Website relaunch with AI agents: what three weeks achieved and where I had to step in

Key points at a glance:

• Relaunch of a recipe website with 5,433 recipes using the AI agent Claude Code, in just over three weeks, with around 20 to 24 hours of my own time and around 220 US dollars for the AI tool and images.
• The AI did the routine work: 4,050 corrections in 2,346 recipes, 255 merged duplicates and recipe data for every recipe. It built the new design after several rounds of feedback.
• A human still had to step in at many points: original texts, food safety, technology after going live and every decision on law and costs.
• Two weeks after the first go-live steps: 35 per cent more clicks from Google Search. Almost half of all clicks now came through recipe results with an image, compared with practically none before.
• By our estimate, the same scope would have taken 470 to 600 working hours the classic way, probably more: around 22,000 to 28,000 euros in pure labour costs for an in-house team, or 42,000 to 54,000 euros at freelance rates.

gekonntgekocht.de has a personal story. When my father, Fritz Grundmann, a qualified master chef (Küchenmeister), retired in 2006, I set the website up for him as a hobby. Over many years he published his recipes there and built up a cooking glossary. Since his illness around three years ago, the content of the website had been left idle: recipes and glossary were no longer maintained, I only kept WordPress and the plugins up to date. New recipes had already become rare before that. He died in June 2026, and we are continuing the website in his spirit.

Technically, the website had been stuck in the early 2010s: at first it ran on a customised WordPress default theme, later on a purchased theme from 2011 with an early responsive design, extended over the years with plugins and customisations. A rebuild by developers with a contemporary design would never have been economical for a hobby project. That is exactly why I wanted to find out how far we could get with AI agents.

The inventory at the end of September 2026:

• Content: 5,433 published recipes, 3,309 glossary entries and 216 blog posts. Almost all recipes date from 2007 to 2014.
• Technology: custom changes sat in the parent theme and would have been lost with the next update. The home page had no H1, the site name was empty and ratings were marked up twice.
• Legacy: tables from plugins deleted long ago, including one statistics table of 92 MB, and two redirect plugins running side by side.
• Images: only 757 of the 5,433 recipes had a photo.

This showed clearly in Google Search. Several Google updates hit the website hard; the most severe drop came at the start of the years without maintenance. After the helpful content update in September 2023, the SISTRIX Visibility Index fell from 0.58 to 0.21 within a week, and after the March 2024 core update to below 0.01 by May. The June 2025 core update brought part of it back. After that the index rose in stages: sharply but not lastingly in May and June 2026, and steadily from mid-August 2026, still before the relaunch. This cannot be attributed to a single update.

Even so, in the two weeks before the first step, clicks were 22 per cent below the previous year, and recipe results with image and rating were practically non-existent.

We worked with Claude Code by Anthropic. It is an AI agent that does not just suggest text but reads files, writes scripts and runs them itself. It had access to a staging copy of the website, meaning files and database, never to the live site.

Three rules applied from the start:

• Back up first, then change: before every change the agent backs up the affected data and then records in the change log what it did and how to undo it.
• Check first, then apply: it checks every larger change in the browser on the staging site. Large packages are also checked by a second agent that sees only the result, not the first agent’s reasoning.
• Only a human takes things live: I carry out the transfer to the live site myself. That happened on 28 September, 30 September and 18 October 2026.

My role was that of a client with subject knowledge: setting goals, answering questions, checking results, fully for important things and by spot checks otherwise, and deciding when goals conflicted. Over a little more than three weeks, this took me around 20 to 24 hours by my own estimate. That it stayed at that is also down to experience: I have been working in web development and online marketing for more than 25 years. I recognised many problems at first glance and knew which solution would hold. Without that background, in my estimation the same relaunch would have taken considerably longer, and some errors would only have come to light once rankings or visitors were missing.

Work on staging started on 26 September 2026. What went live when:

The biggest gain came from tasks that are too large and too fiddly for a team:

• Copy-editing: the agent edited around 600 of the most visited recipes, carefully and in my father’s tone.
• Subject-matter plausibility: it checked all recipes and glossary entries for clear errors. In 2,346 recipes it made 4,050 small corrections: quantity errors such as “3000 g flour”, ingredients missing from the list, times that did not match the steps, and food safety details such as core temperatures for poultry. In the cooking glossary it checked all entries and corrected around 600.
• Duplicates: it merged 255 duplicate glossary entries into one address each and redirected the rest with a 301.
• Structured data: every recipe now has recipe markup with ingredients, steps, times and author. Recipe results in Google Search also need an image, which around 1,100 recipes have today. Until 18 October it was only the 757 recipes with an original photo. Whether Google will accept digitally created sample images as an image of the finished dish in the long run is an open question. We are watching this in the follow-up measurement. The agent calculated the nutritional values from the ingredients using data from the USDA nutrient database. The recipe states that these are calculated guide values.
• Clean-up: 26 old advertorials and competitions and 113 outdated blog posts are offline and return 410. Fonts are hosted locally, embedded maps have been removed.
• New features: a custom search, ingredient pages, a weekly planner and a new design.
• Images: some recipes without a photo have received a sample image. Claude used OpenAI’s image API with the gpt-image-2 model and described each image based on the recipe’s ingredients and steps. Every image is labelled as a digitally created sample photo.

Without agents we would have limited ourselves to the most important pages. With agents, the whole inventory was covered.

It did not run without errors. Working on staging caught most problems; some only showed up afterwards. On top of that come decisions that belong with a person anyway.

AI mistakes and checks after going live

• Original texts: when merging duplicates, the agent initially wrote new texts. But I wanted to keep my father’s original texts. Since then the rule has been: take the most complete original version and only list facts as notes.
• Quotes: older texts contained 24 quotes attributed to my father. I had them checked. We found only one of them in his own texts; the other 23 now appear as tips without a name.
• Technology after going live: cache plugins delay scripts and can change their order, and when moving from staging to live, files such as the robots.txt or settings such as scheduled publication dates are easily lost. That is why we check the live site right after every transfer.
• Programming errors: a new feature sent searches with glossary hits into an endless loop. This was found in the tests on staging and fixed before it went live.
• Design: the new design needed several rounds of feedback: spacing on mobile was off, fonts were sometimes serif and sometimes not, a box in the weekly planner simply did not look good. A person judges more reliably whether something feels right.

What a person has to decide

• Subject-matter borderline cases: for doneness I had to set the rule: chicken, turkey, other poultry except duck, minced meat and wild boar are cooked through. For duck breast, lamb, beef and other game my father’s method stays pink. Every recipe with duck breast or game cooked pink now includes the advice of the German Federal Institute for Risk Assessment, which recommends cooking both through.
• Unpublished recipes: recipes from his estate only go online after a subject-matter check. I had this built in as a fixed lock so that it is not forgotten in everyday work.
• Law and labelling: I decided that AI images are labelled directly on the image, which advertising the website shows and how consent for it is obtained. The AI did the research and the implementation.
• Costs and priorities: old affiliate links without income were removed, the newsletter was discontinued, and I paused new AI images for now until it is clear what they bring. No AI makes decisions like these.

A later review showed how important this part is. The AI found and corrected clear errors such as “3000 g flour” in large numbers. It is weaker on factual statements that require experience or current sources: a second review of the content against Google’s criteria for helpful content afterwards still found 15 wrong or outdated statements among 71 checked factual statements from 35 pages, partly the most visited and partly drawn at random, for example on vitamin B12 in seaweed or on mushrooms in the seasonal calendar. They were in old blog and glossary texts, and the first check had missed them. This review also ran with AI agents, but every flagged statement was checked against primary sources, and I decide which correction goes live. They have been corrected since 18 October. Subject-matter responsibility cannot be handed over to an AI. For topics concerning health, money or law (YMYL), this applies all the more.

For the measurement I compare the two weeks before the first transfer to the live site (14 to 27 September 2026) with the two weeks from 4 to 17 October 2026. The data comes from Google Search Console.

• Clicks from Google Search: from 1,195 to 1,611, an increase of 35 per cent. In the same period of the previous year clicks rose by only 3 per cent, so the season explains only a small part of the increase. Compared with the previous year, the website is now 2 per cent up instead of 22 per cent down.
• Impressions: from 143,758 to 173,445, an increase of 21 per cent.
• Recipe results: results that Google shows as a recipe with image, rating and times had 34 impressions and 1 click before, and 31,682 impressions and 767 clicks afterwards. Almost half of all clicks therefore came through this display. I did not add results that only show rating stars: Search Console lists a result under every appearance it shows, so clicks would otherwise be counted twice. Not all of this is additional traffic: recipe results brought 766 more clicks, while total clicks rose by 416. Some searches already led to the same recipes before, just through a standard result. Such results require structured data for recipes with the required properties image and name. Google also recommends rating and times.
• Visibility: the SISTRIX Visibility Index (weekly values) rose from 0.276 on 28 September to 0.326 on 19 October 2026, the highest value since the drop in September 2023. In mid-September, however, it was already at 0.300; for a site of this size the index fluctuates considerably. It had already been rising since spring, before the relaunch, see the chart above.

Is the increase really down to the relaunch? To check, I split the recipe pages into two groups: recipes that appeared as recipe results with an image or with rating stars after the go-live steps, and all other recipes with impressions. Google shows stars even without an image, which is why the first group is larger than the number of recipes with a photo. I compared the same two weeks as above and, as a check for seasonality, the same days in 2025.

The other recipes behave as they did the year before. The increase comes almost entirely from the recipes that now appear with an image or rating. The same pages gained only 6 per cent more clicks in the same period a year earlier, so the season does not explain the jump.

Two weeks is a short time, and the two go-live steps on 28 and 30 September cannot be cleanly separated. The redesign of 18 October is not yet included in these figures. Google also rolled out a spam update from 24 September to 8 October 2026. Whether it affected the website cannot be separated in the figures. The follow-up measurement by 30 November 2026 will show whether the increase holds.

My own effort was around 20 to 24 hours, spread over just over three weeks: setting goals, answering questions, checking and deciding. The AI tool ran on our Claude Max subscription at 200 US dollars a month. For just over three weeks, that is around 150 US dollars of subscription. Measured by usage, just under a third of it went to gekonntgekocht.de, so around 50 US dollars. The sample images cost 168.53 US dollars at OpenAI. Together that comes to around 220 US dollars.

What would the same scope have cost without AI agents? An experienced team would not read every recipe by hand, but search for recurring errors with check scripts and only work on the hits. Calculated this way, with our developers’ estimate for the technology and my own assumptions for the editorial work, without quoted prices:

• Technology: new design, custom search, ingredient pages and weekly planner on the existing WordPress site, plus recipe data, nutritional values, redirects and the check scripts for quantities, ingredients and times. Developers on our team estimate this at 150 hours at least.
• Corrections: 2,346 flagged recipes at three to four minutes each, about 120 to 160 hours.
• Copy-editing: around 600 recipes with just under one standard page of text each on average. One editing provider quotes five to eight standard pages an hour, so reading alone takes six to ten minutes per recipe. Including checking times and ingredients and entering the changes, that is ten to 15 minutes, about 100 to 150 hours.
• Cooking glossary: merging and redirecting 255 duplicates and correcting around 600 entries, checking the facts against sources, about 70 to 100 hours.
• Coordination and testing: about 30 to 40 hours.

Together that is around 470 to 600 hours, most of it routine editorial work. This is more of a lower bound: it does not include the sample images, going through the remaining glossary entries, the quotes or the data protection changes.

The cost of an hour can be derived from salaries. For web developers with a few years of experience I assume an annual salary of around 55,000 euros, for senior developers around 70,000 euros. This matches the median of 18 and 14 job ads with a stated salary on the German Federal Employment Agency’s job board (54,000 and 69,500 euros, October 2026). The agency’s pay atlas gives a median of 5,026 euros a month for web developers at specialist level, around 60,000 euros a year. With the employer’s social security contributions, statutory levies and accident insurance (around 25 per cent in total) and around 1,670 working hours a year (after holidays, public holidays and sick days), an hour costs around 47 euros on average, without office, software, administration and time that cannot be assigned to a project. That gives pure labour costs of around 22,000 to 28,000 euros. At freelance rates of around 90 euros an hour it would be 42,000 to 54,000 euros.

That was never an option for a hobby project, and many companies also limit themselves to the most important pages at such amounts.

This does not transfer one to one to a client project. Coordination, approvals and review take more time there than on my own website. The savings lie mainly in the routine work.

A relaunch with AI agents is not a push of a button, but it shifts what is feasible in a project. It pays off most for websites with a large existing inventory:

• Many similar pages: products, recipes, glossary or guide pages where every page can be checked and improved individually.
• Legacy technology: plugins and customisations added over many years that nobody fully understands any more.
• Content that needs checking: outdated details, duplicates, missing meta data and structured data.

gekonntgekocht.de was a simple case: one decision-maker, one system, no approval loops. In a company, IT, data protection, specialist departments and often a web agency are involved too. The agents do the routine work there as well. The real work lies in giving them the right rules and spotting errors before they go live. Someone who has seen it many times is more likely to notice that missing images block recipe results or that a cache plugin mixes up scripts. Some things still only show up live. That is why checking the live site right after every transfer is a fixed step. AI agents do not replace this experience; they make it pay off faster. This is the part we take on: setting the rules, checking the results, safeguarding the go-live. You decide what goes live.

I consider three things essential: a staging environment for the agent to work in, backups and a log for every change, and people who make the subject-matter decisions and are accountable for going live. For client projects this includes the question of which data may go into which AI tool. Google says it assesses the quality of content, not how it was produced. In the end, a human has to be accountable for that quality.

Typical relaunch errors are described in our article Typical errors in a website relaunch. How we support a relaunch, with and without AI agents, is described on our SEO relaunch support page.

Next on gekonntgekocht.de:

• Follow-up measurement: by 30 November 2026 I will compare the figures after the redesign and add them here, with the date.
• Loading time: the server’s first response could be faster, especially on mobile. That is not a task for AI agents but a hosting question, which we are tackling next.

And then there is the estate. My father left around 4,000 unpublished recipes. We are preparing them step by step, the agents help with that, and every recipe only goes online after a subject-matter check. That way his website keeps growing, even without him.