How to Use the Google Search Console MCP in Your SEO Process (6 Real Workflows)

A local Google Search Console MCP server lets Claude read your Search Console data through Google's official API. A question like "which pages lost impressions this month?" then gets answered from live property data in under a minute. The real gain comes when it becomes part of the SEO process, with a fixed set of prompts for the weekly health check, the period comparison, the indexing check and the CTR review that used to mean exporting CSVs into a spreadsheet.

Every prompt and output below comes from search.agency's own Search Console property in October 2026. The numbers are ours, including a page that quietly dropped out of search and a set of AI-written queries that turned up in our data.

What a local GSC MCP changes in day-to-day SEO work

MCP, the Model Context Protocol, is an open standard that lets an AI assistant call outside tools. A Search Console MCP server is one of those tools. It runs on your own computer, signs in to Google with your account, and gives Claude a set of functions for the Search Analytics API, the URL Inspection API and sitemaps. We explained the protocol itself in our Screaming Frog MCP walkthrough, and the focus here stays on Search Console.

The server we use is the open-source mcp-gsc project by Amin Foroutan, released under the MIT license. When you ask it what it can do, it lists its tools in four working groups covering properties, analytics and reporting, URL inspection and indexing, and sitemaps. Tools that change your account, such as adding or deleting a property, stay switched off unless you set an environment variable to allow them.

Running it locally matters for agencies. The credentials file and the cached login token stay on your machine, and the data moves from Google into the Claude session you are running without passing through a third-party reporting tool. The query results still land in your Claude conversation. Your agency's rules on client data in AI tools apply to them.

The setup in five minutes

Our full Google Search Console MCP setup guide covers OAuth, service accounts and the errors that break installs. The short version has four steps.

1. Enable the Search Console API in Google Cloud Console and create an OAuth client ID of the Desktop app type.
2. Download the client secrets JSON and keep it at a permanent, absolute path.
3. Install uv, then run which uvx to get the full path for the config file.
4. Add the gscServer block to claude_desktop_config.json, quit Claude Desktop completely and reopen it. The first tool call opens a browser window for the Google sign-in.

Two details prevent most of the debugging. Use the full path to uvx, because desktop apps often do not read your shell profile and fail with a spawn uvx ENOENT error. Use absolute paths for the credentials file too, never a ~/ shortcut.

The smoke test is a single prompt asking Claude to list your Search Console properties. It should return every property your Google account can see, with its permission level. On our account it returned 18 properties, which is a good reason to name the exact property in every prompt that follows.

Workflow 1. A Monday health check in one prompt

The first prompt we run each week asks for the 28-day overview and an explanation of anything unusual. The performance overview tool returns totals plus a daily trend, and Claude reads the trend so nobody has to study a chart line by line.

For https://search.agency/, give me a 28-day performance overview. Show total clicks, impressions, CTR and average position, then point out any days or stretches where position or clicks moved sharply. Keep it to five bullet points.

For 24 September to 22 October 2026 the property recorded 227 clicks, 20,552 impressions, a 1.1 percent CTR and an average position of 16.0. The more useful line in the answer was about position. Average position improved to 10.7 on 30 September, then drifted to 22.0 by 17 October while impressions held between roughly 560 and 920 a day.

A falling average position with steady impressions usually means the site is appearing for more queries further down the results page, which a page-level comparison can confirm. The overview also includes the most recent days, which Google marks as preliminary because the data can still change. We read the last few points of the trend loosely.

Workflow 2. Finding the pages that moved between two periods

The compare tool takes two date ranges and returns the change in clicks, impressions, CTR and position for each page or query. Period 1 is the period you are analyzing and period 2 is the baseline. A positive change means the recent period did better.

Compare https://search.agency/ for 22 Sep to 19 Oct 2026 against 25 Aug to 21 Sep 2026, grouped by page. Show the pages with the biggest click changes in a table, and flag any page that lost more than half its impressions.

The homepage lost 12 clicks, down 9.7 percent, while gaining 983 impressions, which pulled its CTR from 2.3 to 1.8 percent. Two new posts, the ChatGPT Ads guide and our guide to fixing crawled and discovered but not indexed pages, picked up 825 and 375 impressions in their first full month at average positions between 6 and 7. The Screaming Frog advanced workflows post climbed from position 20.7 to 9.6.

The flag caught one page. Our post on SEO and AI search campaign costs in Indonesia fell from 278 impressions to 7. A drop like this is easy to miss in the Search Console interface because the page never earned many clicks, and a fall that steep deserves its own diagnosis.

Workflow 3. Diagnosing a page that went quiet

When a page loses visibility, the first question is whether Google can still index it. The URL inspection tool answers that in one call, returning the index verdict, last crawl date, the canonical Google selected, robots status and any rich results it detected.

Inspect https://search.agency/blog/seo-ai-search-campaign-cost-indonesia in the search.agency property. Then pull its daily impressions from 23 Aug to 19 Oct 2026 and tell me whether this looks like an indexing problem or a ranking problem.

The inspection came back clean. The page was submitted and indexed, crawled on 21 October 2026 as mobile, allowed by robots.txt, and Google's canonical matched ours. The daily data explained the drop. The page first appeared on 7 September with 50 impressions, held between 6 and 42 a day until 19 September, and then fell to almost nothing from 20 September onward.

That pattern fits a new page that Google tested near the top of the results and then moved down. The work shifts to the content and the internal links pointing at it. The inspection listed only one referring page, our post on owning AI search visibility, which gave us a concrete first task. Google's referring list is a sample and may leave some links out. We confirm it with a crawl before acting on it.

Workflow 4. Finding CTR gaps on queries that already rank

Queries with plenty of impressions, a decent position and almost no clicks are among the cheapest wins in SEO, because the page already ranks and only the title or snippet needs work. The advanced analytics tool pulls query and page pairs together and sorts them by impressions, which makes these gaps easy to spot.

For https://search.agency/, pull query + page data for 22 Sep to 19 Oct 2026, sorted by impressions. List queries with 100+ impressions, position better than 15 and CTR under 2%. For each, suggest a new title tag that matches the query intent.

The query "ai search agency" earned 1,700 impressions at an average position of 4.0 for our homepage and only 10 clicks, a CTR of 0.6 percent. Our Ahrefs content gap tutorial ranked at position 3.6 for "ahrefs reports" with 175 impressions and no clicks, which points to a searcher who wants a list of Ahrefs reports and a title that promises a single workflow. "chatgpt ads" sent 308 impressions at position 7.6 to the ChatGPT Ads guide for 3 clicks.

We treat Claude's title suggestions as drafts and check each one against the live results page before changing anything, because the query alone does not show what already ranks above us. The brand query "search agency" makes a useful benchmark on the same site, since a title that matches the query earns a 19.8 percent CTR there.

Workflow 5. Checking indexing after you publish

The batch inspection tool checks several URLs in one call, which turns a post-publish check into a single prompt. Google's usage limits allow 2,000 URL inspections per day and 600 per minute for each property. The batch tool suits a handful of priority pages, and a site-wide check still belongs to a crawler.

Batch-inspect these six URLs in https://search.agency/ and return one row each with index verdict, coverage state, last crawl date and rich results detected. Then check the sitemap status for the property.

All six pages came back as submitted and indexed, with last crawl dates between 29 September and 17 October and breadcrumb rich results detected on each. The sitemap check returned one sitemap with 252 URLs, zero errors and 4 warnings. Google last downloaded it on 19 October 2026 even though it was last submitted in May 2020, which is normal for a sitemap that Google refetches on its own schedule.

We run this check the day after publishing and again a week later. If a page is still missing, our guide to crawled and discovered but not indexed pages covers the usual causes, and pairing the check with IndexNow and the Bing Webmaster MCP covers Bing as well.

Workflow 6. Spotting AI agents in your query data

Some of the queries in our October data were plainly not typed by people. One query that sent impressions to our AI search service page opened with "about me: marketing manager/director series a to series c" and ended with a question asking which agencies help B2B companies get cited by ChatGPT and Google AI Overviews. Others carried long chains of -site:reddit.com -site:twitter.com exclusions or numbered prefixes such as "984: seo competitors tracker".

In https://search.agency/ query data for the last 60 days, find queries longer than 12 words, queries with 3+ search operators, and queries that start with a number and a colon. Group them by landing page and say what each group suggests about who or what is searching.

These look like prompts sent by AI visibility trackers or research agents that run Google searches on someone's behalf, often with a buyer persona pasted in front of the real question. We cannot see which tool sent them. We treat them as signals and keep them out of any demand estimate. They still show which buyer questions are being tested in our category and which of our pages Google returns for them.

For AI search work this is useful raw material. The questions inside those prompts can go straight into the prompt set you use to measure AI search visibility, and pages that keep appearing for them deserve answer-first sections that an AI engine can quote. Our guide on getting cited by AI search engines covers how to write those sections.

Limits to know before you trust the answer

Search Console withholds some queries to protect user privacy, and the gap can be large on a single page. Over 60 days our cost page recorded about 285 impressions at page level, while the page-by-query tool could attribute only 20 of them to named queries. Google explains this filtering in its deep dive on performance data limits, and the practical rule is to take page totals from page-level data and treat query lists as a sample.

Claude can also misread data. Average position in a period comparison shifts with the query mix, CTR on low-impression rows swings wildly, and a model can describe a pattern with more confidence than five impressions allow. We ask for the raw table next to every summary and spot-check one or two numbers in the Search Console interface before anything goes into a client report.

A few practical limits apply to every setup.

• The Search Analytics API returns up to 25,000 rows per request. Large sites need paginated pulls, which the advanced analytics tool supports.
• The server runs on your own computer and works with any MCP client on that machine, such as Claude Desktop, Cursor or Codex CLI.
• Data from client properties lands in your Claude conversation. Confirm each client is comfortable with that before you connect their property.
• Destructive tools stay off by default, and we keep them off for every client property.

Where the GSC MCP fits in a weekly SEO routine

The workflows become a process once they run on a fixed schedule with fixed prompts. We keep the prompts in a shared document. Every analyst asks the same question the same way, which makes the answers comparable from one week to the next.

The same routine extends to Bing through the Bing Webmaster MCP and to crawl data through Screaming Frog's MCP server. One assistant can then tell you whether a page is crawlable, indexed and visible in both engines. Someone still has to decide which of those numbers matter for the business, and rankings on their own make a weak KPI. If you want this routine running on your own properties, or want our team to run it as part of an SEO and AI search audit, get in touch with Search Agency.