We Run SEO Delivery on Claude Code and n8n. Here's the Unpolished Version

Three text files rebuilt a client’s website last quarter. One held the technical fixes, one held the content changes, one held the style rules. We fed them to Claude Code in sequence and reviewed the diffs. That’s the whole method, and it took us about a year of failed experiments to arrive at something that simple.

“AI-native agency” is on our homepage, and we’re aware the phrase is doing heavy lifting across the industry right now. Every agency claims it. Almost none will show you the pipes. So this article shows the pipes: the actual stack, four workflows in production, what each one costs to run, and the two things we deliberately keep manual because automating them would be either reckless or illegal in our German market.

A warning before we start. This is a field report, and field reports contain failures. Ours are in here too.

The thirty-second version of MCP

MCP stands for Model Context Protocol, an open standard from Anthropic that lets an AI model talk directly to outside tools and data. Before it, an SEO analysis meant exporting from five dashboards into one spreadsheet before thinking could begin. After it, the model pulls what it needs and you spend your time on the thinking.

That’s the protocol. It is not the interesting part. The interesting part is what an agency builds on top of it, because the protocol is public and identical for everyone. Your prompts, your templates, your review process, and your accumulated campaign scar tissue are what differ.

Why we didn’t build a dashboard stack

Most agency MCP setups we’ve seen connect Claude to the familiar SaaS suspects: rank trackers, backlink databases, crawlers. Reasonable. We went a different direction for a specific reason.

Our work is unusually implementation-heavy. We don’t hand clients a slide deck of recommendations and wish them luck. We ship the fixes: rewritten meta data, corrected schema, redirect cleanups, restructured pages. A dashboard-centred stack is built for analysis. Ours had to be built for shipping code and content, which is why it centres on three things:

Claude Code with spec files. Every site overhaul becomes a set of sequential specification documents. Technical fixes first, content second, style pass third. Each spec is precise enough that the implementation is reviewable line by line. The spec file is the deliverable the client never sees but benefits from most, because it means nothing ships from a model’s improvisation.

n8n as the automation spine. Webhook routing, content pipelines, event ingestion, WhatsApp Cloud API for approvals. Self-hosted, which matters for the compliance section below.

Creative MCPs where they earn their place. Canva and Magnific connect into our social production flow. Image generation and template autofill are exactly the kind of repetitive work that should never touch a senior person’s calendar.

Search Console data comes in through direct exports rather than a connector, and honestly that’s a deliberate choice, covered below.

Four workflows, as they actually run

1. GSC forensics for ranking-loss diagnosis

A German financial services client came to us after their impressions collapsed by roughly 80% across two 2026 algorithm updates. The old diagnostic approach: an analyst eyeballs Search Console charts, forms a hunch, writes a document defending the hunch.

Our approach now: full GSC export, every query, every page, sixteen months of history, processed through structured analysis with Claude. The output mapped 236 query clusters that had gone completely dead, separated them from clusters that merely declined, flagged homepage cannibalization, and surfaced two high-value topics the site ranked for accidentally with no dedicated page.

Time cost: one working session instead of a week. And here’s the honest part — the speed isn’t the point. At the old cost, nobody analyses 236 clusters individually. You’d sample twenty, generalize, and miss the two accidental opportunities entirely. The thoroughness only exists because it got cheap.

What stayed human: deciding which dead clusters were worth resurrecting. That call needs knowledge of German lending regulation and the client’s actual license scope, and no export contains it.

2. Spec-file implementation through Claude Code

The workflow from this article’s opening. Audit findings become numbered FIND/REPLACE instructions and structural specs. Claude Code executes them against the actual codebase. A senior reviews the diff before anything deploys.

Two failure modes taught us the current shape. Early on we let the model audit and fix in one pass, and it fixed things confidently that weren’t broken. Separating diagnosis from implementation, with a human sign-off between them, killed that class of error. The second failure: specs that said “improve” anything. Improve is not an instruction. Every spec line now states the exact current text and the exact replacement, which is tedious to write and boring to review, and boring is what you want in deployment.

3. The anti-AI style gate

Uncomfortable admission for an AI-native agency: unedited model output has a detectable voice, readers have learned to hear it, and Google’s spam systems increasingly punish it. So every piece of content we produce passes through a codified style filter before publication. A budget for em dashes. A cap on the “not X, but Y” construction models adore. Word repetition limits. Sentence rhythm checks, because model prose settles into a metronome and human prose doesn’t.

The filter is itself a spec file, which means Claude applies most of it and a human editor catches what slips. Yes, we’re using AI to remove the sound of AI. We’ve made peace with the irony. The German posts we ship for financial clients read like a cautious advisor wrote them, because in effect one did — the model drafted, the rules constrained, a person judged.

4. Social production with an approval gate

Our content pipeline runs in two modes. The semi-automated version: a working session where strategy, copy, and visuals come together with Canva and Magnific handling template autofill and image generation. The automated version runs in n8n with separate strategist and copywriter agent steps, and one non-negotiable component: nothing publishes until a human approves it on WhatsApp.

The approval gate looks like a bottleneck. It’s the opposite. Because a person sees every post before it goes live, the agents upstream can run without supervision, and the whole pipeline needed far less defensive engineering. One checkpoint in the right place replaced a dozen guardrails in the wrong ones.

What we keep manual, and why

Client performance data stays out of third-party connectors. Our German clients operate under DSGVO with sector rules on top, and financial services data is nobody’s training corpus. GSC analysis runs on exports we control, processed under commercial API terms with short retention, self-hosted automation, no analytics connector with user-level reach. Slower by maybe an hour per analysis. Cheap insurance against a conversation with a regulator we never want to have.

**Claims and compliance language never ship without human review.** German law (§5 UWG on misleading commercial claims, PAngV on price representation) means a hallucinated statistic in a client’s ad copy is a legal exposure, and not a hypothetical one — we’ve caught a mock dashboard on a site that presented invented live metrics, and flagged it as exactly this risk. A model cannot carry that liability. A person signs off, every time.

The economics, without theatre

Individual analyses that used to take half a day now cost cents in API credits and minutes of wall-clock time. That part of the story every agency tells, and it’s true. Here’s the part fewer tell: the savings didn’t become lower prices, and mostly didn’t become margin either. They became coverage. Sites get full-inventory triage instead of sampling (one recent project meant decisioning 191 blog URLs individually — nobody does that manually). Content gets a style pass it never used to get. Diagnosis gets sixteen months of data instead of ninety days.

Same senior hours in. Considerably more work examined per hour. That’s the trade, honestly stated.

What can’t be copied

Fair challenge: MCP is open, n8n is open source, Claude Code is available to anyone with a subscription. If the stack is reproducible in a weekend, where’s the moat?

Not in the tools. In the spec library and the failure history. Every workflow above is on its third or fourth version, and the versions in between broke in instructive ways. The audit-and-fix-in-one-pass mistake. The “improve” instruction problem. The discovery that a style guide works only when it’s enforceable line by line. Those lessons are encoded in templates we run across every engagement, and each new campaign adds to them.

An in-house team can absolutely build this. It takes about a year and requires making our mistakes personally. Some companies should do exactly that. Most, in our experience, would rather rent the year.

If you want to see what your own site looks like through this pipeline, ask for an audit. The diagnostic layer runs fast enough now that showing you costs us very little, which is precisely why we offer it.

—

FAQ

What is MCP in an SEO context?
Model Context Protocol is an open standard letting AI models like Claude connect directly to tools and data sources. For SEO delivery it removes the manual export-and-combine step, so analysis draws on crawl data, performance data, and content inventories in one session.

What does “AI-native SEO agency” actually mean?
It means the delivery process was designed around AI from the start rather than having AI bolted onto manual workflows. In practice: spec-file-driven implementation, automated pipelines with human approval gates, and senior time concentrated on judgment instead of data assembly.

Does automating SEO delivery reduce quality?
It reduces quality when models run unreviewed, and raises it when the saved hours become wider coverage — full URL inventories instead of samples, longer data histories, style review on every piece. The determining factor is where the human checkpoints sit.

Is client data safe in AI-powered SEO workflows?
It depends entirely on the setup. Key controls: commercial API terms rather than consumer accounts, short data retention, self-hosted automation where possible, and keeping user-level analytics data out of third-party connectors. For regulated markets like German financial services, these are requirements before they’re preferences.

Can I build this stack in-house instead of hiring an agency?
Yes. The tools are public and none are expensive. Budget roughly a year for the workflow design, template refinement, and the instructive failures. Whether that year costs more than an agency retainer depends on how many sites you’ll run it across.