Google Search Console MCP

Audit 1-5 own URLs: HTTP HTML, metadata, directives and exact-page GSC queries. Literal matches only; no JavaScript rendering or automatic edits. On demand; use nextCursor to continue explicitly.

Discover own sitemap URLs and inspect Google's indexed view, up to 10 per explicit call. Indexed state is not a live test; samples do not describe whole-site indexation. Never submits URLs.

Crawl a bounded own-site HTML link graph respecting robots.txt. Prioritize observed incoming links and literal contextual suggestions. Partial graph is not proof of orphan pages. Explicit batches only.

Record/list/delete changes or analyze before/after metrics on request. GSC_HISTORY=1 enables local history independently of exports; analyze can be transient. saveSnapshot requires a saved changeId. No scheduler, notifications or causal attribution.

Audit declared old/new URL mappings across authorized properties: redirects, destination HTML, indexed status and before/after metrics. Explicit batches; no edits or indexing submissions; no ranking guarantees.

Send an Indexing API notification for an updated or deleted URL. Google supports this API for JobPosting and livestream BroadcastEvent pages; a notification does not guarantee crawling or indexing. Requires write access and appropriate credentials. IMPORTANT: Base your analysis ONLY on the data returned. Report exact numbers from the results. Do not speculate about causes (e.g. algorithm updates, competitor actions) unless the data explicitly supports it. If the data does not contain enough information to answer a question, say so clearly rather than guessing.

Send up to 200 Indexing API notifications. Google supports this API for JobPosting and livestream BroadcastEvent pages; notifications do not guarantee crawling or indexing. Requires write access and appropriate credentials; quota applies. IMPORTANT: Base your analysis ONLY on the data returned. Report exact numbers from the results. Do not speculate about causes (e.g. algorithm updates, competitor actions) unless the data explicitly supports it. If the data does not contain enough information to answer a question, say so clearly rather than guessing.

Submit a sitemap to Google Search Console. Submission does not guarantee crawling or indexing. Requires write access and appropriate credentials. IMPORTANT: Base your analysis ONLY on the data returned. Report exact numbers from the results. Do not speculate about causes (e.g. algorithm updates, competitor actions) unless the data explicitly supports it. If the data does not contain enough information to answer a question, say so clearly rather than guessing.

Surface AI-conversation exhaust hiding in your regular query data: bare replies to Google's AI ('yes', 'go on'), 'what about X' pivot follow-ups, conversational questions, AI-visibility tracker probes, and full agent prompts logged as queries. Google counts every AI Mode follow-up as a new query and folds AI Mode/AI Overviews into the web search type, so these fragments carry real impressions, positions and clicks. The dedicated Generative AI report has no query dimension; this is the only query-level AI evidence available anywhere. Classifies every match into seven buckets with landing pages, plus a monthly timeline showing when reply-artefacts first appeared on your site. Treat probe and harness buckets as machine traffic, not demand. Pulls abort over 50k rows per request. IMPORTANT: Base your analysis ONLY on the data returned. Report exact numbers from the results. Do not speculate about causes (e.g. algorithm updates, competitor actions) unless the data explicitly supports it. If the data does not contain enough information to answer a question, say so clearly rather than guessing. PRESENTATION: Always present these results as a rich, interactive visualization using artifacts. Use summary cards for key metrics, colour-coded indicators for severity or change direction (green for positive, red for negative), bar charts or sparklines for comparisons, and tabbed sections to organise different categories of results. Make the output visually scannable and dashboard-like rather than plain text or raw tables. POSITION: position is an impression-weighted average over real impressions, blended across queries, devices and countries. It is not a rank-tracker rank. Deep positions are only recorded when a user actually reaches that part of the results, so values beyond page one rest on sparse data. Treat absolute positions as directional and prefer position deltas between periods when judging change.

Fetches pages from YOUR OWN site and audits every image on them for the on-page factors that drive image-search performance: missing/empty/generic/duplicate alt text, non-descriptive filenames, missing width/height attributes, lazy loading on the LCP candidate, srcset coverage, file format and weight, intrinsic dimensions vs Google's ~250x200 indexing minimum, ImageObject and licensable schema, max-image-preview, inline background images, and optionally metadata inside image files (camera EXIF and GPS that should be stripped, IPTC Creator/Copyright/Caption that should survive, XMP DigitalSourceType on AI-generated images). Find candidate page URLs with gsc_search_analytics using searchType=image and dimensions=[page], then audit selected pages. Page URLs must belong to the instance allowlist and every fetch stays on public hosts with redirects revalidated; images may be fetched from public CDN hosts. Returns a per-image findings table, page-level checks, and an ordered top_fixes list. IMPORTANT: Base your analysis ONLY on the data returned. Report exact numbers from the results. Do not speculate about causes (e.g. algorithm updates, competitor actions) unless the data explicitly supports it. If the data does not contain enough information to answer a question, say so clearly rather than guessing. PRESENTATION: Always present these results as a rich, interactive visualization using artifacts. Use summary cards for key metrics, colour-coded indicators for severity or change direction (green for positive, red for negative), bar charts or sparklines for comparisons, and tabbed sections to organise different categories of results. Make the output visually scannable and dashboard-like rather than plain text or raw tables.

What this server can do: tools, auth state, limits and instance allowlist. Works offline, no credentials needed.

無需輸入參數。

GSC properties visible to this instance. Preserves the exact siteUrl identifier.

無需輸入參數。

Flexible performance query with explicit coverage. Sort applies to the retrieved dataset, not to all Google data.

Index inspection for 1-10 URLs. Per-URL results in input order; one failure never discards the successes.

Registered sitemaps and their reported states. Readonly; never submits.

Page through a previous result with zero new Google calls. Datasets expire after 15 minutes and never cross profiles.

Totals, timeline and optional comparison for a site. Totals come from the aggregate row, never from summing a top.

Deterministic period comparison. Absent keys stay in their own blocks, never as zeros.

Observed queries for one exact page URL. No title/H1/intent inference.

Observed pages for one exact query. Two URLs sharing a query is a candidate, never proof of harm.

Review candidates by transparent rule: low_ctr, striking_distance, declining, ctr_benchmark, low_clicks. Heuristics with versioned thresholds, never diagnoses.

Queries served by several of your own pages. Candidates for editorial review, never consolidation orders.

Compare caller-defined segments (brand terms, folders, countries, devices) over one retrieved dataset. Groups may overlap.

Export a previous dataset to this machine (json/csv/markdown). Requires GSC_EXPORTS=1. Never overwrites, never uploads.

Traffic-drop decomposition: observed signals, hypotheses and missing evidence. Correlation is not causation.

Three-window decline candidates for web or image. Sustained fall only; anything less is not a finding.

On-demand rule evaluation with stable alert ids. Not continuous monitoring.

Observed queries at weak positions with their observed pages. Review, do not auto-create.

Prioritized action candidates with recoverable evidence. Nothing is edited automatically.

Per-property health with isolated failures and a shared budget. Never sums properties.

Deterministic markdown report from tool outputs. No model involved; save is export-gated.

Check one structured claim against comparable evidence: verified, contradicted or unverifiable.

Web vs image surfaces joined by key. Separate universes: no audience sums, no CTR transfer.

Correlate a traffic-drop window with confirmed Google ranking updates (core, spam, reviews) from the public Search Status Dashboard. Reports temporal coincidence only, never causation. Dashboard history starts Dec 2022.