How to Use Claude for SEO: Skills, MCPs, and

Usman Akram, who leads SEO and AEO at Modash, mentions casually partway through this 73-minute Semrush video that he keeps eight to ten Claude Code sessions running simultaneously across different projects. Not eight browser tabs. Eight active AI coding sessions, each building or maintaining a different SEO tool or audit script. At that point, the job description stops being "SEO manager" and starts being something we don't have a clean title for yet. That's the world this video is actually describing.

The setup is aimed at people who already know SEO but feel lost when the conversation turns to Claude, skills, MCPs, and automation. Akram's framing is blunt: most guides teach Claude in general without connecting it to actual SEO work. So he structures everything around three specific layers of capability, each suited to a different kind of task, and demonstrates all five use cases live on screen.

Layer one: Skills (your methodology in a markdown file)

A skill, in Claude's terminology, is an instruction document you write once and load before a task. Akram's analogy: an SOP you'd write for a new team member. The document tells Claude your standards, your output format, and your decision rules. Every time you run that task, Claude reads the doc first.

The practical implication is significant. Without a skill, Claude reasons from scratch each time, and three runs of the same audit can produce three structurally different outputs. With a skill, the shape is locked. Only the variables change.

Akram demonstrates two live: a topical authority gap analysis for Modash against competitors Grin, HypeAuditor, and Upfluence, and a content quick wins audit scoped to ten keyword opportunities. The gap analysis output follows his exact spec, a coverage matrix breaking influencer marketing into subtopics like influencer discovery, fake follower detection, and live subscriber count, stacked against each competitor, with gaps defined strictly as areas where at least two of three competitors have coverage and Modash has none. One competitor doing something counts as a bet, not a gap. That's his rule, and the skill enforces it every time.

Building skills doesn't require writing them from scratch, either. Akram's preferred method: do the task conversationally in Claude for 15-20 minutes, then ask it to synthesize the exchange into a skill document. Claude has a built-in "skill creator" skill for exactly this purpose.

Layer two: MCPs (Claude's reasoning on top of live data)

MCP stands for Model Context Protocol, which Akram explains as the USB-C standard for connecting LLMs to external tools. Instead of building a custom integration for every platform, tools publish an MCP and any compatible model can plug in.

The Semrush MCP, which Akram uses throughout, lets Claude pull keyword rankings, organic traffic data, competitor analysis, and backlink data directly into its reasoning process. The value isn't avoiding a few clicks in Semrush. It's applying Claude's intent filtering on top of raw data.

The example he walks through stuck with me. Modash serves brands running influencer marketing programs, not influencers themselves. A keyword like "how much does an influencer earn" contains the phrase "influencer marketing" but gets searched almost entirely by aspiring creators. No filter in Semrush can cleanly separate brand-intent keywords from creator-intent keywords at scale. Claude can, because it understands the distinction contextually. That's the MCP value: not fetching data faster, but reasoning about it in ways a tool's built-in filters can't.

During the forensic ranking audit demo, when Akram asks Claude to investigate why Modash has been losing organic traffic, Claude's output (surfaced live during the session) found the premise didn't hold. The data showed traffic growing year-over-year and recovering significantly from an earlier trough. December showed some page-level declines, concentrated but not spreading. The shape, as Claude described it, was a recovery curve, not a slide.

The moment that made me reconsider something

I went into this video assuming Claude Code was overkill for most SEO practitioners: a developer tool that got rebranded as accessible but still required real technical lift. Then Akram showed his internal linking audit build prompt. It's a massive document, hundreds of words, specifying scoring rubrics, data sources, orphan page definitions, money page criteria. He wrote maybe 20% of it. Claude drafted the rest during a planning conversation, then saved it to Google Docs via the Drive MCP, and Akram copied it over with minor edits. The "coding" part wasn't Akram writing Python. It was Akram describing his logic in plain English, Claude translating it into a working script, and the whole thing running locally on his laptop. The session during the demo ran for roughly 45 minutes and consumed around 111,000 tokens, per Akram's readout on screen. It produced a prioritized internal linking spreadsheet covering orphan pages, underlinked pages, and commercial-priority pages with suggested anchor text and source URLs, built to his exact scoring rubric. Running that by hand would take a full day.

Claude Code separates from skills and MCPs specifically when scale matters. A 100-page website is manageable in the standard web chat. A 100,000-page site needs a script that can crawl without burning LLM tokens at every step. The script does the lifting; Claude handles analysis at the end. Akram has built tools this way that his entire team accesses via login, actual deployed software with a username and password, not just a saved prompt.

The decision tree that nobody handed my generation

People entering the workforce right now are getting handed AI tools with zero framework for when to use what. Every job posting mentions AI fluency, but the actual guidance usually stops at "use ChatGPT for drafts." Akram's three-tier decision tree is the clearest answer to that problem I've seen spelled out:

One-off question with no recurring need: raw Claude chat plus an MCP. Good for a forensic traffic audit you're running once, or keyword research for a new client you're scoping.

Recurring task with a standard process: build a skill. Topical authority gap analysis every quarter, content brief generation, title tag audits on a schedule. Write the process once; run it consistently.

Scale, or a team-facing tool: Claude Code. Anything touching an entire website, anything that needs to run monthly automatically, anything other people need to use.

He adds two meta-rules: start with the task that costs you the most manual hours today, and don't automate anything you haven't done by hand. The second one has real teeth. "Automation will only amplify what you have defined as good based on your experience and judgment," Akram says. "It cannot create good in itself."

For Claude skills applied to business workflows, that principle holds whether you're in SEO or operations: the system is only as good as the mental model it's encoding. A content audit that looks clean in a spreadsheet but lacks the strategic filter of someone who actually knows which pages matter will produce recommendations that don't move rankings. The output quality is bounded by whoever wrote the skill.

This video is specifically for practitioners who already know what they're doing and want to stop re-explaining their methodology to an AI every single time. If you don't have a methodology yet, no amount of skills, MCPs, or Claude Code will manufacture one.

Yuki Okonkwo covers AI and machine learning for Buzzrag.