Stefan Oversberg and Looy – Toolsammlung für Local SEO: AI Coding interview

Right now I'm working on making the individual tools interlock. Previously each tool was an island: answering reviews, writing posts, creating services. Now that hangs together. The competitor radar reads competitors' profiles, and the AI evaluates what their customers complain about and what my own customers praise. From that, content emerges directly: stories as a basis for texts, Google posts with images that I schedule straight away on Google, and location-related services. On top of that comes the ranking grid. It simulates search queries with phone and location at many points in the area and shows where a profile is actually found. And everything should work for agencies, with their own logo, own customers and own API keys.

AI coding is a real dialogue for me. I'm not a classic developer, I'm a Local SEO practitioner. I know what works in practice and what Google treats as spam. Claude implements that technically, I test immediately live with real profiles and send screenshots back when something sticks. Sometimes I also correct Claude. For example it said Google couldn't schedule posts, when that has long been possible. Or locations don't belong in the name of a service, but in the description. That way what used to take me weeks or paying a developer happens in hours. The domain expertise stays with me, though, and that is the decisive part.

LLMs: Building is done with Claude, currently Claude Opus, directly in the Claude app, which is linked to my machine via the desktop app. Claude thus works directly in my local project folder. Inside Looy itself I deliberately use several models: GPT-4o mini, Claude Haiku and Gemini 2.5 Flash for texts such as review replies, posts, services and analyses. For images I use gpt-image-1 or Gemini Flash Image. Agencies store their own keys, then everything runs through their account.

Harness and working method: The project's memory is a CLAUDE.md in the main folder. It holds every architecture decision, every table and every rule, for example that agencies may only use AI with their own keys. On top of that come project documents on the individual building blocks such as tenants, API options or the audit module. Before every change Claude creates a backup and afterwards checks the syntax of PHP and JavaScript. He renders new interfaces with test data in the browser before I even see them. I then upload and test live with real Google profiles.

Tech stack: PHP 8 without a framework, MySQL/PDO, vanilla JS, Leaflet with OpenStreetMap for the maps. Runs on perfectly ordinary web hosting; I set up the cron jobs in the hosting panel. Connected are:

Google: Business Profile API (reviews, posts, services, opening hours), Performance API, Places API, Search Console DataForSEO for rankings from Maps and the Local Finder as well as for competitor data SISTRIX for keyword data

Sub-agents, skills, MCP:

I use MCP the most. SISTRIX is connected via MCP, as is Claude in Chrome. That way Claude looks at pages himself and takes screenshots, for example of the dashboard for the application. Skills: I have my own skill for non-commodity content in Local SEO, which checks local service pages for content that only that business can deliver. I barely use sub-agents. For my work the continuous dialogue is more important, because every decision builds on the previous one. Research, for example which Google APIs still exist, Claude handles directly.

Honestly I never counted the tokens, because I work with a Claude subscription and don't bill per token. More tangible is the scope: Looy now has more than 340 PHP files and around 120,000 lines of PHP and JavaScript. On top of that come more than 120 database migrations that I applied gradually. The first files in today's project folder date from spring 2026; it really picked up speed in June.

On time: estimated 10 hours per week over 7 months (so about 300 hours). Most of that is not programming, but testing with real Google profiles, making domain decisions and correcting Claude when something is factually wrong. A feature like the ranking grid with map and competitor analysis stands in one morning. Previously I would have paid a developer several weeks for that.

Ongoing AI costs in the tool itself: review replies, posts and analyses run via GPT-4o mini, Claude Haiku and Gemini Flash. Those are cent amounts per call. Agencies use their own keys, so their costs don't land with me at all.

By training I'm a Diplom-Ökonom and not a trained developer. I started automating about ten years ago, with Ubot Studio. With that I built my own bots and small programs, entirely without classic code. It was about browser workflows that you otherwise click through by hand for hours. Later came websites and online marketing for the clients of my agency Semper Superior, and with that PHP and WordPress in a pragmatic sense.

I've been programming with AI since March 2026. It started with a single tool, Looy's review tool. Matching that came a customer card program, then the next tool, and so within a few months Looy grew to more than 40 tools.

The real drive is quite simple: I'm lazy. I'd rather think for four hours than work stupidly. Every task I do by hand for the third time becomes a tool. Previously I was limited to what Ubot offered. Today I describe to Claude what a tool must deliver domain-wise, and I get real software: with database, Google APIs, tenants for agencies and PDF reports. My Local SEO knowledge remains the raw material. I know what Google treats as spam and what businesses really need. AI coding turns that in days into what used to take a developer team.

I'm not a hardware person, so only my view from the software corner. Software works so well with AI because the feedback comes immediately. Write code, test, see errors, correct – that takes minutes. With hardware every round means: order parts, fabricate the board, assemble, measure. That costs days and real money. For design and firmware agents certainly already help a lot, but in the end a human has to check the prototype.

I see this limit with Looy as well. It always gets difficult where the real world comes into play: APIs that behave differently than documented, or Google rules the AI doesn't know. The agent is the fastest developer I've ever had, but no substitute for experience and control.