I Pointed a Swarm of AI Agents at a Dead Domain. Here's What Actually Moved the Needle.

Kemal Esensoy — Sat, 18 Jul 2026 06:26:24 +0000

Most "agentic SEO" posts are either screenshots with no method, or a method with no results. This one has both. And I'm going to be honest about which parts were the agents and which parts were just me knowing the topic.

No pivot. No expired domain. No inherited authority. Just a brand-new site, a weekend, a lot of coffee, and a lot of tokens.

Quick context

I run a German-learning iOS app as a side project. A bit over 4,000 users right now. deutschwunder.com is the companion web property, on a brand-new domain. First commit was July 2025, built on a Next.js blueprint I've used for years. I'm a React dev and SEO by trade, so the technical foundation was never the hard part.

Then real client work ate my entire autumn and winter. I did not touch the web app from roughly August 2025 until March 2026. It just sat there: a correct foundation and almost no content.

The whole revamp you're about to read happened over one weekend in March 2026.

The results first, so you know if this is worth reading

Google Search Console, last 6 months vs the previous 6:

Impressions: 26.5K to 479K, about 18x
Clicks: 558 to 5,420, about 9.7x
Average position: 45.5 to 11
CTR dropped from 2.1% to 1.1%, and that's not a regression. It's the site surfacing for a much wider query set with a lot of page-2 rankings. That's an opportunity bucket, more on it at the end.

Plausible, year to date:

22.8K unique visitors, up about 700% year over year
Top pages are /games (4.9K visitors) and a stack of blog posts
chatgpt.com is my #5 traffic source at 1.6K visitors. Copilot, claude.ai and perplexity all show up too.

Who actually did the work

The agents did not do the hard part.

The hard part was the keyword research and the angle, and that came from me actually knowing this topic. I learned German as a student myself, so I know which steps genuinely help a learner and which ones are filler. That domain knowledge is why the app works, and it's why the content clusters were not generic slop.

So learn this before you jump on a new project:

If you do not know your topic, agents will help you produce a lot of confidently mediocre content faster. Knowing the topic is the moat. The agents are the excavator, not the surveyor.

The workflow

1. Research and clustering. This is where the time went.

Before writing a single post, I built the content clusters by hand:

Ahrefs for volume and difficulty
DataForSEO for SERP data at scale
crawl4ai to scrape competitor pages and understand what was actually ranking
serper.dev for live SERP checks

I molded everything into 3 to 4 categories that match real search intent for German learners: learning German (grammar, listening, pronunciation), German culture, and listicles like best German movies on Netflix or best German learning apps. Every cluster came out of the keyword research, not out of vibes.

2. Batch content generation.

I have a batch content skill running on Claude Code, Opus only. I've never had a quality issue with Opus for this, so I've never bothered with smaller models. I fed it the cluster briefs. In March it produced 97 posts in one batch, which I then drip-published across March and April rather than dumping all at once.

The important detail: the skill does not just "write a blog post." Before writing anything it checks the steps. What internal links exist, what content already exists so it does not cannibalize or duplicate, which cluster this belongs to. When the knowledge base and the guardrails are set correctly, I genuinely did not have content I had to throw away. The pre-writing checks are what prevent the garbage.

3. Human review, every single piece.

I go over every blog post, every quiz, every game before it goes live. This is not a "wake up to a ranked site" story. The agent does the volume, I am the editor and the taste layer. I think that's the only version of this that isn't a liability.

4. Publishing.

Content goes straight into a Directus CMS through a CLI wrapper the agents call. The same Directus backend powers both the iOS app and the web app, which matters for the next point.

5. The quizzes and the actual funnel.

There are 403 quizzes on the backend. I'm constantly adding quizzes for the iOS app, and the web app shows a truncated version, only a handful. That's deliberate. The web quizzes and games are there to rank and engage, and the goal is to push people toward the iOS app download.

So the funnel is: blog acquires strangers, games and quizzes engage them and keep them on the web app, the app is the conversion. /games is my single biggest page.

One more lesson before we move on:

Building is not the hard part anymore. Direction is. Point a swarm of agents at nothing and you get a flock of chickens: motion without a destination. A goal with a plan gets you where you want to go, and it makes the distribution part easier too.

6. The SEO-audit loop.

I also built an SEO-audit skill with a CLI wrapper that connects to Google Search Console via a service account, and to Plausible. It flags the high-impression, low-click pages that need meta rewrites, the crawled-but-not-indexed pages that need to be rewritten or strengthened, and so on. That's the ongoing maintenance loop now that the content exists.

Every month: seo-audit, CTR check, check why pages are not being indexed, rinse and repeat.

7. Answer-engine visibility.

In July I added an llms.txt. Is it a must? I'm not sure. Even without it, chatgpt.com was already my #5 source of traffic. But remember when Google said domain authority is not a ranking factor, and then the leak happened. I added llms.txt because why not. What if they're not telling the whole story on this one either.

The caveat I'm not going to hide

The April traffic inflection is as much "the domain aged past roughly 9 months" as it is any single thing I did. New domains sit in a low-trust sandbox for roughly 8 to 9 months, then break out. I happened to load the content backlog in March, right before that trust threshold hit, so the backlog ignited at the same moment the domain matured.

So the causal story is not "agents made a rocket." It's "I had a correct foundation from day one, I used agents to load a large, well-clustered content backlog during the sandbox, and it was all sitting there ready when Google finally trusted the domain." The timing was partly luck. I'd rather say that than pretend I cracked a code and everyone will get the same outcome if they copy the steps.

TL;DR

Brand-new domain, correct technical SEO from day one, then one weekend in March using a batch content skill on top of real keyword research and real domain knowledge. Agents did the volume, I did the strategy and reviewed every piece, and domain age did more of the heavy lifting than any single tactic. Zero to about 479K monthly impressions and 22.8K visitors year to date, funneling toward a 4K-user iOS app.

That's the honest version. Including the boring parts.

I'm Kemal, and I run Wunderlandmedia, a one-person web development and SEO shop in Augsburg. If you want to talk about building agent workflows that produce content you don't have to throw away, that's the kind of thing I do.