On-Page Audit

Crawl a chosen set of pages on one property, run the technical on-page checks on each, and report the issues grouped by severity with the affected URLs and the fix. The user's request may contain a property, a list of URLs, a sitemap URL, a section path (for example "/blog/"), or a page count; use it to skip resolution and to narrow the URL set. This skill reads the pages' own HTML. For whether Google indexed them, run the indexing-health skill.

Workflow

1. Resolve the property. If the user's request or the conversation names a property, use it. Site URLs use GSC formats: sc-domain:example.com or https://www.example.com/. Otherwise call list_sites and pick the obvious match. If several plausible properties exist, ask the user which one before continuing.
   • The property also bounds the crawl: audit_onpage_seo fetches only URLs that belong to it (a domain property covers its subdomains, a URL-prefix property only its own origin and path prefix) and lists any other URL in skipped with a reason, so a URL on another host needs its own property.
2. Choose the URL set. Take the first source that applies, and keep the whole workflow to about 25 to 50 pages unless the user asks for more, because every page is fetched live from the user's site.
   • The user named URLs: use them as given, at most 25 per call (the urls cap).
   • "My top pages" or no hint: call query_top_pages with the siteUrl and limit: 25, omitting startDate and endDate for the last 28 settled days, and take the page values. Rows are sorted by clicks, highest first; ctr is a 0-100 percentage, never multiply it by 100; position is lower-is-better. GSC data lags about 2 to 3 days, so the window ends before today.
   • No list and no traffic read: omit urls and let the tool auto-pick the top pages by impressions with maxUrls (default 10, max 25) and minImpressions (default 0, counted over the same 28 days).
   • A sitemap or a section: call get_sitemap_performance with sitemapUrl (on the property host; one level of sitemap-index expansion) and maxUrls (default 500, max 2000). Audit the URLs with clicks or impressions first.
   • URLs with zero impressions are an index-coverage question for the sitemap-audit and indexing-health skills, not an on-page one, unless the user asks about them.
3. Run the audit in batches. Call audit_onpage_seo with siteUrl and urls (up to 25).
   • The tool fetches each page live, honouring robots.txt Crawl-delay per host and the property's crawl rate, and stops after about 50 seconds. When truncated is true, the URLs in notAudited were not fetched: call again with urls set to that list and repeat until notAudited is empty.
   • On a slow or throttled site, batches of 10 finish more reliably than batches of 25. A site that keeps answering 429 can be crawled more gently with update_site (lower crawlMaxConcurrency, raise crawlMinDelayMs) once the user confirms the change; a read-only connection refuses it.
   • Read auditedCount, skipped, contextRange (the 28-day window behind impressions and clicks) and summary: total_pages, total_issues, by_severity (critical, high, medium, low, info), indexable_pages, non_indexable_pages. summary covers one call, so merge the batches yourself before reporting totals and say so.
   • Read pages[], sorted by issueCount descending then impressions. Crawl fields per page: statusCode, redirected, redirectChain, responseTimeMs, bodyBytes, impressions, clicks.
   • On-page fields per page: indexable, indexabilityStatus, noindex, canonicalUrl, canonicalIsSelf, title, titleLength, metaDescriptionLength, h1Count, wordCount, internalLinks, externalLinks, imagesMissingAlt, imagesCount, hasStructuredData, structuredDataTypes, schemaIssues, hasViewport, lang, hreflangCount.
   • Each entry in issues[] has id, category, severity and message; issueCount is the length of that list and the primary sort key.
   • crawlNote: some pages answered a blocked-looking status because a bot rule rejects the crawler; relay the whitelisting advice instead of reporting broken pages.
   • robotsNote: robots.txt disallows GSCWizard-Bot, so those pages were not fetched and carry the blocked-by-robots issue.
   • metricsUnavailable with contextNote: impressions and clicks are null, meaning unknown, not 0; the on-page checks still stand.
   • If pages is empty, read emptyBecause and do what nextStep says.
4. Interpret by category and weight by traffic. Severities are critical, high, medium, low, info; categories are indexability, canonical, meta, headings, content, links, images, structured-data, technical. Read them with this table.

Always rank a finding by the impressions and clicks of the pages it affects, not by its count: a noindex on a page with clicks is an emergency, a short title on a page with no impressions is housekeeping. Issue message texts carry the measured value (the title length, the word count, the HTTP status), so quote them rather than restating thresholds.

5. Confirm a canonical mismatch (optional, quota-aware). Only for pages with impressions where canonicalIsSelf is false or canonicalUrl is null.
   • First call list_url_inspections with urlContains set to the path and latestPerUrl: true. It spends no quota and returns the stored google_canonical and user_canonical per row; a row whose google_canonical, user_canonical and inspection_result_link are all null predates canonical storage.
   • If nothing is stored, call get_inspection_quota and read remaining (Google allows 2,000 inspections per property per day, shared with the app; unavailable: true means the counter could not be read and the figures are placeholders, so treat the quota as unknown). Only when plenty remains, call inspect_url with inspectionUrl for at most 3 to 5 pages and compare googleCanonical, userCanonical and the audit's canonicalUrl. A null canonical means Google did not report one, which is common for unindexed pages.
   • Stop if a response carries quotaExhausted: true. For a wider inspection run, hand over to the indexing-health skill.

Output

Open with one line of scope: pages audited, pages skipped or not audited, the contextRange, and the merged issue counts by severity with indexable versus non-indexable pages. Then present, in this order:

1. Issues by severity (audit_onpage_seo, with the context range): one table, critical first, with columns Severity, Issue (id), Category, Affected pages, Example URL, Fix. Group identical ids across pages into one row and give the fix in one sentence.
2. Worst pages (top 5 to 10 by issueCount then impressions): URL, impressions, clicks, status code, issue count, the most severe issue. Keep prose for interpretation, not for restating numbers.
3. Canonical check (only if step 5 ran): URL, declared canonical, Google's canonical, verdict.
4. Not yet audited: the URLs still in notAudited or skipped, with the reason, if any remain.
5. Crawl notes: the whitelisting or robots.txt advice from crawlNote and robotsNote, when either is present, since those pages need the user's action before a rerun.

End with a short prioritized list of 3 to 7 concrete recommendations ordered by traffic at stake and effort, each naming the pages and the action. Present every multi-row result as a compact markdown table (top rows only, typically 5 to 10). Do not fabricate rows the tools did not return; if a batch returned no issues, say the pages passed the checks that were run.

Error handling

If any tool call fails with an authentication or subscription error, tell the user to run /mcp to (re)authenticate, and that a GSC Wizard account with an active plan or trial plus a connected Google Search Console account is required. If the property is not found, re-run list_sites and offer the closest matches instead of guessing a URL format. If a single tool fails for another reason, report the gap in the relevant section and complete the rest of the workflow with the remaining data.

Grounding

Every number you report must name the tool it came from and the date range it covers, for example "clicks 12,340 (get_site_summary, 2026-09-21 to 2026-09-27)". For a table, state the tool and date range once on the line above it. If you derive a figure yourself (a delta, a sum, a share), say so and name the tool outputs it was computed from. Never report a number no tool returned.