Semrush splits an agentic SEO setup into four layers, and runs eight workflows on a live site with Claude, MCP and Search Console

Semrush published a piece on 18 September by Carlos Silva, an SEO writer at the company, and Faizan Ali, its SEO and AI Search Strategist. The article sets out a four-layer method for handing an SEO workflow to an AI agent and describes eight workflows the authors say they ran against a live site.

The four layers

Semrush frames agentic SEO as splitting the work into a project, skills, connectors, and a prompt. The project holds the business context the agent needs on every run: business plan, content guidelines, branding, and ideal client profile. Skills carry the method for each step, the checks that get applied whether or not the user thought to ask. Model Context Protocol servers pull live data from Semrush, Google Search Console, or Google Analytics instead of waiting for someone to paste an export. The prompt only carries the current ask.

Faizan Ali explains why the split matters: "With one long prompt I'd be re-explaining everything every time, and the longer it gets the more Claude drifts, loses context, and starts filling gaps with hallucinated stuff. Plus if something breaks, I know exactly which layer to fix instead of debugging a wall of text."

What separates it from an AI-assisted chat

Semrush contrasts agentic SEO with two neighbours. Workflow automation is perfect at repeating steps you decided in advance, and breaks on any input the rules did not anticipate. AI-assisted SEO depends on what you paste into the chat and how you phrase the question. Agentic SEO retrieves its own data, follows a method that lives outside the conversation, and returns the same analysis every time the workflow runs.

The company is honest about the limit: "If a task differs every time you do it, there is no method to write down, and an agentic workflow adds overhead with no benefit." A 50,000-URL status-code check belongs in a script, not in an agent.

The eight workflows and what Claude's free tier can run

The article names eight recurring workflows the authors built as Claude skills, including a technical triage, a competitor investigation, a search opportunity brief, and AI citation mapping. Semrush notes that as of September 2026 the free Claude plan is limited to one custom connector, and lists four of the eight workflows a reader can run inside that limit using only the Semrush MCP: technical triage, competitor investigation, search opportunity brief, and AI citation mapping. The other four workflows need a paid Claude subscription.

The stack outside Claude uses Semrush's seven-day free trial, Google Search Console, Google Analytics, SerpAPI, and Firecrawl. SerpAPI's free plan gives 250 credits a month, Firecrawl's gives 1,000. Semrush says that is enough for most smaller sites.

The stack is Semrush's own product, so read the recommendations as such: a company selling the data connector at the centre of the setup is describing a setup that depends on that connector. The layer split, and the argument for pulling live data instead of pasting exports, are useful on their own.

The real value here is a working method for recurring SEO work that a team can copy: name your context once, write the checks in a skill, plug in a live data source, keep the prompt to what changed. A team that has been rewriting the same audit instructions in every chat has a documented alternative to try.

Source

Semrush: What is agentic SEO? 8 workflows run on a live site, 18 September 2026, by Carlos Silva with Faizan Ali.