Claude for SEO: 3 Workflows I've Automated as a Practitioner

A while back I gave a talk for the Brand SEO Hub community on how I actually use Claude in my day-to-day SEO work at SE Ranking. This post is the written version of that session — the same three workflows, the same setup choices, just cleaned up so you can copy what works and skip what doesn't.

One disclaimer up front, the same one I opened the talk with: there is no magical prompt, no perfect MCP, and no universal workflow that solves SEO for everyone. The best automation is the one that removes a routine task you personally keep doing. Treat what follows as a starting point, not a recipe.

Want the full workshop recording?#

The unlisted video link goes out to my newsletter subscribers. Drop your email below to get it — or follow me on LinkedIn (and message me to ask for the recording / workshop language - Ukrainian), where I share the same updates.

Claude vs. Claude Code: Which One I Actually Use#

Most SEOs I talk to are using Claude through the chat or the desktop app. That's fine for one-off questions, but for any workflow that involves multiple data sources, multiple steps, or writing files, I use Claude Code.

Three reasons:

• No tool-use limits in the middle of a run. In Claude chat I regularly hit the tool-use cap and have to click "continue" five or six times during a long workflow. Claude Code just keeps going.
• The agent is genuinely autonomous. It writes its own to-do list, opens files it needs, creates and deletes them, and works through the steps without me babysitting.
• The CLAUDE.md file. This is the one feature I'd bring up first. It's a markdown instruction file scoped to whatever folder you open the agent in. You write down which MCPs to use, how to handle data, which filters to apply, where to save results — and Claude Code picks it up automatically every time you launch. There's a global version for cross-project rules too.

The same end-to-end backlink workflow I'll describe below took ~10 minutes in Claude Code. The same flow in Claude chat took close to 30 minutes with multiple "continue" clicks.

What MCP Is, in One Paragraph#

MCP (Model Context Protocol) is just a standard way to plug external tools into the model. Instead of pasting a CSV into the chat, you install an MCP server for, say, Google Search Console, and Claude can query it directly. Most useful MCPs for SEO are local — they authenticate either through a service account or a normal OAuth login, and the providers ship step-by-step setup guides. The one-time chore is usually a Google Cloud Console project plus a config file. After that, the model has hands.

A practical token tip: don't enable every MCP you own inside every project. Each connected tool eats context just by existing. Spin up a project per workflow with only the MCPs that workflow needs.

With that out of the way, here are the three workflows.

Workflow 1: Automated Content Briefs#

This was the first workflow I built and the one that paid for itself fastest. A proper content brief — the kind a writer can actually rank with — usually takes me hours: open the SERP, read the top 10, extract subtopics, build a keyword cluster, check related questions, pull what we already rank for, structure an outline.

Here's how I have Claude Code do it:

1. Pull the top 10 SERP results for the target keyword via a SERP MCP.
2. Scrape each competitor URL — headings, word count, schema, internal linking patterns.
3. Expand the keyword cluster using related-keywords, questions, long-tails, and similar-keywords endpoints. I cap the list at 50 keywords because Claude later builds a master keyword strategy from them — primary keyword, secondary keywords, long-tails, and a recommended usage frequency for each.
4. Connect to Google Search Console via MCP and check which of my own pages already rank for those queries.
5. Map those existing rankings to specific sections of the new outline as internal-link recommendations, with anchors pre-suggested. I prioritize linking from service / product / feature pages first, not blog-to-blog.
6. Write everything to a structured brief: target intent, title patterns, H2/H3 outline, must-cover entities, internal link plan, FAQ block, word-count target.
7. Save to a Google Doc or markdown file.

The output isn't 100% perfect — nothing automated is — but it's a 70–80% draft I can finalize in 15 minutes instead of three hours. Because it lives as a saved Claude Code command, the next keyword is just a variable swap.

Workflow 2: Technical SEO + AI Bot Traffic Analysis#

This is the workflow I get the most questions about, and it's the one most teams aren't doing yet. The premise is simple: AI crawlers (GPTBot, ChatGPT-User, ClaudeBot, PerplexityBot, etc.) are now a meaningful slice of your traffic, and almost nobody is correlating their behavior with what's actually performing in Google.

I point Claude Code at a server log file and a Google Search Console export for the same date range, and have it:

• Segment all crawler hits by user agent.
• Pull clicks and impressions from GSC via MCP for the same range.
• Run a statistical correlation (Spearman) between AI-bot event volume per URL and GSC clicks per URL.
• Surface the anomalies — pages with high AI-bot activity but low Google clicks, and the inverse: high Google clicks but barely any AI-bot interest. Both lists tell you something different about your site.
• Break the analysis down per user agent so you can see whether ChatGPT, Claude, or Perplexity is treating your content differently.

Two things make this work where naïve "ask the model about my logs" attempts fail. First, the agent has filesystem access, so it processes the data in chunks instead of trying to hold it all in context. Second, I always make it show its math — every conclusion comes with the underlying numbers and the formula it used, so I can verify before I act on anything.

This one is more time-consuming to set up than the brief workflow, but the insights you get out of it aren't really available anywhere else right now.

Workflow 3: Link-Building Prospecting (Almost) on Autopilot#

The third workflow is the one I demoed live, and the numbers are real. Inside the project's CLAUDE.md I have an instruction that says: when I type out, run the outreach prospecting workflow. The flow is:

1. Take a target keyword + location/language.
2. Pull the SERP and strip the noise — YouTube, Play Market, social networks, Reddit, Amazon, my own domain, irrelevant niches.
3. Pull the referring domains for the target site (so we know what to compare against).
4. Find the gap: domains that link to ranking competitors but don't yet link to us.
5. Apply quality filters — Domain Rating, traffic, keyword overlap. I keep the filter set deliberately simple; you can tune it to your own thresholds.
6. Run a per-domain check (small sub-agent calls) to validate relevance.
7. Output a clean CSV: prospect domain, the page that earned the competitor link, anchor text, and a suggested outreach angle.

Live numbers from a fintech run during the talk: SERP returned 47 unique referring domains → quality filter cut to 30 → gap pass kept 30 (none were already linking to us) → final filters produced a usable shortlist of 25 prospects. Total runtime in Claude Code: around 10 minutes.

The agent doesn't replace a link builder — it replaces the hour of SERP-staring that has to happen before a link builder can start.

A Few Principles That Cut Across All Three#

• Orchestrate, don't monolith. A "manager" prompt that delegates to focused sub-tasks beats one giant mega-prompt almost every time. Fewer hallucinations, cleaner outputs.
• Build QC into the workflow. Before I publish any AI-assisted output, I run it back through Claude with an explicit checklist: facts verified, no fabricated numbers, internal links make sense, tone matches. QC as its own step is what separates content that ranks from content that earns a manual action.
• Content vs. technical SEO is a false choice. With AI flooding the SERPs with mediocre content, technical foundations are what let genuinely good content actually be found. You need both.
• Re-brief in the target language, don't translate. For multilingual SEO, give Claude the keyword cluster in the target language plus local-market context (competitors, terminology) and let it write natively. Translated content is almost always worse.

How to Start Without Overcomplicating It#

If you've never set up MCP, don't try to wire ten integrations on day one. The trap is real and I see people fall into it constantly.

1. Install Claude Code.
2. Add one MCP. Google Search Console is the highest-leverage starting point — every SEO already has the data and the API is free.
3. Pick one workflow from the three above. Content briefs are the easiest to evaluate because the output is obviously good or obviously not.
4. Save the prompt as a reusable Claude Code command, run it on five real keywords, and compare the result to what you'd have produced manually.
5. Only after that's working, add a second MCP and a second workflow.

Ship one thing, then expand. The afternoon you spend on this pays for itself the first week.