diff --git a/.github/workflows/site.yml b/.github/workflows/site.yml
index fbe2e86f..63512886 100644
--- a/.github/workflows/site.yml
+++ b/.github/workflows/site.yml
@@ -147,6 +147,10 @@ jobs:
       # `astro check` or the a11y pass -- see the header of the script.
       - name: Changelog dates
         run: npm run check:changelog
+      # JSON-LD renders as nothing, so a malformed breadcrumb shipped on every
+      # docs page until Search Console flagged it. See the script's header.
+      - name: Structured data
+        run: npm run check:structured-data
       # Runs axe against both themes; see site/scripts/a11y.mjs for why
       # both are needed, and why it prefers the runner's matched
       # chromedriver over the one the npm package downloads.
diff --git a/site/package.json b/site/package.json
index 5f3304c4..5cf16227 100644
--- a/site/package.json
+++ b/site/package.json
@@ -12,6 +12,7 @@
     "preview": "astro preview",
     "check": "astro check",
     "check:changelog": "node ./scripts/changelog-dates.mjs",
+    "check:structured-data": "node ./scripts/structured-data.mjs",
     "check:css": "node ./scripts/css-shadowing.mjs --max 0",
     "check:regions": "node ./scripts/css-regions.mjs",
     "css:equivalence": "node ./scripts/css-equivalence.mjs",
diff --git a/site/scripts/structured-data.mjs b/site/scripts/structured-data.mjs
new file mode 100644
index 00000000..5a0ffa03
--- /dev/null
+++ b/site/scripts/structured-data.mjs
@@ -0,0 +1,85 @@
+// Is the structured data on every built page something Google accepts?
+//
+// JSON-LD renders as nothing. A breadcrumb with a nameless, URL-less middle
+// step shipped on every docs page and was only noticed when Search Console
+// raised it as a critical issue: Missing field "item" (in "itemListElement").
+// This reads the built HTML and checks the rules that would have caught it.
+// Run after `astro build`.
+//
+//   node ./scripts/structured-data.mjs
+
+import fs from 'node:fs';
+import path from 'node:path';
+import { fileURLToPath } from 'node:url';
+
+const dist = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '..', 'dist');
+if (!fs.existsSync(dist)) {
+  console.error(`No build at ${dist}. Run \`npm run build\` first.`);
+  process.exit(1);
+}
+
+const pages = [];
+const walk = (dir) => {
+  for (const entry of fs.readdirSync(dir, { withFileTypes: true })) {
+    const full = path.join(dir, entry.name);
+    if (entry.isDirectory()) walk(full);
+    else if (entry.name.endsWith('.html')) pages.push(full);
+  }
+};
+walk(dist);
+
+const problems = [];
+let blocks = 0;
+let breadcrumbs = 0;
+
+// Google's rules for a BreadcrumbList: at least one ListItem, positions
+// counting up from 1, a name on each, and an `item` URL on each (Google
+// excuses the last; this does not, since the last is the page itself and
+// always has one). Repeating a URL is a step that goes nowhere.
+function checkBreadcrumb(list, where) {
+  breadcrumbs += 1;
+  const items = list.itemListElement;
+  if (!Array.isArray(items) || items.length === 0) {
+    problems.push(`${where}: BreadcrumbList has no itemListElement`);
+    return;
+  }
+  const seen = new Set();
+  items.forEach((step, index) => {
+    const at = `${where}: breadcrumb step ${index + 1}`;
+    if (step.position !== index + 1) problems.push(`${at} has position ${step.position}`);
+    if (!step.name) problems.push(`${at} has no name`);
+    if (!step.item) problems.push(`${at} ("${step.name}") has no item URL`);
+    else if (seen.has(step.item)) problems.push(`${at} repeats ${step.item}`);
+    else seen.add(step.item);
+  });
+}
+
+function visit(node, where) {
+  if (Array.isArray(node)) return node.forEach((child) => visit(child, where));
+  if (!node || typeof node !== 'object') return;
+  if (node['@type'] === 'BreadcrumbList') checkBreadcrumb(node, where);
+  for (const value of Object.values(node)) visit(value, where);
+}
+
+const ldPattern = /<script[^>]*type="application\/ld\+json"[^>]*>([\s\S]*?)<\/script>/g;
+for (const page of pages) {
+  const where = path.relative(dist, page).replaceAll('\\', '/');
+  const html = fs.readFileSync(page, 'utf8');
+  for (const match of html.matchAll(ldPattern)) {
+    blocks += 1;
+    try {
+      visit(JSON.parse(match[1]), where);
+    } catch (error) {
+      problems.push(`${where}: JSON-LD does not parse (${error.message})`);
+    }
+  }
+}
+
+if (problems.length) {
+  console.error(problems.join('\n'));
+  console.error(`\n${problems.length} structured-data problem(s).`);
+  process.exit(1);
+}
+console.log(
+  `Structured data OK: ${blocks} JSON-LD block(s) on ${pages.length} page(s), ${breadcrumbs} breadcrumb list(s).`,
+);
diff --git a/site/src/components/starlight/Head.astro b/site/src/components/starlight/Head.astro
index 1133fbfc..63bb4f5c 100644
--- a/site/src/components/starlight/Head.astro
+++ b/site/src/components/starlight/Head.astro
@@ -13,8 +13,12 @@
  *
  *   TechArticle    the page itself, with its own frontmatter title and
  *                  description, pointing back at the software it documents.
- *   BreadcrumbList  Docs > <section> > <page>, matching the sidebar group
- *                  the page sits in, so a result can show where it lives.
+ *   BreadcrumbList  Docs > <page>. Every step but the last needs an `item`
+ *                  URL, and the sidebar groups ("Start", "Interfaces")
+ *                  have no page of their own, so they are not steps.
+ *                  They were, as a nameless middle rung, and Search
+ *                  Console flagged every docs page: Missing field "item"
+ *                  (in "itemListElement").
  *
  * `isPartOf` ties every docs page to the same SoftwareApplication `@id` the
  * landing page declares, rather than restating the software's details on
@@ -22,7 +26,6 @@
  */
 import Default from '@astrojs/starlight/components/Head.astro';
 import { meta } from '../../data/site-content/meta';
-import { docsSidebar } from '../../data/site-content/docs';
 
 const route = Astro.locals.starlightRoute;
 const entry = route.entry;
@@ -35,22 +38,18 @@ const description = entry?.data?.description ?? meta.description;
 // identity for one page.
 const pageUrl = new URL(Astro.url.pathname, meta.domain).href;
 
-// Which sidebar group the page sits in, for the breadcrumb's middle rung.
-// Matching on the link rather than the slug because docs.ts writes links
-// with the leading and trailing slash this compares against.
-const section = docsSidebar.find((group) =>
-  group.items.some((item) => item.link === Astro.url.pathname),
-)?.label;
-
 const softwareId = `${meta.domain}/#software`;
 
+const docsUrl = `${meta.domain}/docs/`;
+// The docs index is the "Docs" step itself, so it gets one step, not the
+// same URL twice.
+const steps = [
+  { name: 'Docs', item: docsUrl },
+  ...(pageUrl === docsUrl ? [] : [{ name: title, item: pageUrl }]),
+];
 const breadcrumb = {
   '@type': 'BreadcrumbList',
-  itemListElement: [
-    { '@type': 'ListItem', position: 1, name: 'Docs', item: `${meta.domain}/docs/` },
-    ...(section ? [{ '@type': 'ListItem', position: 2, name: section }] : []),
-    { '@type': 'ListItem', position: section ? 3 : 2, name: title, item: pageUrl },
-  ],
+  itemListElement: steps.map((step, index) => ({ '@type': 'ListItem', position: index + 1, ...step })),
 };
 
 const techArticle = {