Semrush gave Claude eight SEO tasks on a live site and wrote up what happened

Semrush ran eight typical SEO tasks through Claude on a live site and wrote down what came of it. The account is on the Semrush Blog.

The idea is simple: instead of writing a long prompt every time, you describe the procedure once, connect the agent to Search Console, analytics and Semrush, and then give it a short task.

The four layers the agent stands on

The first layer is the project: a document with business context, competitors and priorities. The second is skills — procedures written once and polished afterwards.

The third is MCP connections, live access to Semrush, Google Search Console, Google Analytics, SerpAPI and Firecrawl. The fourth is the prompt itself, which can be a single line, because the system already knows the rest.

A full set of connections needs a paid Claude plan: the free one holds only one custom connector.

For a Ukrainian agency or in-house team this is a familiar pain: the analysis done on Monday is repeated differently on Thursday because someone forgot one of the steps. A written procedure removes exactly that. If you are still choosing what to build such a stack on, see our review of AI SEO tools for small business.

What the agent found on a live site

The examples in the piece are down to earth, and that is the most interesting thing about them.

• A content decay diagnosis surfaced three pages losing traffic: one was a deliberate consolidation, another had simply dropped out of the sitemap. Nothing had to be rewritten.
• A technical sprint grouped errors by cause: ten "temporary redirects" all traced back to internal links missing a trailing slash.
• An internal link hunt showed that the site's second-largest revenue source had zero internal links, and suggested specific sentences to place them in.
• A release regression check caught two critical failures: temporary cross-domain redirects and a page that was never removed during a migration.

Note the character of the findings. This is not "a strategy for the quarter" but blind spots: what a person stops seeing because they are looking at the same report for the tenth time.

An agent fits that niche well, because it does not get bored.

Where the limits are

Semrush states honestly that the content portfolio procedure, which sorts every URL into "keep, update, merge, redirect or investigate", requires properly configured conversion tracking.

Without it the agent will recommend selectively: it sees traffic but does not see which pages bring in money. The same goes for citation mapping: 32 prompts are a sample, and you still have to ask whether that sample resembles your customers' questions.

A separate word about what the original does not say outright. All the examples come from one site, and this is a publication by a company selling a tool that is built into the scheme.

There is no comparison with the human effort on the same tasks, so "time saved" here remains an impression without a single figure.

For our market there is one more practical point: MCP connectors to Search Console take ten or fifteen minutes to set up, while access to client accounts usually takes an agency weeks to agree, and that bottleneck is not going anywhere. Teams already reporting in GA4 will feel the difference sooner — we covered the groundwork in our guide to GA4 reports for SEO.

I'd start with the dullest procedure of the lot: the post-release regression check. That is where mistakes are expensive and where a human tires fastest.

What struck me most was the example of a page with no internal links at all that brings in the second-largest revenue. I wonder how many of those are sitting on sites where internal linking was last touched two years ago.

Related reading: SGE and AI Overview: Optimising for AI Search · SEO Website Promotion