How to Run an SEO Audit With Claude and the Semrush MCP

Most SEO audits stall in the same place. The crawler finishes, the report says there are 3,000-odd issues, and someone has to spend a day turning that wall of errors, warnings and notices into a short list a developer will actually work through. That translation step is where audits lose momentum, and it is exactly the step that a language model connected to live crawl data can speed up.

In October 2026 Semrush published a walkthrough showing how to pull Site Audit results into Claude through the Semrush MCP (Model Context Protocol) connector and question them in plain English. That is useful, but it covers one slice of the job. A real audit also has to judge on-page quality, the backlink profile and the competitive gap, and it has to separate what the AI read correctly from what it guessed.

This guide is the version we use with clients: how to connect the two tools, what data Claude can and cannot reach, a prompt sequence that runs from crawl health to competitors, a verification routine for every AI finding, a scoring model that turns findings into a ranked fix list, and an honest look at API-unit costs. If you want the traditional, tool-agnostic process first, start with our step-by-step SEO audit checklist and come back here to speed it up.

What the Semrush MCP connector actually exposes to Claude

MCP is an open protocol that lets an AI assistant call tools on another service. Semrush's server is listed as an official app in Claude, ChatGPT and Perplexity, and its knowledge base says it also works with Gemini, Cursor, VS Code and Claude Code. The important point for auditors is that the connection reads data; it does not change anything in your Semrush account.

“The connection uses read-only access, meaning your AI assistants can retrieve Semrush data but cannot modify or delete it.”

— Semrush Knowledge Base, Understanding Semrush MCP

The Semrush knowledge base groups what you can query into five families, and Semrush's own audit walkthrough adds Site Audit on top. Here is how that maps to the parts of an audit:

Audit area	Data Claude can request via the MCP	What it cannot do
Crawl health and technical issues	List of completed Site Audit runs, audit summaries (Site Health score, issue counts by severity, audited and unaudited pages), affected URLs per issue, all issues on one page, Site Health across several audits	Start a new crawl, change crawl settings, read Search Console data
On-page	Issues Site Audit already flags (missing or duplicate titles, missing meta descriptions, broken links and so on), keyword and ranking data for the domain	Render your live pages or judge content quality unless you paste the content in
Backlinks	Authority scores, referring domains, backlink quality signals, anchor text	Decide whether a link is manipulative; that needs a human reading Google's policies
Competitors	Keyword overlap, traffic share, competitor backlink profiles, organic and paid metrics	See competitors' private analytics
Traffic and market data	Traffic by source and geography, market trends	Only with a separate Trends API plan

Sources: the Semrush MCP knowledge base page for the data families and the Trends API requirement, and the Semrush Site Audit with Claude walkthrough (published 15 October 2026) for the Site Audit functions.

Two limits shape everything that follows. First, Site Audit access through the MCP only supports GET requests, so Claude works with audits that have already finished. Second, every request spends API units, which makes prompt design a cost decision as well as a quality one.

Connecting Claude to Semrush in about five minutes

You need a Claude account (Semrush says any plan works for this setup) and a Semrush subscription with MCP access. According to Semrush, Site Audit and MCP access come with Semrush One Starter and Pro+, or SEO Classic Pro and Guru, which include 50,000 MCP API units that refresh on your renewal date. Semrush One Advanced and SEO Classic Business need an added API units package.

1. Run a fresh Site Audit first. In Semrush, open your Site Audit project and check the "Last Update" column. If the crawl is older than your last significant release, re-run it, because Claude can only read completed audits.
2. Open Claude's connector settings. On claude.ai or the desktop app, go to the connectors area (Semrush's guide describes Settings, then Connectors; Anthropic's current help page labels it Customize, then Connectors).
3. Add the Semrush connector. Search the directory for "semrush" and add it. Semrush is listed in the directory, so you do not need to paste a server URL.
4. Authorise. Sign in to Semrush if prompted and approve the access request. The knowledge base says the connection uses OAuth or an API key.
5. Check status. A checkmark in the Status column means the connection is live.
6. Test in a new chat. Enable the Semrush toggle from the plus menu and ask, "Give me a list of website audits for example.com in Site Audit." Approve the tool permission when Claude asks.
7. Tighten permissions. If you only ever audit, consider blocking tools you do not use from the connector's settings.

Anthropic's documentation is blunt about connector hygiene, and it applies to official connectors as much as custom ones:

“Custom connectors allow connections to unverified services.”

— Anthropic, Add a connector that isn't in the directory

The same page recommends reviewing tool approval requests carefully, using "Always allow" only for trusted servers, and turning off connectors you are not using. In an agency setting, that means one Semrush seat per analyst rather than a shared login, so the API query log shows who spent which units.

If you are weighing whether to build an MCP server for your own business, rather than consume one, our guide on whether your business should publish an MCP server covers that decision.

The six-prompt sequence we run on every audit

Asking Claude to "audit my site" produces a generic essay. The better approach is a fixed sequence where each prompt narrows the next. These are the prompts we use, adapted from Semrush's examples and extended to cover backlinks and competitors. Replace the bracketed parts.

Prompt 1: establish the baseline

Get [domain]'s Site Health scores for the latest 2 audits in Site Audit. Then summarise the latest audit: date, pages audited vs not audited, and issue counts by errors, warnings and notices. Flag anything that changed by more than 10% between the two audits.

The "pages not audited" number matters more than people think. If a crawl covered 4,000 of 9,000 URLs because of a page limit or blocked sections, every percentage that follows is about a sample, not the site.

Prompt 2: group errors by root cause, not by issue name

For every error-level issue in the latest audit, get the affected URLs. Group the issues by likely shared cause (template, CMS setting, redirect rule, plugin, content). For each group give: the issues included, number of URLs, the URL pattern, and the single fix that would clear most of them.

This is where Claude earns its keep. Site Audit might report 600 broken internal links, 600 4XX pages and 600 "pages with only one incoming link" separately; Claude will often notice they share a /blog/tag/ path and are one template problem.

Prompt 3: on-page issues on pages that matter

Using the domain's organic ranking data, list the 30 URLs with the most estimated organic traffic. For each, get all issues Site Audit found on that page. Show only pages with title, meta description, heading, duplicate content or internal-link issues.

Fixing a missing meta description on a page with no traffic is busywork. This prompt joins crawl issues to ranking value, which is the step most manual audits skip.

Prompt 4: the backlink health check

Summarise [domain]'s backlink profile: referring domains, authority score, anchor text distribution (top 20 anchors with share), and any referring domains that look low quality. Then list the 10 strongest referring domains pointing to pages that now return 4XX errors in the latest Site Audit.

The second half of that prompt is the valuable part: links to dead URLs are the cheapest wins in any audit because a single 301 redirect recovers them. Our backlink audit guide explains how to judge the rest of the profile by hand.

Prompt 5: the competitive gap

Compare [domain] with [competitor 1] and [competitor 2] on organic keyword overlap. List 25 keywords where both competitors rank in the top 10 and we rank 11-30 or not at all, with search volume and keyword difficulty. Then compare referring domain counts for the three sites.

Prompt 6: the fix list

Combine findings from this conversation into a table: issue group, URLs affected, estimated impact (high/medium/low with a one-line reason), estimated effort (hours), owner (developer, content, SEO), and dependencies. Sort by impact divided by effort. Do not include any issue you have not retrieved data for in this chat.

The last sentence of prompt 6 is not decoration. It tells Claude to stay within the data it actually pulled, which cuts down the plausible-sounding filler that language models are prone to. For more prompt templates outside the audit context, see our 40 ChatGPT and Claude prompts for SEO.

Where Claude reads the data well, and where it guesses

After running this workflow on a range of sites, the pattern of strengths and weaknesses is consistent. Claude is very good at structure: grouping, counting, spotting shared URL patterns, comparing two audits and writing fix tickets. It is weaker at anything that requires seeing the page as Google sees it, or knowing your business context.

Task	Claude's reliability	Why	Human check
Counting and grouping issues	High	Works directly on returned data	Spot-check two groups against the Site Audit UI
Comparing Site Health between audits	High	Simple arithmetic on returned scores	Confirm both audits used the same crawl scope
Explaining what an issue means	High	Well documented issue types	None, unless it contradicts Google's docs
Guessing the root cause	Medium	Infers from URL patterns, cannot see your CMS	Developer confirms before a ticket is raised
Judging whether a page should be indexed	Low to medium	Depends on business intent	SEO lead decides, using Search Console
Calling links "toxic"	Low	Metrics are proxies, not proof of manipulation	Manual review against Google's spam policies
Estimating traffic impact of a fix	Low	No access to your conversion data	Use Search Console and analytics

Two examples illustrate the gap. A Site Audit "noindex" flag means nothing on its own; Google itself says not every URL needs to be indexed.

“You should not expect all URLs on your site to be indexed, only the canonical pages, as described above.”

— Google Search Console Help, Page indexing report

Claude might recommend removing noindex from filtered category URLs because the audit flagged them, when noindex was the correct decision. Similarly, a low authority score for a referring domain is not evidence of link spam. Google defines link spam by intent: "creating links to or from a site primarily for the purpose of manipulating search rankings" (Google spam policies). A metric cannot see intent, so a person has to.

A ten-minute verification routine for AI audit findings

Treat every Claude finding as a hypothesis until it passes these checks. On a typical audit this takes ten to twenty minutes and catches the handful of errors that would otherwise go to a developer.

1. Sample three URLs per issue group. Open them in a browser and view source. Confirm the problem exists on the live page today, not just in a crawl from last week.
2. Cross-check with Search Console. Use the URL Inspection tool on one URL per group and compare with the Page indexing report. If Search Console says a page is indexed and serving fine, a crawler warning may not matter.
3. Check crawl scope. Confirm the audit covered the sections Claude is drawing conclusions about. A conclusion about "the blog" from a crawl that only reached 200 of 1,500 posts is weak.
4. Ask Claude for its evidence. Follow up with: "For each recommendation above, show the exact data point you retrieved that supports it." Anything without a data point gets removed or marked as an assumption.
5. Check performance claims separately. If Claude makes claims about speed, test the template in PageSpeed Insights. Google's Core Web Vitals guidance uses LCP within 2.5 seconds, INP of 200 milliseconds or less and CLS of 0.1 or less, measured at the 75th percentile of page loads.
6. Read rewrites before publishing. If Claude drafts new title tags or meta descriptions, read them against the page. Google says it may change title links when the title does not reflect the page, and that snippets are "primarily created from the page content itself" (Google on meta descriptions).

The final output of verification is simple: each finding is marked confirmed, adjusted or dropped. Only confirmed and adjusted items move into prioritisation.

Scoring fixes with an impact, reach and effort model

Claude's own priority labels are a good first draft, but they are not consistent between runs. We give it a fixed formula instead, so the ranking is repeatable and a client can see why item 4 sits above item 9.

Priority score = (Impact x Reach x Confidence) / Effort

• Impact (1-5): how much the issue blocks crawling, indexing, ranking or conversion on affected pages. Indexing blockers on money pages score 5; a missing alt attribute on a decorative image scores 1.
• Reach (1-5): the share of valuable traffic touched. Use organic traffic of affected URLs, not URL count.
• Confidence (0.5, 0.8 or 1.0): 1.0 if verified on the live page and in Search Console, 0.8 if verified on the page only, 0.5 if still an assumption.
• Effort (1-5): 1 is a CMS setting change; 5 is a template rebuild needing a sprint.

Priority band	Score	What happens
Do this week	8 or above	Ticket raised immediately, owner named
Next sprint	3 to 7.9	Scheduled with the development backlog
Batch later	1 to 2.9	Grouped with similar small fixes
Ignore or monitor	Below 1	Logged, not worked

You can paste this model straight into prompt 6 and ask Claude to apply it. For a wider view of how audit output becomes a quarterly plan, see our guide to building an SEO roadmap that actually ships.

Worked example: a 1,400-page homewares store

This is a hypothetical example built to show the workflow and the arithmetic. The numbers are illustrative, not client results.

The store runs on a hosted ecommerce platform with a blog. Its latest Site Audit crawled 1,400 pages, reported a Site Health score that had fallen six points since the previous crawl, and listed 41 error types, 58 warning types and a long tail of notices.

Prompt 1 showed that the drop came almost entirely from a jump in 4XX errors and broken internal links between the two audits. Prompt 2 grouped 312 broken internal links and 96 4XX pages under one pattern: products discontinued in a range refresh, still linked from older blog posts and "related products" blocks. A second group of 180 duplicate title tags came from paginated category URLs.

Prompt 3 joined issues to traffic. Of the 30 highest-traffic URLs, nine had duplicate or truncated titles, and four category pages had more than 150 internal links to dead products. Prompt 4 found that 14 referring domains, including two trade publications, pointed to discontinued product URLs. Prompt 5 found 25 keywords where both main competitors ranked top 10 and the store sat between positions 11 and 30, mostly buying guides.

After verification, the scored list looked like this:

Fix	Impact	Reach	Confidence	Effort	Score
301 discontinued product URLs to closest category or replacement	4	4	1.0	2	8.0
Remove dead products from "related products" logic	3	4	1.0	2	6.0
Rewrite nine high-traffic titles	3	3	0.8	1	7.2
Self-referencing canonicals and unique titles on pagination	2	2	0.8	3	1.1
Brief five buying guides for the competitor gap	4	3	0.5	4	1.5

The redirects scored highest because they solved three problems at once: 404s, broken internal links and the 14 referring domains pointing at dead pages. The content gap scored lower, not because it matters less, but because confidence was only 0.5 until someone checks search intent by hand. That is the kind of trade-off a scoring model makes visible.

Unit budget for this example. Semrush says most requests cost between 100 and 1,000 units and pulling a past audit costs 10,000. If the six prompts trigger roughly 15 requests at an average of 500 units, that is about 7,500 units, or 15% of a 50,000-unit monthly allowance. Pulling one historical audit for comparison would add another 10,000. Exact costs depend on request type, so check the API query log after your first run.

API units, plan limits and other costs to plan for

The cost of this workflow is mostly unit spend and analyst time, not software. Here is what Semrush documents, as of October 2026:

• Included units: 50,000 MCP API units on Semrush One Starter, Semrush One Pro+, SEO Classic Pro and SEO Classic Guru, refreshing on your subscription renewal date.
• Per-request cost: Site Audit requests use 100 to 10,000 units depending on type; a past-audit pull costs 10,000.
• Extra volume: unit packages of 2M, 5M, 10M or 20M units, and Advanced or Business plans need a package for MCP access.
• Traffic and market data: needs a Trends API plan.
• Monitoring: remaining units sit on the API Units tab of your subscription page, and past requests in the API query log.

Three habits keep spend predictable. Narrow every query (specify the database, the number of results and the date). Pull the audit summary once and reuse it in the same chat instead of asking again. And avoid historical audit pulls unless you need a trend.

The non-software cost is review time. In our experience, the AI-assisted version of an audit moves effort from data wrangling to verification and decision-making, which is where senior time should go anyway. If you would rather hand over the whole process, our team runs these audits as part of ongoing SEO services; to scope it, request a custom quote.

Gaps this workflow leaves open, and how to fill them

The Semrush MCP is strong on what Semrush measures. Several audit areas sit outside it, and skipping them would make the audit incomplete:

• Search Console data. Indexing status, actual clicks and queries come from Google, not Semrush. Export the Page indexing report and Performance data and paste summaries into the same chat for Claude to join with crawl data.
• Field speed data. Site Audit flags performance issues, but real-user Core Web Vitals come from the Chrome UX Report. Our page speed optimisation service works from field data first.
• Rendering problems. JavaScript-dependent content needs checking with URL Inspection's rendered HTML. Our technical SEO guide covers the manual checks.
• AI search visibility. A clean crawl does not mean AI assistants cite you. Track that separately with an AI visibility dashboard built in Semrush, and check whether assistants describe your brand accurately, since many lag behind rebrands; see our piece on the LLM positioning lag.
• Branded results. If an AI Overview appears for your brand name, audit what it says; our guide to AI Overviews on brand searches explains how.

Many of the fixes this audit surfaces also help with AI answers. Clean indexing, clear titles and fast pages are covered in our list of SEO practices that work for both Google and AI search. And if the competitor prompts reveal pages AI tools cite instead of yours, the next step is AI citation outreach.

Questions SEOs ask about auditing with Claude and Semrush

Can Claude run a new Site Audit crawl for me?

No. Semrush states that the MCP connection cannot start new audits or change audit settings. Start or schedule crawls in the Semrush dashboard, then ask Claude about the completed result.

Do I need a paid Claude plan?

Semrush's walkthrough says a Claude account on any plan works for this setup. Anthropic notes that on the Free plan you can add one custom connector; Semrush is in the directory, so check your plan's connector limits.

How many audits can I analyse with 50,000 units?

It depends on request type. Most requests cost 100 to 1,000 units, but a past-audit pull costs 10,000, so five historical pulls would use the whole allowance. A focused audit on the latest crawl usually needs far fewer units.

Is it safe to let Claude see client data?

The Semrush connection is read-only. Still, follow Anthropic's advice: connect only trusted servers, review permission requests, and turn off connectors you are not using. Check your client contracts for rules about AI processing.

Will Claude replace a crawler like Screaming Frog?

No. Claude analyses data a crawler collected. It does not crawl through this connector, so you still need Site Audit or another crawler for raw data.

Can Claude write the fixes, not just list them?

Yes, for text fixes such as title tags, meta descriptions and structured data drafts. Review every draft against the live page before publishing.

Running your first Claude-assisted audit this week

Pick one site you know well, so you can tell when Claude is wrong. Run a fresh Site Audit, connect the Semrush connector, and work through the six prompts in a single chat. Verify each issue group, score the confirmed ones, and send the top three to a developer. Then compare the effort with your last manual audit; the time saved shows you whether the approach suits your team.

If you would like experienced eyes on the findings, or want the whole audit and fix programme handled for you, get a free, no-obligation quote from My SEO Experts.

Further reading on seo audit with claude

• Building an AI Visibility Dashboard With Semrush
• Google AI Overviews Now Show for Most Major Brands: What to Do
• Zero-Click Marketing: How to Execute and Measure It

Sources for this Claude and Semrush audit guide

1. Semrush Blog: How to run an SEO audit with Claude and Semrush (15 October 2026)
2. Semrush Knowledge Base: Understanding Semrush MCP
3. Anthropic: Add a connector that isn't in the directory
4. Google Search Console Help: Page indexing report
5. Google Search Central: Influencing title links in Google Search
6. Google Search Central: How to write meta descriptions
7. Google Search Central: Spam policies for Google web search
8. web.dev: Web Vitals