How to use Claude for a technical SEO audit (with MCP)

Running a crawl is the easy part. The slow part comes after: working out which of the 23 issues actually matter for this site, putting them in order, and writing them up so a developer can act without a call to explain them.

That’s where an AI assistant like Claude helps, but only if it can see your real data. Paste a CSV export into a chat and you get generic “best practices”. Let the assistant query the audit directly and you get answers that name your URLs, your templates and your numbers.

Crawlens includes an MCP server for exactly that. This guide walks through a typical session.

What can Claude see in a Crawlens audit?

MCP (Model Context Protocol) is an open standard that lets an AI app call tools. Once connected, Claude can call eight read-only Crawlens tools:

The server is read-only and local. It can’t start a crawl or change anything, and Crawlens itself doesn’t upload your data. Setup takes a couple of minutes: see Connect Crawlens to your AI tools.

Step 1: ask for a summary, not a list

Start broad:

Summarise the latest Crawlens audit of example-store.com and give me a prioritised fix list.

Claude calls get_overview, then list_issues, then get_issue_urls for the issues it wants to check. Because it sees counts and affected URLs, it can spot what a flat issue list hides. It might notice, for example, that 37 broken internal links all come from the same footer block: one template fix, not 37 tickets.

Crawlens also ships a ready-made prompt, Summarise the audit & prioritise fixes, that asks for exactly this: an executive summary, then fixes ordered by impact against effort, each with who should fix it (developer, content or SEO) and example URLs.

Step 2: rank issues by traffic, not just severity

An issue on a page nobody visits can wait. The same issue on a page with 2,000 clicks a month can’t. If you’ve connected Search Console or GA4 in Crawlens, Claude can see that difference too.

Two checks are especially worth asking about:

• Pages with search traffic return errors. URLs with Search Console clicks that now return 4xx or 5xx.
• Pages with search traffic are not indexable. Pages people find in Google that now carry a noindex or point their canonical elsewhere.

Try:

Which pages with Search Console clicks have critical or warning issues? Sort by clicks.

That’s a query_urls call filtered and sorted the way you’d do it by hand in the URLs table.

Step 3: write it up for the person who’ll fix it

This is the part that usually eats an afternoon. Pick one issue and ask:

Explain the redirect chains to our developer.

The built-in Explain an issue for developers prompt pulls the affected URLs, opens two or three with get_url_details, and writes an explanation using this site’s actual redirect hops: why it matters, how to reproduce it with curl or browser devtools, how to fix it, and a checklist to confirm the fix in the next crawl.

For content issues, Rewrite titles & meta descriptions drafts replacements for pages with missing, duplicate, too-long or too-short titles and descriptions. It reads each page’s H1 and headings first, keeps titles to 30–60 characters and descriptions to 70–160, and writes in the page’s own language.

Step 4: check what changed after the fixes

After the fixes ship, crawl again and ask:

What got better and what got worse since the previous crawl?

compare_crawls returns the health score change, new and fixed issues per check, and URLs that were added, removed or changed. It’s a quick way to catch the regression that came with the fix, like the redirect cleanup that accidentally noindexed a section.

Tips for better answers

• Name the project. “The example-store.com audit” is enough. If the name doesn’t match, Crawlens lists the projects that are shared.
• Ask for URLs, not just counts. Recommendations get much better when the assistant has looked at real examples.
• Mind client data. Crawlens doesn’t upload anything, but your AI app sends what it reads to its own service as part of the chat. If that’s a problem, hide Search Console and Analytics data in the MCP settings, or only share some projects.
• Treat the output as a draft. Claude reads the same data you would, fast and usually correctly, but you know the site and the client. Check its priorities before they become tickets.

What do you need to get started?

Crawlens 0.20 or later, at least one audited crawl, and an MCP-capable app: Claude Desktop, Claude Code, Codex, Cursor or VS Code. Turn sharing on in Settings → Integrations → AI tools (MCP), follow the setup guide, and ask your first question.