By Olivia Hart

A live-site test of eight "agentic SEO" workflows caught real problems fast: a page worth 27% of a site's traffic had no internal links pointing to it at all, and 10 "temporary redirect" audit rows traced back to one root cause. The setup, published free on GitHub by Semrush's Faizan Ali, hands defined SEO workflows to an AI agent that pulls its own live data and returns the same analysis every run.

That last part is the point. An agent that produces a brilliant competitor analysis once and an unrecognizable one next month has automated nothing.

What is agentic SEO, exactly?

It's the practice of giving an AI a documented method, live data connections, and your business context — then letting it execute a workflow end to end. It differs from ordinary prompting on two axes:

• The system retrieves its own data via Model Context Protocols (MCPs), so it reads this week's Search Console numbers rather than an export you downloaded last month.
• The method lives outside the conversation, so important checks run whether or not you remembered to ask.

Ali, SEO & AI Search Strategist at Semrush, explained the payoff: "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."

Takeaway: structure, not prompt length, is what makes AI SEO output repeatable.

How the four layers work

The architecture splits into four parts, each with one job:

• Project — a folder holding business context: market, competitors, editorial rules, priority definitions, actions requiring human approval. It loads on every run.
• Skills — written procedures the model loads and follows, turning a good prompt into a repeatable workflow.
• MCPs — server connections pulling live data from Semrush, Google Search Console, Google Analytics, SerpAPI and Firecrawl.
• Prompt — just this run's task, kept short because the other layers carry the rest.

Cost is near zero for small sites. SerpAPI gives 250 free credits per month; Firecrawl gives 1,000. Claude's projects and skills features work on every plan, including free, though free accounts cap at five projects and one custom connector.

Takeaway: the free tier is enough to test this on a smaller site today.

What did the live runs actually find?

These are practitioner-reported results from a single test site, not a controlled study — treat them as illustration, not benchmark.

• Content decay: across a 90-day year-over-year comparison, one 230-click drop turned out to be a deliberate 301 from a planned consolidation, caught by an indexation check. Another page was indexed but missing from the sitemap — a five-minute fix.
• Technical triage: of five reported errors, two were the same javascript:void(0) link on a zero-traffic page. The 10 temporary redirects shared one root cause: internal CTA links missing a trailing slash.
• Competitor growth: a competitor drew most high-intent traffic from just two pages, despite a similar backlink profile.
• Internal linking: four pages brought in 96% of site traffic; the second-biggest page sat in the sitemap and nowhere else.
• AI citations: 32 questions asked across four AI tools produced 693 cited sources. Unless the brand was named in the prompt, it never appeared in answers.
• Migration check: the regression guard found cross-domain and trailing-slash redirects wrongly issued as temporary, plus an approved page removal that never shipped.

Takeaway: the workflows mostly surfaced evidence-backed fixes a human review could plausibly miss.

Where agentic SEO stops working

The comparison table in the source is blunt. To check 50,000 URLs for a status code, a script beats an agent. If a task differs every time you do it, there's no method to write down, and the agent adds overhead with no benefit. One workflow — the content portfolio decision engine — came back mostly useless because the test site had no GA4 conversion tracking set up.

The brief-generation run showed the same limit from the other side: the agent found four things a human brief missed, including three recently emerging search questions, but couldn't decide how to structure the article. A human made that call.

Takeaway: agents handle recurring analysis that needs judgment; scripts handle volume, humans handle structure.

What to watch next

Start small, Ali's guide advises: one MCP connection, one skill, one run — the technical sprint triage proves out fastest against an existing Site Audit. For teams tracking AI visibility, the Semrush AI Visibility Toolkit handles ongoing prompt monitoring and share of voice beyond the citation-gap workflow. The open question is consistency at scale: whether these workflows hold up across verticals and larger URL counts the way they did on one site.