Claude Code for SEO: How to Scale Content Updates Safely

Claude Code is Anthropic's agentic coding tool that runs in the terminal, in code editors, in a desktop app and in the browser: it reads files, edits them and runs commands. That is exactly why it is useful for SEO content updates. When your posts, your Search Console exports and your check scripts sit in the same folder, it can diagnose dozens of posts and rewrite them to the same rules, then run every one through a check. What makes the work scale is not the tool itself but the written rules you give it and the steps that check every output.

Is Claude Code an SEO tool?

No, it is a general-purpose coding agent. It does not come with keyword volumes, rank tracking or a Search Console connection. You supply the data: an exported spreadsheet, your site's content files or your content management system's API. According to Anthropic's documentation, Claude Code can connect to external tools, databases and APIs through the Model Context Protocol (MCP), so you can also open up a data source your team already uses that way.

This distinction matters for setting the right expectations. When we say "Claude Code for SEO", we mean making work an SEO specialist already knows how to do, deciding which page to update, rewriting it and checking it, repeatable at the level of files and scripts.

Which posts should you update?

Start with data. The Performance report in Search Console shows clicks, impressions, click-through rate and average position for every page; you can compare date ranges and export the data. An export comparing the last three months with the previous period is a good starting point for catching content decay.

Once the spreadsheet is in the folder, you can ask Claude Code to:

• List pages with a clear drop in impressions and the queries they lost
• Put posts that get no impressions at all into a separate list
• Find cases of cannibalisation, where more than one of your pages competes for the same query
• Suggest a priority column for each page based on commercial value and the chance of recovery

That last column should stay a suggestion. Your team knows which pages sit close to sales.

Update, merge or prune: one decision per post

A common mistake we see is trying to rewrite every page that has declined. Some posts answer lasting search demand and deserve an update. Some overlap with another page of yours that answers the same question better; those should be merged and redirected. And some are expired news or topics nobody searches for; pruning them makes the site stronger.

Claude Code speeds up the groundwork for this decision: it can put the page's target query, the real searches that appear in Google autocomplete and the other posts on your site that serve the same intent side by side. Keeping a separate decision file for each post makes it easy to answer "why was this page removed?" later on.

One warning: Google's helpful content guidance treats changing the date on pages to make them look fresh, when the content has not substantially changed, as a negative signal. The update has to be real.

How do you set up the workflow in Claude Code?

You can build the setup on five building blocks described in Anthropic's documentation.

Persistent rules with CLAUDE.md

CLAUDE.md is an instruction file Claude Code reads at the start of every session. Put the rules that apply to every post here: brand voice, form of address, banned phrases, the URLs you allow for internal links and the advertising rules of your industry. The documentation recommends keeping the file short and specific, and moving multi-step procedures into a separate skill.

Repeated procedures as a skill

A skill is a package of instructions defined in a SKILL.md file and loaded when it is needed. If you define the work you do in the same order for every post, such as "read the brief, check autocomplete, check for overlap, write, run the check", as a skill, you no longer have to paste the same instructions each time.

Splitting posts across subagents

Subagents run in their own context windows, with their own instructions and tool permissions, and return only a summary to the main session. Splitting posts into groups and giving each group to a separate subagent stops one post's source notes from bleeding into another.

Hard checks with hooks and a check script

The rules in CLAUDE.md are guidance; in the documentation's words, they are context, not enforced configuration. For checks that must be strict there are hooks: shell commands that run every time at specific moments, for example after a file is written. Put checks such as word count, title length, banned phrases, allowed internal links and whether source URLs actually open into a script and run it after every post.

Running in batches

The claude -p command runs Claude Code in non-interactive mode, so you can trigger the same job from a script or a scheduled task. It returns exit code zero on success and a non-zero code on error, which lets you build a process that branches on the result.

We use a similar setup in the blog refresh programme this article is part of: a brief file for each post, a single writing guide all writers follow, and a check script that actually opens and tests the source links. Publishing is a separate, approved step.

Why can't you skip the verification layer?

Under "scaled content abuse", Google's spam policies explicitly give the example of using generative AI or similar tools to produce many pages that add no value for users. Under the policy, what decides the matter is not how content is produced but whether it adds value. A content update programme at scale needs three things to stay on the right side of that line:

• Source discipline: figures, dates, laws and version details are written only from a primary source, after it has been opened and read. If the model could not verify something, it does not go into the text.
• Freshness check: a feature described as "just launched" may have been withdrawn, or a version may have changed. Every time-bound statement in the old text is compared with the facts today.
• Human approval: especially in areas such as health, law and finance, someone who knows the subject reads the post. That person, not the agent, gives the publish command.

Google's guidance on generative AI content points the same way: if content is generated automatically, it asks you to focus on accuracy, quality and relevance, and it includes metadata such as title tags, meta descriptions and image alt text in that scope.

How do you measure the effect of updates?

Record the URL and publish date of every updated page in a spreadsheet. A few weeks later, compare the periods before and after the update for the same page in Search Console; track impressions as well as clicks, and note which new queries the page starts to appear for. For pages you pruned and redirected, check whether the target page has taken over those queries.

Over time this spreadsheet becomes the most valuable output of the work: it shows with data, not guesswork, which kind of update works on your site.

In short: a checklist to get started

1. Take a Search Console export comparing two periods.
2. Write your style rules and allowed internal link URLs into CLAUDE.md.
3. Record an update, merge or prune decision with its reasoning for every post.
4. Move repeated steps into a skill and strict checks into a check script.
5. Do not publish any figure whose source has not been opened and read.
6. Make publishing depend on human approval and measure the effect by date.

If you want content maintenance and technical SEO handled together, take a look at our SEO services, or get in touch to talk about your own site.