diff --git a/package.json b/package.json
index f851fd1..0cf0db1 100644
--- a/package.json
+++ b/package.json
@@ -22,7 +22,8 @@
     "validate-links": "tsx scripts/validate-links.ts",
     "validate-redirects": "tsx scripts/validate-v2-redirects.ts",
     "validate-content": "tsx scripts/validate-content-api.ts",
-    "validate-mermaid": "tsx scripts/validate-mermaid.ts"
+    "validate-mermaid": "tsx scripts/validate-mermaid.ts",
+    "triage-404s": "tsx scripts/triage-404s.ts"
   },
   "dependencies": {
     "fumadocs-core": "16.6.0",
diff --git a/scripts/triage-404s.ts b/scripts/triage-404s.ts
new file mode 100644
index 0000000..76d324a
--- /dev/null
+++ b/scripts/triage-404s.ts
@@ -0,0 +1,208 @@
+#!/usr/bin/env tsx
+/**
+ * 404 triage for a Vercel log export.
+ *
+ * The raw export is unreadable: in the 2026-08-26 → 08-31 window, 4,771 of
+ * 5,000 rows were Next.js client segment-cache prefetches, so the ~230
+ * failures worth acting on sat under 95% noise. This classifies the rows,
+ * drops the noise, and prints what is left with requester attribution.
+ *
+ * Usage: bun run triage-404s <export.csv>
+ * Export from the Vercel dashboard: Logs → filter status 404 → Export CSV.
+ *
+ * Not wired into prebuild — it reads a file that only exists when someone
+ * has pulled an export.
+ */
+import fs from "node:fs";
+
+interface Row {
+  requestId: string;
+  path: string;
+  status: string;
+  userAgent: string;
+  time: string;
+}
+
+/**
+ * Requests the Next.js client segment cache makes on hover/viewport. They
+ * 404 because segment data is only generated when `cacheComponents` is on;
+ * the router falls back to a normal prefetch, so nothing is broken for the
+ * reader. Benign, high volume, and the reason raw 404 reports are useless.
+ */
+const SEGMENT_PREFETCH = /\.segments\/|\.rsc$/;
+const ASSET = /\.(?:js|css|png|jpe?g|gif|webp|svg|ico|woff2?|map)$/;
+/** `file.ts:44` in docs prose, resolved as a link by an agent. */
+const SOURCE_REF = /\.[cm]?[jt]sx?(?::\d+)?$/;
+const MACHINE_ENDPOINT =
+  /\/(?:mcp|robots\.txt|openapi\.json|api\.json|\.well-known\/[^/]+)$/;
+
+type Bucket =
+  | "segment-prefetch"
+  | "source-ref"
+  | "markdown-suffix"
+  | "machine-endpoint"
+  | "asset"
+  | "page";
+
+function classify(path: string): Bucket {
+  if (SEGMENT_PREFETCH.test(path)) return "segment-prefetch";
+  if (/\.mdx?$/.test(path)) return "markdown-suffix";
+  if (SOURCE_REF.test(path)) return "source-ref";
+  if (MACHINE_ENDPOINT.test(path)) return "machine-endpoint";
+  if (ASSET.test(path)) return "asset";
+  return "page";
+}
+
+// Self-identified AI agents and scripted clients. A browser user-agent is
+// not proof of a human: agents that do not identify themselves land in
+// "browser", and bursts of sequential machine-shaped requests show up there.
+const AI_AGENTS =
+  /GPTBot|OAI-SearchBot|ChatGPT-User|ClaudeBot|Claude-User|Claude-SearchBot|anthropic-ai|PerplexityBot|Perplexity-User|DuckAssist|Amazonbot|Applebot|Bytespider|Google-Extended|meta-external|cohere|curl\/|python-requests|node-fetch|Go-http-client|axios|undici/i;
+const SEARCH_CRAWLERS = /Googlebot|bingbot|DuckDuckBot|YandexBot/i;
+
+type Requester = "ai" | "search" | "bot" | "browser";
+
+function requester(userAgent: string): Requester {
+  if (AI_AGENTS.test(userAgent)) return "ai";
+  if (SEARCH_CRAWLERS.test(userAgent)) return "search";
+  if (/bot|spider|crawl/i.test(userAgent)) return "bot";
+  return "browser";
+}
+
+/** Minimal RFC 4180 reader: the export quotes user-agent strings, which contain commas. */
+function parseCsv(text: string): string[][] {
+  const rows: string[][] = [];
+  let row: string[] = [];
+  let field = "";
+  let quoted = false;
+  for (let i = 0; i < text.length; i++) {
+    const char = text[i];
+    if (quoted) {
+      if (char === '"') {
+        if (text[i + 1] === '"') {
+          field += '"';
+          i++;
+        } else quoted = false;
+      } else field += char;
+      continue;
+    }
+    if (char === '"') quoted = true;
+    else if (char === ",") {
+      row.push(field);
+      field = "";
+    } else if (char === "\n") {
+      row.push(field);
+      rows.push(row);
+      row = [];
+      field = "";
+    } else if (char !== "\r") field += char;
+  }
+  if (field || row.length > 0) {
+    row.push(field);
+    rows.push(row);
+  }
+  return rows;
+}
+
+const file = process.argv[2];
+if (!file) {
+  console.error("usage: bun run triage-404s <export.csv>");
+  process.exit(1);
+}
+
+const table = parseCsv(fs.readFileSync(file, "utf8"));
+const header = table[0];
+const index = (name: string) => header.indexOf(name);
+const columns = {
+  requestId: index("requestId"),
+  path: index("requestPath"),
+  status: index("responseStatusCode"),
+  userAgent: index("requestUserAgent"),
+  time: index("TimeUTC"),
+};
+for (const [name, position] of Object.entries(columns)) {
+  if (position === -1) {
+    console.error(`✗ export is missing the ${name} column`);
+    process.exit(1);
+  }
+}
+
+// The export repeats rows; requestId is the only stable key.
+const seen = new Set<string>();
+const rows: Row[] = [];
+for (const cells of table.slice(1)) {
+  if (cells.length < header.length) continue;
+  const requestId = cells[columns.requestId];
+  if (seen.has(requestId)) continue;
+  seen.add(requestId);
+  // requestPath carries the host: "public-docs.vercel.app/docs/x?y=z".
+  const raw = cells[columns.path];
+  const slash = raw.indexOf("/");
+  rows.push({
+    requestId,
+    path: (slash > 0 ? raw.slice(slash) : raw).split("?")[0],
+    status: cells[columns.status],
+    userAgent: cells[columns.userAgent],
+    time: cells[columns.time],
+  });
+}
+
+const duplicates = table.length - 1 - rows.length;
+const buckets = new Map<Bucket, Row[]>();
+for (const row of rows) {
+  const bucket = classify(row.path);
+  const list = buckets.get(bucket);
+  if (list) list.push(row);
+  else buckets.set(bucket, [row]);
+}
+
+const times = rows.map((r) => r.time).sort();
+console.log(
+  `\n${rows.length} unique requests (${duplicates} duplicate rows dropped)`,
+);
+console.log(`window: ${times[0]} → ${times[times.length - 1]}\n`);
+
+console.log("bucket             requests");
+for (const [bucket, list] of [...buckets].sort(
+  (a, b) => b[1].length - a[1].length,
+)) {
+  const note =
+    bucket === "segment-prefetch" ? "  (benign Next.js prefetch — ignore)" : "";
+  console.log(
+    `  ${bucket.padEnd(17)} ${String(list.length).padStart(5)}${note}`,
+  );
+}
+
+const actionable = rows.filter((r) => classify(r.path) !== "segment-prefetch");
+const byPath = new Map<string, Row[]>();
+for (const row of actionable) {
+  const list = byPath.get(row.path);
+  if (list) list.push(row);
+  else byPath.set(row.path, [row]);
+}
+
+console.log(
+  `\n${byPath.size} distinct paths worth triaging, ${actionable.length} requests:\n`,
+);
+console.log("hits  ai search bot browser  path");
+for (const [path, list] of [...byPath].sort(
+  (a, b) => b[1].length - a[1].length,
+)) {
+  const counts: Record<Requester, number> = {
+    ai: 0,
+    search: 0,
+    bot: 0,
+    browser: 0,
+  };
+  for (const row of list) counts[requester(row.userAgent)]++;
+  console.log(
+    `${String(list.length).padStart(4)}  ${String(counts.ai).padStart(2)} ${String(counts.search).padStart(6)} ${String(counts.bot).padStart(3)} ${String(counts.browser).padStart(7)}  ${path}`,
+  );
+}
+
+const fromAgents = actionable.filter(
+  (r) => requester(r.userAgent) === "ai",
+).length;
+console.log(
+  `\n${fromAgents} of ${actionable.length} came from self-identified AI agents.`,
+);