A client emails at five on a Friday, asking whether it's true that a competitor now outranks them. Answering means a rank check, a look at the page involved, maybe a check of the AI answer, and then a reply written in the client's language. An assistant connected to live SEO data runs those checks mid-conversation, which turns the questions clients actually ask from a queue of small jobs into minutes of work.
An assistant that measures before it writes
On its own, Claude has no window onto today's results page. Ask it where a client ranks and the honest answer is a hedge, because there is nothing for it to measure. The previous lesson, AI in the agency workflow, drew the line: AI is reliable where its output can be checked against something real. Tools move that line, by giving the assistant something real to check.
MCP, the Model Context Protocol, is the open standard that does it. Connect Claude to an MCP server and it discovers the server's tools, calls one whenever a question needs a measurement, and writes its sentence around the data that comes back. The connection is one URL in your assistant's settings, and Connect Claude to your SEO data walks through the setup, so this lesson stays on what an agency does with it once it works.
Three lookups cover most client questions
The connector exposes three read-only tools, listed on the connector page.
- A live rank check. Give it a domain, a keyword and a country and it returns the domain's live Google organic position and the URL holding it; a zero means the domain isn't in the top 100, the range the tool's own description says it checks. This is the Friday email answered with a measurement, and it works just as well on a prospect's domain before a first call.
- A page's metadata. Give it a URL and it returns the title, meta description, H1 and headings. Use it to confirm a client's developer actually shipped the title change, or to see how a competitor's ranking page describes itself before you write the brief to beat it.
- AI presence. Give it a domain and a keyword and it reports whether Google shows an AI Overview or AI Mode answer for that query, and whether the domain is among the cited sources. "No AI surface for this query" is a real finding, common on local and transactional keywords, and it belongs on the AI page of the report as exactly that.
Chain the lookups like an analyst
Single lookups answer single questions. The analyst move is the chain, and Claude will run one from a single request. "Where does the client rank for their main keyword, what do the title and headings of the page holding that position say, and does the AI answer for the query cite them" is three tool calls and one summary. The pattern generalises: walk a client's tracked keywords and flag the ones sitting just off the first page, then pull the metadata of each ranking URL, and you have a quick-wins list with the evidence attached.
The summary is where agencies save real time. Ask Claude to phrase the results as a three-sentence client update and it will, in plain language, with the numbers in place. That draft isn't sendable yet. Ask for the raw tool results and confirm every number in the prose appears in the data, because an assistant can round, mix up two lookups, or paper over a failed call. The habit costs a minute, and it is what makes the workflow client-safe.
Keep it honest before it reaches a client
Four habits keep connected answers defensible.
- Report a lookup as what it is: one market, at one moment. A domain can sit near the top in Singapore and be absent in the United States at the same time, and the country you pass decides which you see. Your own browser also personalises results by location and history, so a neutral lookup that disagrees with what you saw this morning isn't necessarily wrong.
- Know what stays in the product. The lookups are read-only; full audits, content generation and anything else that spends real credit runs inside the product, and the same rate limits as the public tools apply, so a conversation firing dozens of lookups per minute will be told to slow down.
- Keep client confidentiality in mind, then relax. The endpoint is public by design, carries no account secrets and can't read your workspace, so the only client information in the conversation is what you type and what a lookup returns.
- The causal sentence stays yours. Claude can report that a position changed; whether your work changed it is a claim the data in one conversation can't carry.
What to take away
- Connected over MCP, Claude answers ranking, page and AI-visibility questions with live lookups instead of plausible prose.
- Three read-only tools, a rank check, page metadata and AI presence, cover most day-to-day client questions; audits and content generation stay in the product.
- Chained lookups plus a drafted summary turn a client question into a checked answer in minutes, and the raw tool results are how you check it.
- Every lookup is one market at one moment, and the sentence about why something moved is still written by you.
Next
With delivery, reporting and an AI analyst in place, the last chapter turns to growth: The first five clients playbook covers picking and keeping your first accounts.