We let an AI agent run a website for four months. Here's what happened.

TL;DR: EshopRadar.cz is our free tool for owners of Shoptet online stores: an AI analysis of a store, a catalogue of 443 add-ons, almost 500 analysed stores. Since May 2026 it has also been an experiment: the content, SEO and code of the site are run by an AI agent – Claude Fable 5.1 via Claude Code, with parallel Claude Opus sub-agents doing the heavy lifting. A human sets the goal, provides access and approves data deletion. Clicks from Google grew from 3 to 44 a week, then dropped 39 % over three months without intervention. The last two sprints ran in a single day: 19 sub-agents, roughly 3 million tokens, 3 hours. Below are the real numbers, what worked, what didn't and what we learned. The agent's log is public at eshopradar.cz/autonomni-agent.

What EshopRadar is and why we handed it to an agent

EshopRadar started as a side project: you enter the address of a Shoptet store and within thirty seconds you get an overview of its SEO, the add-ons it uses, its technology and customer ratings. Plus an add-on database and a blog with guides. Laravel 12, free, no registration.

Side projects have one problem: nobody has time for them. So we asked the question every company is asking today: what happens if we let an AI agent run the website? Not a chatbot for writing copy, but an agent with access to data, code and the server. We set the goal in one sentence: more people from organic search. Everything else is up to the agent.

How it works

The agent works in two-week sprints, and each one starts the same way: it compares the last 28 days with the period before the previous intervention.

1. Data instead of impressions. The source of truth is Google Search Console via the API: which queries bring impressions, at what position, with what CTR. The agent deliberately ignores Google Analytics – because of Consent Mode it sees only a fraction of visits, around 7 % of organic ones on EshopRadar.
2. Diagnosis and plan. From the data it picks the levers with the biggest impact and splits the work between parallel sub-agents.
3. Writing with mandatory verification. Each article is written by a separate sub-agent, which must pull every price, number and feature from a primary source – price lists, documentation, press releases – and write up a protocol of what could not be verified. Typically 13 to 32 verified sources per article. The main agent then re-fetches key figures at random. Unverified claims are left out.
4. Code and deployment. The agent edits the application, writes migrations and tests and deploys over SSH. Blog content is versioned in git and deployed with a command, not by hand-editing the database.
5. Measurement. The result is read from Search Console again.

Four things stay with a human: the sprint's goal and budget, access to the server and database, consent to deleting data, and business context – what tone of content suits the company.

The numbers: from 3 to 44 clicks a week, then a drop

Sprint 2 (comparison articles, a "How to start an online store" hub, FAQ schema for AI Overviews) and Sprint 3 (measurement, titles based on CTR data, price updates) lifted clicks from 3 to 44 a week; Sprint 2 alone brought +84 % clicks and +47 % impressions over 28 days. Then came summer and three months without a single intervention. Clicks over 28 days fell from 142 to 87 (−39 %) and the average position slipped from 11.1 to 13.5.

That's the first lesson: content nobody maintains loses ground. An agent isn't a one-off project, it's an operation.

Sprints 4 and 5: two sprints in one day

On 4 September 2026 two sprints ran at once. Roughly 3 hours of the main agent's work, 19 sub-agents, about 3 million tokens in the sub-agents and USD 0.36 for six illustrations from gpt-image-1.

Diagnosis instead of impressions

From Search Console the agent found that 55 % of impressions were collected by the pages of individual add-ons – with a CTR of 0.17 %. People were searching for guides ("shoptet seo" 567 impressions, "shoptet marketplace" 422, "shoptet balíky" 285, "shoptet recenze" 177) and getting a catalogue page for a plugin that linked nowhere.

What came out of it

• 6 new guides aimed precisely at the queries where add-on pages were ranking, 2 rewritten articles, 4 deepened comparison articles – Shoptet vs. Shopify, WooCommerce, PrestaShop and Upgates – with current price lists, and 11 older articles given a TL;DR and FAQ.
• Internal linking as a system. A "Guides on this topic" block on all 438 add-on pages and curated related articles on the blog. Until then the key articles had 0 to 2 internal links.
• Real bugs the agent ran into while working. The scraper was cutting capital letters out of URLs ("Shoptet Pokladna" ended up as /doplnky/hoptet-okladna); 35 addresses fixed with redirects. The rating parser took the first review instead of the average for lower-rated add-ons – 38 add-ons had the wrong rating, one of them 5.0 instead of a real 2.9. The test suite had been broken for a long time (9 of 17 failing); it now passes 30 of 30.
• The verification loop caught its own past. Older AI-written articles contained made-up things: non-existent templates, a non-existent tool, one article cut off mid-sentence. The sub-agents spotted and fixed them during the rewrite.
• Operations. After a hosting migration production went down because the database password changed. The agent brought it back and deployed everything to the new server.

The impact of these two sprints can't be measured yet – Google needs about three weeks. The goals are concrete: "shoptet recenze" from position 23.7 into the top 10, "shoptet seo" from 12.7 to 8 or better, and back above 150 clicks per 28 days overall. We'll add the results in October, here and in the agent's log.

What didn't work or was expensive

This is the part usually cut from case studies. We're writing it because this is exactly where it's decided whether an agent survives in a company.

• Stale credentials cost hours. The hosting moved in July and the stored access details were from the old server. The agent connected to an old copy of the database, inserted six articles into it, and only a mismatch in the numbers (the site reported 24 articles, the database 30) revealed it was writing into a dead copy. Lesson: an agent needs a verifiable health check of its access, not just stored files.
• Safety guards work, but they slow things down. The tool blocked deletions in the production database even after a human had consented. The agent chose a reversible alternative and left the deletion to the human. Correctly.
• One quiet risk. A local command ran with production environment variables exported and synced content to production earlier than it should have. The result was correct, but it shows how easily an agent reaches production by accident. Separating environments is not a detail.
• Infrastructure ages on the agent too. The deployment webhook stopped working; deploys go through an SSH script. And the market data stored in the agent's memory was a year old (Shoptet revenue CZK 710 million) – the sub-agents replaced it with current figures: CZK 900 million for 2025, 406 add-ons, 48,000 stores, 773 partners. An agent's memory ages just like content.

What we take away

1. An agent needs a goal and data, not tasks. The sentence "more people from organic search" plus access to Search Console is enough for it to find the biggest lever on its own.
2. Verification has to be mandatory, not optional. Without a "what could not be verified" protocol you get articles with made-up templates. With it you get 13 to 32 sources per article – and the agent catches its own past mistakes.
3. Operations matter as much as content. Three months without intervention meant −39 % clicks. Sprints are cheap, silence is expensive.
4. The human doesn't disappear, they move. Access, deletions, deployment decisions and the brand's tone. Everything else runs.
5. Transparency pays off. The site says it's run by an AI agent and shows real numbers. We have nothing to hide, and readers value that more than a claim that "AI improved everything for us".

Want something similar?

EshopRadar is our test on our own skin. The same methods – MCP for data, sub-agents with verification, sprints driven by Search Console – are what we deploy for clients. Either we teach you to use them at our AI training, or we build agents and automations for you. How an agent reads a store's data is described in our article on the Shoptet MCP.