How to use Claude for SEO: live data through MCP, Claude Code and honest limits

Claude is strongest at the parts of SEO that are reading, reasoning and writing over material you give it. Hand it a table of queries and pages, a crawl export or a page's text, and it does in minutes what takes an afternoon of filtering:

• Analysis. Explain a change between two periods, group queries by intent, spot two pages competing for the same search, and turn a crawl export into a list ordered by how many pages each issue affects.
• Writing. Briefs built from the queries a page already gets, title and meta description options that match those queries, answers for a FAQ, outlines for missing sections.
• Technical drafting. JSON-LD from a product's fields, redirect maps, robots.txt rules, regular expressions for filtering queries and URLs.
• Review. Reading a page against a query and saying what it leaves unanswered.

What it cannot do on its own is know anything about your site. Without a connection to your data, Claude answers from its training: it can explain how Search Console works, but not what happened to your clicks last week.

Pasting an export works for small questions and runs out quickly, because the Search Console interface exports at most 1,000 rows and many opportunities sit in the long tail. The better route is a connector: an MCP server that lets Claude query the data itself. The guide to SEO agents explains how MCP works; in Claude, Anthropic's help centre describes it like this.

• Plans. Custom connectors using remote MCP are available to Free, Pro, Max, Team and Enterprise users, with Free users limited to one.
• Adding one. On Pro and Max, open the Connectors page, click Add and choose Add custom connector, then enter a name and the server's URL. Claude detects the sign-in method; a server that uses an API key takes it as a request header.
• Teams. On Team and Enterprise, an Owner adds the connector for the whole workspace, and each member then connects their own account.
• Using it. Turn the connector on for a conversation from the + button at the lower left of the chat, under Connectors.
• Reachability. Claude connects from Anthropic's cloud, not from your computer, so the server must be reachable over the public internet.

Anthropic also warns that custom connectors reach services it has not verified, and that a malicious server may include hidden instructions that try to make Claude perform unintended actions. Add only servers you trust. The guide to connecting Search Console to Claude walks through one connection from start to finish.

Claude Code is Claude working in your terminal and your site's repository. For SEO, that means it can fix a cause rather than describe it: find the template that prints a relative canonical URL, change it, run the build and show you the diff. It suits work such as:

• auditing templates for titles, canonicals, hreflang, robots meta tags and structured data;
• writing scripts that pull Search Console data, generate sitemaps or test redirects;
• making the same change across hundreds of content files, such as adding a missing field to every article.

You add a data source with claude mcp add --transport http <name> <url>, adding an Authorization header with --header for a server that uses a key, or signing in through /mcp for one that uses OAuth. A server added at project scope is stored in .mcp.json at the project root, so your team shares it through version control.

In Manual mode, Claude Code reads files without asking but stops before most edits, shell commands and network requests. Newer versions start interactive sessions in auto mode instead, where a second model reviews each action in your place, so when the work touches your live site's templates, start with claude --permission-mode manual and approve each step. For scheduled jobs, claude -p runs a prompt without the interactive session, so a daily report can be one line: claude -p "List pages that lost over 30% of their clicks" --permission-mode dontAsk --allowedTools "mcp__gsc__*", where gsc is the name you gave the Search Console server. In dontAsk mode Claude Code refuses anything that would otherwise ask for approval, so the job can read files and query that server but cannot edit anything. Add --output-format json and each run also reports an estimated cost.

Requests work best when they name the data, the comparison and the output you want. With Search Console connected, for example:

• "Compare clicks by page for the last 28 days with the 28 before. List pages that lost more than 30% and had at least 100 clicks, with the queries that lost the most for each."
• "Find queries where two of my pages both get impressions, and say which page should own each one."
• "For this URL, list its top 30 queries and the questions in them that the page does not answer."
• "Rewrite this page's title and meta description to match the queries it actually gets. Give three options."
• In Claude Code: "Check every page template for a self-referencing canonical and list the ones without one."

Picture a freelance consultant who connects a client's Search Console to Claude on a Friday afternoon and asks the first question above. Claude lists six pages, and for the largest drop it shows the clicks moving from an old guide to a newer page on the same site that answers the same queries. She reads both pages, agrees they overlap, and recommends merging them with a 301 rather than rewriting the old one.

If you have caught a chatbot inventing a statistic, you are right not to take Claude's numbers on trust. This section covers where it fails in SEO work and the habits that catch each failure.

• Invented or misread facts. Anthropic's own guidance says even the most advanced language models, Claude included, can generate text that is factually incorrect, and that its techniques reduce hallucinations without eliminating them. Give Claude permission to say it does not know, ask it to quote the data behind each claim, and tell it to use only the data you provided.
• Numbers without context. Ask for the date range, filters and row counts behind every figure. Query rows rarely add up to Search Console's chart totals, because anonymised queries are left out of the rows, so a summary built by adding rows undercounts.
• Publishing unreviewed text. Google says it is critical to manually fact-check and review all AI-generated content before publishing, titles, meta descriptions and structured data included. The article on whether Google penalises AI content covers the policies behind that.
• Servers you cannot vouch for. Claude Code's documentation tells you to verify that you trust each server, and notes that servers fetching outside content can expose you to prompt injection.
• Cost and limits. Long exports and many tool calls use up your plan's limits or your API budget. Ask for filtered data rather than whole tables.

AutoPSEO's MCP server works as a Claude connector and in Claude Code, giving Claude Search Console, GA4, research and content tools at one endpoint. If you want Claude to write for your site, the content engine can use your own Claude Code as its writer: the same topic queue, brief and checks as AutoPSEO's writer, no articles counted against your plan's quota, and a runner that can start it on a schedule. The guide to writing with your own AI shows the setup.