && buf.length >= 10 && ["GIF87a", "GIF89a"].includes(buf.toString("ascii", 0, 6))) {
+		return { width: buf.readUInt16LE(6), height: buf.readUInt16LE(8) };
+	}
+
+	// WebP: validate the RIFF container and read dimensions from the payload used
+	// by its lossy, lossless, or extended encoding.
+	if (
+		ext === ".webp" &&
+		buf.length >= 20 &&
+		buf.toString("ascii", 0, 4) === "RIFF" &&
+		buf.toString("ascii", 8, 12) === "WEBP"
+	) {
+		const chunk = buf.toString("ascii", 12, 16);
+		if (chunk === "VP8 " && buf.length >= 30 && buf.subarray(23, 26).equals(Buffer.from([0x9d, 0x01, 0x2a]))) {
+			return { width: buf.readUInt16LE(26) & 0x3fff, height: buf.readUInt16LE(28) & 0x3fff };
+		}
+		if (chunk === "VP8L" && buf.length >= 25 && buf[20] === 0x2f) {
+			return {
+				width: 1 + buf[21] + ((buf[22] & 0x3f) << 8),
+				height: 1 + ((buf[22] & 0xc0) >> 6) + (buf[23] << 2) + ((buf[24] & 0x0f) << 10),
+			};
+		}
+		if (chunk === "VP8X" && buf.length >= 30) {
+			return { width: 1 + buf.readUIntLE(24, 3), height: 1 + buf.readUIntLE(27, 3) };
+		}
+	}
+
+	return undefined;
+}
+
+/**
+ * What's wrong with one image reference. Errors mean the card is broken and the
+ * image won't appear at all; warnings mean it still renders, just worse. Only
+ * errors fail the build, because a small cover is a judgement call the author
+ * gets to make and an unrenderable one isn't.
+ */
+function checkImage(key: string, value: string): { errors: string[]; warnings: string[] } {
+	const none = { errors: [], warnings: [] };
+
+	let url: URL;
+	try {
+		url = new URL(value);
+	} catch {
+		// Relative paths resolve against the scraper's idea of the page, which is
+		// often nothing at all. They have to be absolute.
+		return { ...none, errors: [`${key} is not an absolute URL: ${value}`] };
+	}
+
+	if (url.protocol !== "https:") {
+		return { ...none, errors: [`${key} is not https: ${value}`] };
+	}
+
+	const ext = (url.pathname.match(/\.[^./]+$/)?.[0] ?? "").toLowerCase();
+	if (!RASTER.has(ext)) {
+		return { ...none, errors: [`${key} is ${ext || "extensionless"}, which scrapers won't render: ${value}`] };
+	}
+
+	// Every image the sites reference is served from the same build, so the
+	// pathname doubles as its location in dist. An off-origin image would need
+	// fetching instead, and none exists yet.
+	const path = join(DIST, url.pathname);
+	if (!existsSync(path) || !statSync(path).isFile()) {
+		return { ...none, errors: [`${key} has no file at ${path}: ${value}`] };
+	}
+
+	const size = dimensions(path, ext);
+	if (!size || size.width === 0 || size.height === 0) {
+		return { ...none, errors: [`${key} is not a valid ${ext.slice(1).toUpperCase()} file: ${value}`] };
+	}
+
+	if (size.width < MIN_WIDTH || size.height < MIN_HEIGHT) {
+		return {
+			...none,
+			warnings: [
+				`${key} is ${size.width}x${size.height}, under the ${MIN_WIDTH}x${MIN_HEIGHT} needed for a large card, so it'll show as a thumbnail: ${value}`,
+			],
+		};
+	}
+
+	return none;
+}
+
+function main(): number {
+	if (!existsSync(DIST)) {
+		console.error(`no ${DIST}/ -- run \`astro build\` first`);
+		return 1;
+	}
+
+	const broken: string[] = [];
+	const degraded: string[] = [];
+	const files = pages(DIST);
+
+	for (const file of files) {
+		const found = metas(readFileSync(file, "utf8"));
+		const errors: string[] = [];
+		const warnings: string[] = [];
+
+		for (const key of REQUIRED) {
+			if (!found.get(key)) errors.push(`missing ${key}`);
+		}
+
+		// twitter:image is optional -- without it a card falls back to og:image, which
+		// is fine. It just has to be valid when a page does set it.
+		for (const key of ["og:image", "twitter:image"]) {
+			const value = found.get(key);
+			if (!value) continue;
+
+			const result = checkImage(key, value);
+			errors.push(...result.errors);
+			warnings.push(...result.warnings);
+		}
+
+		const list = (problems: string[]) => `${file}\n${problems.map((p) => `  ${p}`).join("\n")}`;
+		if (errors.length) broken.push(list(errors));
+		if (warnings.length) degraded.push(list(warnings));
+	}
+
+	if (degraded.length) {
+		console.warn(`Social cards that render smaller than they could, in ${degraded.length} of ${files.length} pages:\n`);
+		console.warn(`${degraded.join("\n\n")}\n`);
+	}
+
+	if (broken.length) {
+		console.error(`Broken social cards in ${broken.length} of ${files.length} pages:\n`);
+		console.error(`${broken.join("\n\n")}\n`);
+		return 1;
+	}
+
+	console.log(`Social cards OK across ${files.length} pages.`);
+	return 0;
+}
+
+if (import.meta.main) process.exit(main());
diff --git a/src/layouts/global.astro b/src/layouts/global.astro
index 878eaee..4cd274d 100644
--- a/src/layouts/global.astro
+++ b/src/layouts/global.astro
@@ -62,9 +62,12 @@ const proUrl = import.meta.env.MODE === "staging" ? "https://moq.wtf" : "https:/
 		<meta property="og:description" content={fullDescription} />
 		<meta property="og:url" content={pageUrl} />
 		<meta property="og:site_name" content="Media over QUIC" />
+		<!-- No og:image:width/height: covers live in public/ at whatever size they
+		were drawn, so any hardcoded pair is a lie for most pages. Scrapers measure
+		the image themselves when the dimensions are absent; when they're present and
+		wrong they're believed, and a card claiming to be smaller than Twitter's
+		300x157 minimum gets downgraded to a thumbnail. -->
 		<meta property="og:image" content={ogImage} />
-		<meta property="og:image:width" content="163" />
-		<meta property="og:image:height" content="150" />
 
 		<!-- Twitter Card -->
 		<meta name="twitter:card" content="summary_large_image" />
diff --git a/src/pages/blog/open-for-business.mdx b/src/pages/blog/open-for-business.mdx
index 6899383..4e96402 100644
--- a/src/pages/blog/open-for-business.mdx
+++ b/src/pages/blog/open-for-business.mdx
@@ -3,7 +3,7 @@ layout: "@/layouts/global.astro"
 title: "Open For Business"
 author: kixelated
 description: I have officially achieved the rank of "Capitalist Scum".
-cover: "/blog/open-for-business/viewers-1000.svg"
+cover: "/blog/open-for-business/splash.png"
 date: 2026-08-17
 ---