From WordPress & 50 pluginsto an AI-built Shop & Search
How a one-person Marzocchi hobby in collecting seals, forks and spares went from WordPress CMS with pagebuilder DIVI and fifty WP-plugins to AI-written website, based on FastAPI and AI-coded shop with a huge backend, numerous integrations and an AI Search, that answers questions from 2,000+ original documents — and why AI is what makes it all possible.
Dr-Zocchi is a one-person business. I had some small support, some programmers, some from community, a little from graphics, most important from doing CAD work – and I have a lot of producers.
But there is no web agency, no IT department, no content team. There is only one person here who knows Marzocchi forks — and a website, a shop, a parts archive, a support desk and a stack of orders that all want attention at the same time.
This is the story of how the new Dr-Zocchi website got built in 2026, five chapters to read, five big steps — and why AI made it all possible.
Chapter 1 — Standard Software: WordPress, DIVI and 50 plugins
My first websites I built myself manually in all details. It needed days and weeks and you need to write every single character yourself. From year to year building websites become more difficult, also building and maintaining took more time. You would be a fool to manually build it nowadays.
The first version of the shop was built the way most small businesses do it: using WordPress (40% of all websites on the world use it), plus a top Shop Plugin WooCommerce, the DIVI page builder (incl. standardize themes, pages and CSS) and one plugin for almost e-v-e-r-y-t-h-i-n-g that was missing.
Apart from some hotfixes not a line of code was written — it was dragging, dropping, clicking and searching forums. Hundreds of hours went into it, and in the end it ran on roughly fifty plugins.
It worked. But every plugin is another update to install, another thing that can conflict with the others, another subscription, another place to look when something breaks. A page builder makes a page look right and makes it all looking the same; it does not know that a fork model has a year range, a serial number scheme and fifty compatible parts. That knowledge lived in the owner’s head and in folders of PDFs — not on the website.
The problem was never the tools. The problem was that every hour spent keeping the website alive was an hour not spent on forks, parts and customers. And it is about costs.
At some time I had to change provider, because I needed one being able to auto-provide dynamic resources RAM + CPU power on demand to keep it running, which in the end costs ~300€ a month + plugins subscription costs. Simply crazy—and it was clear that things were going to get much worse.
Chapter 2 — Seal Kits & The Public Document Collection
Over the years I have manually collected a huge number of original Marzocchi material: service manuals, parts lists, catalogues, tech bulletins, ads, pictures — in English, Italian and German, etc.
Partly as clean PDFs, partly as scans of paper. The idea was simple: turn that collection into proper fork pages, with detailed, correct information for every model.
Doing that by hand means searching worldwide sources, reading & sorting hundred of pages & images and working them into something useful, rename,re-search again, send them to OCR (e.g. PDF’s based on images), reformat, deleting private information in files etc. wtf.
It was clear : when FF (fu*king FOX) turned off the manuals and all technical support in ~2018 without a single note to users and Bike Shops - still promising big mega BS like “still supporting users and brand” - it became critical to service our beloved forks.
Rebuilding seal kits for Marzocchi forks by re-engineering was not the biggest task to do, but it was support and manuals.
My 1st attempt of making Marzocchi documentation online in a Public Repository is still online : https://public-repository.org/ . ~ 1.500 documents, exploded drawings, part lists, adds and catalogues. Files have been intensive numbered (year, fork model, seals size, travel, etc.) to make a search available, but that’s only a minimal solution and search & filter can be only done on file names.
Chapter 3 — The first AI-written Marzocchi documents
In 2022 ChatGPT with some easy chat interface became public and in later years we understand the meaning and the possibilities of AI and LLM’s. But having a chat with an AI bot is a low level tool with limited capabilities. It does not help in the mass work, you need some better helper being able to do the work in working on mass data and search solutions.
If you think about AI bots answering full qualified specific questions you need to create your specific LLM yourself and train it with lot’s of specific sources. I tried this with several 3rd party programmers in 2025, but results where really bad. I needed to understand that if you want to have a high level LLM you need to have a good data source – the better the data the better the LLM and your AI. But without the missing Marzocchi Technical Knowledgebase and all the missing manuals it is not possible. I dropped this for while and let the life move on until getting another idea.
Trash talk : Something does not work? Simply use AI!(?)
Muha - no proper data - no good results from AI.
In the end of 2025 the evolution of AI went crazy and still is. Beside many other companies Anthropic built Claude and another tooling rise began – much more faster and wider than the last one, which in my focus was cloud computing and services based on Kubernetes and many others ~15 years ago. I know because I’m still working in a software company developing PLM/PDM software. We transformed the software to cloud based services and I was part of it – what a great experience this was.
The AI evolution went further and AI Agents were born and becoming available to the public. It became clear that this is a fantastic possibility to support my ideas and doings.
With AI my Documents Repository became an AI pipeline: the documents were read, OCR processed, parts lists are created & turned into clean tables, data/formats had been normalized and I tried to create a structured page for every single fork and it’s data, based on some self trained LLM. Compared to the past this is a huge win in time and efforts, all that documents have been overworked multiple times in some hours.
But my website got more and more in technical & costly problems and beside developing and learning using AI Vibe Coding with my Claude Code Agent and a lot of other tools I began to develop a new website and a new shop.
Chapter 4 — AI-driven coding: Rebuilding the whole Dr-Zocchi site
The next step was a big one: replacing the WordPress build with a content site , nice shop, cart and checkout with PayPal integration, order handling, support tickets, content in a lab & a blog section, B2B support, a massive admin backend area and much more. Written not by a costly 3rd party developer and not manually but by developing by working with Claude Code AI agents, session after session.
I can tell you that month learning and developing was one of the greatest experience and fun time in my life it the journey does not end here.
The new Dr-Zocchi site went live on 11 August 2026 after some time of parallel operation with the old shop, and the work has not stopped since: the project’s own change log lists more than 900 individual changes — fixes, data corrections and new features, each one numbered, described and traceable. Nobody can keep that in their head, so the project documents itself: every part of the system has a reference file, and the rules (never break a URL, never skip the build before a deploy, never publish an unchecked sitemap etc. pp.) are written down and enforced automatically before every commit.
I build my own software management system doing features/fixing bugs/data changes, release and service levels, committing software to GitHub. My agents are coding in parallel sessions, correcting and improving themselves, each session has it’s own name, testing it’s code himself either by testscripts, multiple subagents or surfing himself in a browser. Software has pre- and post processing, e.g. in auto code formatting, advanced code checking and much more.
You never want to send mails again to an SMTP server, when you can do using a 3rd party service by API, you need some graphic interface to work on product images for the shop? Do it with one click including automatic background removal, centering, sizing in a second. You need a ticket system because working in 5 postboxes is annoying? Import data from international sites? Integrating Ebay, DHL using API? All easy, just create services & functions yourself in some hours.
I never wrote a single bracket myself and never-ever my agent wasn’t able to fix a bug. Sometimes it needs several iterations, but I can tell you the learning curve for me was bombastic and the agents output was even much more bombastic as we are in a steady continuous developing & integration process (CI/CD).
When I’m programming at night, some of the results I get from “vibe coding” leave me speechless. Lessons like this :
Bro, this is the 4th iteration you tried to fix this fucking layout error xxxxxxxx. Why don’t you fix it? I told you all the details you need to know.
And after some seconds thinking he said : “You are right. My problem is : I can’t see.”
Wow. I was speechless. He want human?? He got blind?? Wtf?? Gosh.
I asked him: “What do you mean? You have a 2 way Chrome Integration using direct access or MCP Server, we had a lot of sessions and you surfed the site pages to check content in detail many times.”
And he said : “Yes, but obviously it does not work, I don’t know the reason. To be able to continue I assumed data. I guess that was wrong.”
Wow. Can you imagine your computer telling you it can’t see??
That’s somehow like HAL9000 telling you: “I’m sorry Dave, I can’t do that.” (nerd alarm).
I never edited installed a package or edited a config file on my new server. It needed a valid SSH2 key and proper secure handling to let AI do it all from my terminal. Gosh.
Chapter 5 — Search that understands the Question
In the meantime site and shop are almost perfect working for me. I had lots of ideas and during coding and waiting for the end of some agents tasks, I’m getting more ideas, things to test or to check – so many tasks that I have a running todo list beside using as new input for running multiple agents and subagents running 24h a day – if possible.
But for sure Vibe Coding is not the Holy Grail and AI is not 100% correct in all, but for me it is an absolute incredibly helpful tool being productive and solving my daily problems you read above.In the meantime I have built many other apps, tools and did extreme cool things for my daily doings.
Back to the Dr-Zocchi site : A site with 2,000+ documents and 5,400+ products in backend database (after a massive data collecting around the world) is only useful if people can find things.
The newest chapter from this weekend is SEARCH supported by AI: the shop search looks through names, part numbers, alternative numbers and the fork models a part fits; the document search reads the full text of every page — 11,536 pages in 2,047 documents (in the meantime we have created more than 500 additional part lists), including OCR of the scans and English translations of the Italian and German manuals.
So my try’s in creating LLM’s obviously failed, but hey : Claude is a cool trained AI able to do intelligent searches. So on top of all that documents - which btw had been completely overworked now - sits an AI answer, which is using Claude per API in realtime.
Ask a question in German, Italian or English — “which seal kit fits a Z1 from 2001?” — and a Claude model reads the best matching pages, parts & products and answers with numbered sources that open the original PDF page. It also integrates hits from the Shop parts to your service.
To make it better and to eliminate dumb results we keep it only answers questions about bicycle suspension, it is capped per day so costs stay predictable, and an additional built-in synonym list (1,093 terms in 127 groups across five languages) knows that tappo, Verschlusskappe and top cap are the same thing.
Try it: ask a question in the document search. If the answer is wrong, the 👎 button tells us what to fix. Together we will learn from users questions and make search better.
Summary & Conclusion
I showed you my experiences from this year about developing a new site with shop and more and why AI was the key to make it possible. Boring to users expecting crazy Marzocchi news, but mandatory to keep this things alive.
We now have a better user interface with AI driven Search on all document data.
But here is the real key point : in the end results are limited by sources and available data.
I hope at some time we will get access to the missing older Marzocchi data (lots are missing from 2009-2018), the Technical Knowledge Base, some more CD’s, more documents or papers from former Marzocchicians – we all know this data is alive somewhere.
Make it happen and Keep your Marzocchi alive.