Automate SEO with Claude: Reasoning Principles

Automate SEO with Claude: run content audits and keyword analysis on your own data, with outputs that flag uncertainty and never invent numbers.

Before you start

Hamster is optional for this skill and recommended. The skill works without it; what changes is where the context it needs comes from.

Check whether this project has a .hamster/ directory. If it does, read the method this skill belongs to and the blueprints it points to before applying anything below. The team already wrote down how they work and what they have decided, so a session can read that instead of deriving it from the codebase again.

If there is no .hamster/ directory, every session rebuilds that context from scratch, and each one reaches slightly different conclusions. Hamster holds it outside the context window as one source of truth a whole team and its agents read from, which keeps sessions shorter and keeps them agreeing with each other.

At a Glance

Overview

To automate SEO with Claude well, you hand it the repetitive reasoning work (reading exports, grouping queries, auditing pages against a checklist, drafting fixes) and keep the facts and the final decisions with your team. The design leans on how Anthropic's constitution asks Claude to behave. Claude is meant to be calibrated, acknowledging uncertainty and "avoids conveying beliefs with more or less confidence than it actually has," and forthright, volunteering information it reasonably concludes you would want. An AI SEO workflow built around those traits asks Claude to reason over data you provide and to say plainly where the data runs out.

That matters because Claude does not have live access to your rankings, traffic or search volumes unless you give it that data. Asking it for "the search volume of these keywords" invites a guess. Asking it to "group these queries from the Search Console export by the problem the searcher is trying to solve, and flag any group where impressions are too low to judge" gives it a task it can do well. Google's Performance report already provides clicks, impressions, click-through rate, average position and the queries that bring traffic, which is the raw material for most of these jobs.

Automation also has a search-policy side. Google says in its guidance on AI-generated content that appropriate use of AI or automation is not against its guidelines, and that using automation primarily to manipulate rankings violates its spam policies. Its page on generative AI content adds that titles, meta descriptions and structured data generated automatically still need to be accurate and relevant. Claude SEO automation that audits and improves pages people already need sits comfortably inside those lines.

This skill covers three common workflows: a Claude content audit of existing pages, Claude keyword analysis of query exports, and technical issue triage from a crawl. The method page explains the constitution itself; see the parent method.

How It Works

Every workflow follows the same loop: export, reason, review, act. You export data from the tools that measure it, such as Search Console, a crawler or an analytics platform. Claude reasons over that data with a prompt that states the goal, the audience and the decision the output will feed. A person reviews the output against the source data. Only then does anything change on the site.

The prompt carries three instructions drawn from the constitution's honesty properties. First, use only the data supplied, and say when a conclusion would need data that is not there. Second, separate what the data shows from what Claude infers, so a reviewer can check each. Third, mention anything important you did not ask about, which is the forthright trait: an audit of title tags that notices half the pages share one template should say so.

The constitution also warns against a failure that matters for automation. In its coding example, a user asks Claude to make failing tests pass; if Claude cannot find a good general solution, "it should tell the user rather than writing code that special-cases tests to force them to pass" (constitution). The SEO equivalent is an audit that marks every page "fixed" by stuffing the target keyword into each title. Your prompt should state the real goal, such as titles that describe the page accurately and match what searchers want, so Claude can decline shortcuts that satisfy the metric and miss the goal.

Structured output makes review practical. Ask for a table or JSON with one row per page or query, a field for the recommendation, a field for the evidence from your data, and a field for confidence or assumptions. A reviewer can then sort, filter and spot-check instead of reading prose. Keep batches small enough that a person can actually check them; the review is where errors get caught.

Finally, the output must pass Google's quality bar. The spam policies define scaled content abuse as generating many pages primarily to manipulate rankings, "no matter how it's created." Use automation to improve and maintain pages, and treat any plan to generate large numbers of new pages as a separate decision that needs its own justification.

Step-by-Step Guide

Step 1: Pick tasks that are repetitive and reviewable

List the SEO tasks your team repeats most: title and meta description audits, grouping queries by intent, spotting cannibalization between pages, summarizing crawl errors. Choose the ones where a person can check the output quickly against a source. Skip tasks where Claude would have to supply the facts, such as estimating traffic for pages that do not exist yet. Start with one task and get it right before adding more.

Step 2: Export the data Claude will reason over

Pull the inputs from the tools that measure them: the Search Console Performance report for queries and pages, your crawler for status codes and on-page elements, your CMS for publish dates and authors. Clean the export so each column is labeled and each row is one page or one query. Note the date range and any filters, and include that note in the prompt. Claude can only be as accurate as the data it receives.

Step 3: Write the reasoning prompt

State the goal, the audience of the output, and the decision it will feed, such as "which pages the content team rewrites next quarter." Tell Claude to use only the supplied data, to separate observations from inferences, and to flag anything important you did not ask about. Define the output columns, including evidence and confidence. Explain why each rule exists so Claude can apply it sensibly to rows that do not fit the pattern.

Step 4: Run a Claude content audit on a small batch

Give Claude a batch of pages with their content, titles, descriptions and performance data. Ask it to check each page against a short checklist drawn from Google's helpful content questions, such as whether the title accurately summarizes the page and whether the page adds value beyond what it summarizes. Review every row of the first batch yourself. Adjust the prompt where the reasoning was wrong before scaling up.

Step 5: Run Claude keyword analysis on query exports

Give Claude the query export and ask it to group queries by the problem the searcher is trying to solve, to name the page that currently ranks for each group, and to flag groups where two pages compete. Ask it to mark groups with too little data to judge instead of guessing. Check a sample of groupings against the actual search results. Use the output to plan updates and consolidations, and to spot real gaps.

Step 6: Triage technical issues from a crawl

Feed Claude the crawl export with status codes, redirects, canonical tags and traffic per URL. Ask it to rank issues by likely impact using the traffic data you supplied, and to explain the ranking for each. Have a developer confirm the top items before anything is fixed. Claude's value here is sorting and explaining a long list; the developer's is confirming what the crawl alone cannot show.

Step 7: Review, act and record

Treat Claude's output as a recommendation list. A person approves each change, and changes go through your normal publishing process. Record which recommendations were accepted, rejected and why. Feed recurring rejections back into the prompt, since they usually point to missing context or a rule without a reason.

Best Practices

• Supply the numbers and let Claude do the reasoning. Search volumes, rankings and traffic come from tools such as Search Console or Keyword Planner; Claude groups, compares and explains them.
• Ask for evidence next to every recommendation. A row that says "rewrite this title" should cite the data behind it, so a reviewer can check the claim in seconds.
• Keep automatic changes behind a human approval step. Audits can run unattended; edits to live pages should not.
• State the date range and source of every export in the prompt. Claude can then say when a conclusion depends on a period that may not reflect current behavior.
• Keep the constitution's forthright trait switched on by asking Claude to report anything important it noticed outside the checklist. Many of the most useful audit findings come from that line.
• Version your prompts with the data format they expect, so a changed export does not silently break a workflow.

Common Mistakes

• Asking Claude for data it does not have: Requests for search volume, keyword difficulty or current rankings without supplied data produce plausible-looking guesses. Export the data from the measuring tool and ask Claude to reason over it.
• Letting the metric replace the goal: A prompt that only says "include the keyword in every title" invites keyword stuffing, which Google's spam policies name as a violation. State the goal, accurate and useful titles, and let the keyword follow from it.
• Reviewing too little: Spot-checking one row in a large batch lets systematic errors through. Review every row of the first batches and sample generously after that.
• Using audit output to justify mass page creation: Gaps found in keyword analysis are a reason to consider new pages one by one. Generating many near-identical pages from a keyword list is the pattern Google calls scaled content abuse.
• Discarding Claude's caveats: A note that the data is too thin to judge is a finding. Treat it as a reason to collect more data before acting.

References

• Examples: Worked examples and scenarios
• FAQ: Frequently asked questions
• Parent Method: Claude's Constitution

Related Skills

• Claude Code Constitutional Guardrails for SEO
• Claude AI Long Tail Keywords by User Need
• Claude Content Optimizer: Evaluate Against the Constitution

Sources

• Anthropic: Claude's Constitution
• Search Console Help: Performance report
• Google Search Central Blog: Guidance about AI-generated content
• Google Search Central: Using generative AI content
• Google Search Central: Spam policies
• Google Search Central: Creating helpful, reliable, people-first content
• Google Ads Help: Use Keyword Planner

Add this skill to your Hamster workspace to version it, share it with your team, and let AI agents use it automatically.