MCP playbook

Grounding a prompt set in Search Console, from any client

The hardest part of AI visibility is choosing the questions. Search Console is the one first-party record of what buyers actually type, and the connector exposes it as a tool. This recording shows the honest response when a project is not connected, the two-step connection, and the comparison your client runs once it is.

Recorded 4 September 2026 on the Linkeddit connector, project 441e8df6 (linkeddit.com), which had no Search Console connection on the day. The connected response is quoted from the tool contract, not invented.

The transcript

  1. 1

    You ask: Which questions do real buyers type before reaching our site?

    Tool call: gsc_queries_get

    { "project_id": "441e8df6-…", "limit": 15 }

    Response, abridged

    {
      "project_id": "441e8df6-…",
      "connected": false,
      "queries": [],
      "explanation": "This project has no Google Search Console connection, so there is no first-party search data to report. This says nothing about the site's search demand. Connect it from the project's page in Linkeddit, which asks you to add Linkeddit's address as a user on the property in Google Search Console."
    }

    This is the real response for our own project on the recording day, and it is the point of the playbook: the tool refuses to guess. An empty list plus 'connected: false' plus a sentence saying it says nothing about demand is what an honest data tool returns. A tool that returned zero rows without the flag would let your client conclude nobody searches for you.

  2. 2

    You ask: How do I connect it?

    Tool call: The connection is not an MCP call. It is two steps in two places.

    1. In Google Search Console, open the property for the project's domain:
       Settings > Users and permissions > Add user.
       Paste Linkeddit's service-account address (shown on the project's Overview page; it begins
       linkeddit-search-console). Restricted permission is enough for reading.
    
    2. In Linkeddit, on the project's Overview page, press "Check for properties" and choose the property.

    Response, abridged

    What the app verifies before it records the connection:
    - that our service account can actually read the property you chose (read-only Search Console scope)
    - which property the project uses
    
    No OAuth redirect, no customer token stored. You can disconnect from the same panel at any time and
    remove the address from your property's users in Google to revoke the grant itself.

    The service account is Linkeddit's own. You grant it access inside Google, so no customer credential is ever held by Linkeddit. The help article on connecting Search Console has the panel-by-panel version.

  3. 3

    You ask: Once connected: the same question again.

    Tool call: gsc_queries_get

    { "project_id": "441e8df6-…", "days": 28, "limit": 50 }

    Response, abridged

    Response shape, from the live tool description (we cannot show real rows for this project until
    it is connected, and we will not invent them):
    
    "Get the search queries real buyers used to reach this project's site, from the customer's own Google
    Search Console, ranked by impressions … Returns the query, its impressions and clicks, and its average
    position."
    
    Parameters: project_id, days (default 28), limit (default 50), offset.

    Ranked by impressions, so the first rows are the questions your site is already shown for. Position tells you where you rank; a high-impression query at a poor position is the question to measure and the page to fix.

  4. 4

    You ask: Check the approved prompt set against those queries.

    Tool call: answer_prompts_list

    { "project_id": "441e8df6-…", "status": "approved" }

    Response, abridged

    {
      "items": [
        { "text": "What is the best competitor analysis tool for product marketers?", "intent": "commercial", "priority": 86 },
        { "text": "Which competitive intelligence tools use Reddit and review site data?", "intent": "commercial", "priority": 87 },
        { "text": "What are the best AI visibility tools for B2B SaaS?", "intent": "commercial", "priority": 88 },
        { "text": "What is the best competitor monitoring software for tracking pricing changes and product launches?", "priority": 90 },
        { "text": "What are the best answer engine optimization (AEO) tools?", "priority": 92 },
        { "text": "What are the best Crayon alternatives?", "priority": 93 },
        { "text": "What are the best AI citation tracking tools?", "priority": 98 },
        { "text": "What are the best AI brand monitoring tools?", "priority": 100 },
        "… 13 more"
      ]
    }

    This call is real. The grounding step is a comparison your client does: which approved questions match how buyers actually type, which high-impression queries have no approved question yet, and which approved questions nobody asks. Add the missing ones with answer_prompts_update and remove the dead ones.

Why first-party queries, and what they are not

Two things Search Console gives an AI visibility programme, and one it does not.

Real phrasing

Buyers type queries in words you did not choose. Approved prompts that mirror those words measure the questions that exist; prompts that mirror your positioning measure questions you wish existed.

Striking distance

A high-impression query at a poor position names both the question to measure and the page to fix. Add it as a prompt, run the audit, and the content gap view says what that page misses.

Not AI prompt volume

Search Console records Google searches. How often a private assistant is asked the same thing is not published by anyone, and the does-not-do page says so. Grounding narrows the guess; it does not remove it.

Frequently asked questions

Why publish a playbook whose first call fails?

Because the failure is correct behaviour and the reason people should trust the tool. A visibility product that fills an empty Search Console with a guess would be doing the thing every review in this category complains about. The transcript shows what you will actually see before connecting, and the exact two steps to connect.

Which client does this need?

Any. gsc_queries_get and answer_prompts_list are ordinary read tools on the connector. Claude Desktop, Claude Code, Cursor, VS Code and Windsurf all work; the connection itself happens in Google and in the Linkeddit project page, not in the client.

Can Linkeddit write to my Search Console?

No. The scope requested is read-only. We do not submit sitemaps or change settings, and the credential cannot.

Connect your property and run the comparison

The panel-by-panel steps are in the help centre. Once connected, gsc_queries_get answers from your own data in every MCP client.