SEO work on a Nuxt site usually happens in two windows. In one, Search Console shows which page is weak. In the other, an editor shows a route file that does not look anything like that page. The translation between the two is where the afternoon goes. With Search Console connected to Claude Code, both live in the same terminal: the agent reads the data, finds the file, proposes the change, and you ship it.

Setup, once

Claude Code speaks MCP over HTTP, so a hosted connector is one command:

claude mcp add --transport http loopthink https://mcp.eu.loopthink.ai/mcp

Then /mcp inside Claude Code, sign in, done. The Google sign-in itself happened once in Loopthink with the read-only scope, so there is no key file in the repository and nothing in .env that could be committed by accident. The agent gets three kinds of calls: list properties, query search analytics by date, page, query, country or device, and read the submitted sitemaps.

The loop

  1. Ask the data. Which pages have impressions and no clicks, which queries sit on page two, which queries land on two URLs.
  2. Find the source. The agent maps a URL to the route file that renders it. In Nuxt that is file-based routing plus whatever feeds the page: a content file, a composable, a template string.
  3. Change it. Title, description, headings, internal links, canonical, prerender list, sitemap.
  4. Build and check. Generate the site and read the rendered HTML, not the source. What Google sees is the output.
  5. Write down the date. Four weeks later, ask the same question with the release date as the dividing line.

None of the steps is new. What is new is that the first step no longer ends in a spreadsheet that someone has to carry over to the fourth.

A real case: one template line, a hundred pages

This website is a Nuxt site, prerendered, with one dynamic route for every connector and one more for every connector and client. Two files, well over a hundred pages. We asked Claude Code which queries land on more than one of our URLs.

The answer: for connect google search console to claude the overview page /connect/google-search-console was shown 275 times at position 20, and the page written for exactly that sentence, /connect/google-search-console/claude, 227 times at position 31. The general page outranks the specific one, and both are on page two.

Then the second question, in the same session: why? The agent opened both route files and found it in a minute. The overview page builds its title from a template:

const title = `Connect ${system.name} to ${clientNames.join(' or ')} | Loopthink`

That renders "Connect Google Search Console to Claude or ChatGPT". The overview page claims the query of both of its children, word for word, on every one of our connectors. One line in one file, repeated across the whole catalog. No amount of staring at Search Console would have shown that, because Search Console does not know our pages come from a template.

The candidate fix is as small as the cause: give the overview a title about the system and the team, leave the client names to the client pages, and make sure the overview links to its children with the full phrase. We have not shipped and measured it yet, so we are not claiming a result. The point here is the path from a number to a line of code.

Nuxt places worth checking with data

  • Titles from templates. Anything in useSeoMeta or useHead built from a string template applies to every page of that route. Search Console data by page tells you whether the template works. One edit fixes all of them, or breaks all of them.
  • Prerendered routes. With nitro.prerender, a dynamic route that nobody links and nobody lists is simply not built. Ask the agent to compare the routes in the build output with the pages that have impressions. A page that had impressions and is gone from the output is a 404 you shipped.
  • The sitemap. If your sitemap is a server route rendered to a file, like ours, it is code, and code drifts. Let the agent read the sitemap status from Search Console and compare the submitted URLs with the generated ones.
  • Canonical and language alternates. A mirrored /de tree with a missing or wrong alternate shows up as the wrong language ranking in the wrong country. Query the data by country and page and it is visible in one table.
  • What is in the HTML. Content that only appears after hydration is content Google may not credit. Have the agent grep the generated HTML for the sentence that is supposed to rank.

What to keep in mind

Search Console data lags about three days and moves slowly. A title change needs weeks, not a deploy, to show. So this is not a loop you run ten times a day, and an agent that reports success an hour after a release is reporting noise. Give it the release date and ask again in a month.

The search terms are also the reason to think about where they go. They are typed by real people. In Loopthink the query dimension is off per connection until you switch it on, the scope at Google is read-only, and every call the agent makes is in the audit log under your name. The questions themselves, without the editor, are in SEO with Claude and Search Console.

Search Console in Claude →