What does the new Semrush MCP library cover?

Semrush published a 16-prompt library that runs through Anthropic's Model Context Protocol. The library connects Semrush data to Claude, ChatGPT, Cursor, VS Code, Gemini, and Perplexity through one endpoint.

The endpoint is https://mcp.semrush.com/v2/mcp for non-Claude or non-ChatGPT clients. Users paste prompts, swap their domain and country into the {brackets}, and receive structured tables back from the model.

The 16 prompts split into four workflows:

• Keyword strategy
• Competitive intelligence
• Content creation
• Diagnostics

Each workflow chains two or three follow-up prompts in the same conversation. In a documented test run, a 36% traffic drop, a competitor at parity, and a duties page losing to a single law firm article all surfaced through the prompts after the author ran the diagnostics and competitive workflows in sequence. Caveat follows: a working integration proves plumbing, not plan fit.

Which plans qualify, and what do API units cost?

MCP access requires one of four Semrush plans. Each plan ships with 50,000 API units. The four qualifying plans are:

• Semrush One Starter
• Semrush One Pro+
• SEO Classic Pro
• SEO Classic Guru

Traffic and Market prompts pull from a separate Trends API subscription. Users without the Trends add-on see low_data in bounce rate, time on site, and pages per visit. The model falls back to organic-traffic estimates and labels the gap.

API unit cost scales with the data a request pulls, so a 50-keyword US pull costs less than a global pull spanning every Semrush regional database. Users on the Guru plan who run all 16 prompts in one session will burn through unit headroom faster than consultants running diagnostics on a single domain. Caveat follows: budget unit cost against the question you are actually answering, not the volume of tables you would like to print.

How does the keyword workflow find real demand?

The first keyword prompt returns an eight-row table of clusters ranked by combined monthly volume. The second prompt finds keywords competitors rank for that the target domain does not. It splits the result into missing and weak shared rows. The third prompt estimates intent mix per cluster as percentages and ranks the clusters by business value.

Excluded domains include any with Competitor Relevance of 0.00 and organic traffic above 10 million, which removes YouTube, Reddit, Wikipedia, and Facebook from a typical niche audit. The library labels intent shares as estimates. Users must spot-check two SERPs before trusting the roadmap. Caveat follows: the same exclusion list, if misapplied, hides legitimate vertical competitors from a niche audit.

Which competitive intelligence prompts move beyond dashboards?

The competitive workflow ranks the top 10 organic competitors by Semrush's Competitor Relevance score. A traffic-share prompt computes total market traffic, the market leader, and top-two concentration. It accepts multiple domains in a single call. A growth prompt pulls 12 months of organic traffic history per domain and ranks the cohort by absolute change.

Any domain with traffic_12m_ago below 100 is flagged as low_data to suppress misleading percentage swings from tiny bases. The library ships a watch-list prompt and a ruleset prompt. The ruleset generates at least eight alert rules: five for competitor gains, two for own losses, and one optional. The MCP writes the rules; the user implements them inside Semrush as Position Tracking campaigns, because the connector cannot create alerts on its own. Caveat follows: the ruleset is a starter framework, not a deployed monitoring system, and requires a manual handoff to Semrush before any threshold fires.

How does the content workflow decide what to refresh?

The content workflow starts by ranking the 20 lowest-traffic pages against a six-month-old snapshot. Each row records traffic change, top keyword, current position, and position six months prior. A second prompt combines current page content with keyword baselines. It flags thin, outdated, or missing-keyword issues against defined rules.

Pages count as thin when word count drops below 800. Pages count as outdated when content references a year more than two years before the current calendar year. A demand-mapping prompt returns a 12-row table across five intent clusters:

• Branded
• Category
• Use-case
• Comparison
• Transactional

A gap-to-plan prompt then classifies the top three gap clusters as refresh or create, with a one-sentence case per row. Caveat follows: one bad competitor domain in the input set can pollute every downstream cluster in the workflow.

What does the diagnostics workflow tell you about traffic losses?

The diagnostics workflow begins with a 50-row quick-win table. The table covers keywords ranking in positions 5 through 20. A row counts as high when position sits between 5 and 10, volume reaches 500, and keyword difficulty stays at or below 60. The Semrush guide's author noted the high tier often surfaced low-value dictionary queries while real wins hid in the medium tier.

The traffic-loss prompt compares the target domain against five direct organic competitors during the drop window and returns an explicit industry-wide or site-specific verdict in the same response. If three or more competitors also show a negative percentage change, the prompt labels the event industry_wide_signal: true and points the user to Semrush Sensor above 3.5. If fewer than three drop, the prompt concludes the loss is site-specific. Semrush Sensor sits outside the MCP, so the volatility check remains manual. Caveat follows: a clean industry-wide verdict still requires a human to read the Sensor timeline before any action is taken on affected pages.

What should SEO teams monitor next?

The Semrush MCP library turns Semrush data into conversational queries, but the model still does not act. Alerts come out as rules, not deployed campaigns. Workflow order matters: diagnostics fed by an empty competitive pass will produce generic thresholds that need re-calibration.

Watch for Trends API subscription gaps in the engagement columns. Watch for low_data flags that hide real changes behind small baselines. Per the Semrush MCP guide: "Pick a use case from the Semrush MCP prompt library and find the data hiding under your nose." The library exposes the data, but the analyst still owns the interpretation.