/content-rankings948 words

Semrush Publishes Seven-Step AI Content Workflow With Two Human Gates

Semrush's new AI content workflow pairs MCP-connected search data with two human approval gates taking 40-50 minutes, seven tasks, and five reusable Claude skills.

How to use AI for content creation (step-by-step workflow)
How to use AI for content creation (step-by-step workflow)Etienne Boulanger / unsplash
  • Semrush's workflow uses seven tasks and two human approval gates that take 40-50 minutes combined.
  • Semrush MCP ships with 50,000 API units on several plans for live keyword and SERP data.
  • Google's helpful content guidance, updated October 2026, names four rater qualities: effort, originality, talent or skill, accuracy.
  • Claude skills are available on Free through Enterprise plans as of November 2026, with sharing on Team and Enterprise.
  • The example 'website marketing' research selected a primary keyword at 5,400 searches per month.

Semrush has published a step-by-step AI content workflow built around two human approval gates that take 40-50 minutes combined — a small time cost, the company argues, against a structural error that takes hours to untangle once an article has been drafted around it. The guide runs a piece through seven tasks: research, outline, edit, draft, review, verify, and ship, and it works for blog posts, social captions, scripts, and email alike.

Why most AI drafts fail

A prompt like "Write a 2,000-word SEO-optimized blog post on {topic}" produces clean grammar and sensible structure, but no point of view and no firsthand experience. The failure starts with inputs: a blank prompt gives the model no real search data, no brand standards, and no human review of its decisions along the way.

What does the workflow change?

The core move is connecting AI to live data through Model Context Protocol (MCP), which plugs tools like Semrush directly into Claude or ChatGPT so the AI pulls real numbers rather than guessing. Semrush MCP provides keyword, SERP, organic ranking, and competitor SEO data; several Semrush plans include it with 50,000 API units, and every request draws down that balance. Users without MCP access can run the same research manually in Keyword Magic Tool and Keyword Overview.

The guide assigns each research question a connected source:

  • Search demand — Semrush MCP: search volume, keyword difficulty, related keywords, intent
  • Existing rankings — Semrush MCP Organic Rankings data: keywords a page or domain ranks for, plus positions
  • Competing pages — Firecrawl MCP: full-page content for comparing structures, arguments, and gaps
  • Actual page performance — Google Search Console: queries, impressions, clicks, and changes over time

GSC lacks a one-click MCP connection, so the recommended path is exporting Performance data and uploading it, or setting up an API connection for repeatable access.

Standards as files, and where humans stop

The setup separates persistent standards — brand and style guidelines, a brief template, the sitemap, product rules, and a deprecated-tools list — from article-specific context such as target keywords, SERP findings, and expert input. Stored as files, these rules are written once. In Semrush's own workflow, this check caught a draft calling its social scheduling tool "Semrush Poster" and corrected it to Social Poster.

Fully automated publishing fails because research errors compound, the guide says: a misread of search intent propagates from outline into every drafted section. The workflow therefore inserts two stops — after research and after the edited outline — before drafting begins:

Research → [Gate 1: approve research] → Outline → Edit outline → [Gate 2: approve outline] → Draft → Review and revise → Verify → Ship

Inside the steps

The research stage opens with a seed keyword or an existing URL. For Semrush's "website marketing" article, the research pass selected "website marketing" at 5,400 searches a month as the primary keyword plus four secondaries, deliberately excluding agency-hire queries like "website marketing company" that belonged on a service page. It also caught an intent mismatch: Semrush tagged the keyword commercial while the live SERP showed mostly informational guides, so an informational article was planned. Semrush has since updated the tag.

Outline editing takes about half an hour and is, in the guide's words, the cheapest point to fix structure. In the "website marketing" example, review moved a UX section from position 12 earlier, so the first four tactics stayed focused on-site before shifting to social and third-party channels, and promoted Digital PR from a subheading under Guest Blogging to a standalone section.

Verification against Google's 2026 guidance

The review stage leans on Google's updated helpful content guidance from October 2026, which names four qualities search quality raters look for: effort, originality, talent or skill, and accuracy. Google's generative AI content guidance, also updated that month, says a person should fact-check AI-generated content before publication — a rule the workflow operationalizes with a verification skill that flags claims, statistics, links, and product details for manual checking. The model that wrote a claim, the guide notes, is a bad judge of whether it is correct.

That caution proved itself during editing: AI flagged Semrush's Claude skills plan details as wrong, but Anthropic's help center confirmed the draft was right — skills are available on the Free plan, and Team and Enterprise users can share them directly. Taking the flag at face value would have published wrong plan details.

Updates, skills, and limits

Updates differ from new URLs: the starting point is an existing URL plus performance data, the job is diagnosing what broke, and the main risk is disrupting a page that already ranks. The advice is to leave headings that support current rankings alone and change only what the diagnosis calls for.

Once the workflow runs by hand, each stage becomes a Claude skill — as of November 2026, available on Free, Pro, Max, Team, and Enterprise plans with code execution enabled. The guide packages five skills (research, brief, draft, verify, update) with defined inputs and outputs, and notes ChatGPT offers skills to eligible Business, Enterprise, Healthcare, and Edu users.

On penalties, the guide is clear: Google does not penalize content based on how it was produced, and the real risk is scaled content abuse — large volumes of low-value pages aimed at manipulating rankings, whatever produced them.

Four things stay human: original opinion, firsthand experience, deciding what matters, and knowing when the SERP data itself is wrong. Teams adopting this setup should start with one content type, run it manually a few times, and watch whether the approval gates survive after the first few good runs — their removal, the guide warns, is where the workflow most often breaks.

via docs.firecrawl.dev (Original)

More from Tom Whitfield

Tom Whitfield

Show full bio

News editor covering regulation, policy and their impact on Search rankings and algorithm updates, generative engine optimization at SERP Journal.

183 articles