Claude Code for SEO: the agent alone vs with Serpkin

A coding agent on its own can do most of the on-page and technical SEO work on a site: audit the code, write titles and descriptions, add canonicals, JSON-LD, a sitemap and robots.txt, fix content that only renders in the browser, and add redirects. What it can't do by itself is see outside your repo: how many people search for what, where you rank, what Search Console says, who Google's AI Overviews cite, whether last month's fix moved anything, or whether a preview deploy broke a canonical. Serpkin supplies that missing context and memory. This applies the same way to Claude Code, Codex and Cursor.

We make Serpkin, so we have a stake in the answer. That is why this page also says when the agent alone is enough, and shows a do-it-yourself setup that closes part of the gap for free.

#Agent alone vs agent with Serpkin at a glance

"Partly" means the agent can do it, but only when you ask, or only with data you wire up yourself.

#What your agent can do alone

Quite a lot. Everything that lives in your code or in the HTML your server returns is within reach of an agent that can read files and run commands:

• Audit the codebase for missing metadata, duplicate titles, missing canonicals, client-only content and pages nothing links to.
• Write metadata per route: titles, descriptions, Open Graph tags and canonicals, from the page content.
• Add structured data (Organization, Article, Product, FAQPage, BreadcrumbList) that matches what is visible on the page.
• Generate a sitemap and robots.txt from your routes, and add permanent redirects when URLs change.
• Check the live HTML with curl, and run Lighthouse locally for performance hints.
• Draft content for pages you describe, with headings and internal links to your other pages.
• Analyze data you hand it. Given Search Console CSV exports, it can find queries in positions 4 to 20, pages with many impressions and a low CTR, and two pages competing for the same queries, then map each URL to the file that renders it.
• Review diffs for SEO regressions before you merge: removed canonicals, a new noindex, renamed routes without redirects.

A single prompt gets you a useful first pass. For example:

Audit this site for technical SEO. For every public route:
1. Check the server HTML (curl the production URL) has the H1 and main copy.
2. Make sure it has a unique title, a 140–160 character description
   and a self-referencing canonical.
3. Add JSON-LD that matches the visible content.
4. Generate app/sitemap.ts and app/robots.ts from the routes.
List what you changed, file by file, and anything you couldn't check.

For a brand-new site, this foundation pass is most of what matters. Google says there are no extra requirements to appear in AI Overviews beyond being indexed and eligible for a snippet, so the same basics cover AI search too. Our Next.js SEO checklist is a good reference to hand your agent.

#What it can't see

The limits aren't about skill. They are about data the agent doesn't have and time it doesn't span. A session starts, works on your repo and ends.

Search demand

The agent can guess what people search for, but it has no search volume, difficulty or intent data. It can write a great page for a query nobody types, or pick "project management software" over the "kanban board for freelancers" query you could actually win. Ask it for search volumes anyway and it may produce plausible-looking numbers. Treat any metric without a named source as made up.

Rankings and SERPs

Some agents can search the web, but a web search tool is not a rank tracker. It has no fixed country, language or device, and no history, so "we're on page two" one day can't be compared with the next.

Search Console

Your real queries, clicks, impressions and positions live in Google Search Console. An agent can call the Search Console API only if you set up OAuth or a service account and give it the credentials. Most setups don't, so the agent never sees which pages already get impressions and are close to page one.

AI Overview citations

Search Console counts traffic from AI features inside the normal Web search type. Its newer generative AI performance report shows how often links to your site appeared in AI Overviews and AI Mode, by page, country, date and device, but not by query, and not which other sites were cited instead. Knowing whether an AI Overview appears for a given query, and who it cites, takes SERP data for each query.

Whether a fix worked

Search impact shows up over weeks. Answering "did the new title help?" needs a baseline before the change and measurements 14 and 28 days later, ideally compared with pages you didn't change. The agent won't remember the change, let alone come back to check.

Regressions on preview deploys

A stray noindex, a removed canonical or a sitemap that stops validating usually ships in an unrelated PR. The agent only checks if someone asks, on that branch, every time.

#When the agent alone is enough

Skip extra tools when:

• The site isn't live yet, or just launched. There is no search data to read. Do the foundation pass, submit the sitemap in Search Console, and wait for pages to be indexed.
• Search isn't how people find you. An internal tool, a product sold through sales calls, or a site that lives on referrals needs clean basics, not a research loop.
• You want a one-off cleanup. One audit, one PR, done. The prompt above plus a curl check covers it.
• You already read Search Console weekly and are happy to paste the interesting queries into your agent yourself.

#The do-it-yourself middle ground

Between "agent alone" and a dedicated tool there is a setup many developers build themselves. It closes three of the gaps above with an afternoon of work:

1. Pull Search Console data with a script

Create a Google Cloud service account, add its email as a user on your Search Console property, and call the Search Analytics API with the read-only scope (webmasters.readonly). Each request returns up to 25,000 rows; page with startRow. Write the result to a git-ignored seo-data/ folder and the agent can read fresh data any time, without the 1,000-row limit of the web interface.

2. Keep a change log the agent maintains

## SEO log
After any SEO change, append a line to SEO.md:
date | URL | what changed | target query | clicks, impressions, CTR, position (last 28 days, from seo-data/)
When asked "did it work?", compare the same 28-day metrics now, and say that the result is a correlation.

3. Check previews in CI

#!/usr/bin/env bash
# Usage: scripts/seo-diff.sh https://preview.example.app https://example.com /pricing /blog
set -uo pipefail
preview=$1; prod=$2; shift 2; status=0
tags() { curl -s -m 20 "$1$2" | grep -oiE '<(title|link[^>]*canonical|meta[^>]*(robots|description))[^>]*>' | sed "s#$1#HOST#g"; }
for path in "$@"; do
  if ! diff <(tags "$preview" "$path") <(tags "$prod" "$path"); then
    echo "SEO tags differ on $path"; status=1
  fi
done
exit $status

That is crude (it flags every intentional change too, and preview hosts often send a header-level noindex this won't see), but it catches the removed canonical. What the DIY route still lacks is search demand, rankings over time, AI Overview citations and a ranking of what to do next.

#What Serpkin adds to the agent

Serpkin connects over MCP and gives the agent the outside view and the follow-through. The loop, as the agent runs it:

1. serpkin_next_mission returns one mission: why now, target pages and queries, the files to change, ordered steps with snippets, and acceptance checks.
2. The agent implements it in your repo and runs your tests.
3. npx serpkin check <preview-url> compares the preview with production and fails on regressions (errors, a new noindex, removed titles or canonicals, a broken sitemap or robots.txt).
4. serpkin_submit_result and serpkin_verify_mission record the change and verify it, in preview and again in production after you deploy.
5. serpkin_get_results later shows the baseline and the 14- and 28-day measurements.

npx serpkin          # sign in, detect framework and agent, map routes, connect MCP,
                     # run the first scan and print your first mission
# then in Claude Code (with the Serpkin plugin):
/serpkin:next        # get the next mission and implement it

The MCP server has 28 tools and 5 prompts on every plan, so Codex and Cursor get the same loop without the Claude Code plugin. Missions go beyond the foundation pass the agent can do alone:

• Fix the basics: small title, description, canonical, H1, Open Graph and JSON-LD fixes bundled into one mission instead of six.
• Free tool pages when the top results for a query are calculators, checkers or generators.
• Comparison and alternatives pages for competitors you name or "x vs y" queries.
• Backlink recovery (links pointing at broken pages, lost links) and directory listings that fit your site. You submit listings yourself.
• AI Overview citations: pages where an AI Overview cites competitors but not you.

If a mission doesn't fit, dismiss it with a reason ("not relevant", "disagree") and similar missions rank lower. Serpkin never edits, merges or deploys your site.

What it costs, compared with doing it alone

The agent alone costs nothing beyond the subscription you already pay for, plus your time: an afternoon for the DIY setup above, then reading Search Console and briefing the agent each week. Serpkin Free covers 1 site, 3 missions a month and 25 research credits. The Builder plan is $29 a month (or $290 a year) for 3 sites, 30 missions and 500 credits; Pro is $79 for 15 sites. Research uses credits (a keyword seed is 2); missions, scans, checks, verification and results don't.

#Which setup fits you?

Whichever you choose, check these before you call the work done:

• The production HTML (via curl) has each page's H1, title, description and canonical.
• /sitemap.xml lists only 200, canonical, indexable URLs, and robots.txt points to it.
• No page you want found has noindex in its robots meta tag.
• The sitemap is submitted in Search Console.
• You wrote down what changed and when, so you can look at the result in 4 weeks.

For setup details see Serpkin with Claude Code, or start free.